(function(){const e=document.createElement("link").relList;if(e&&e.supports&&e.supports("modulepreload"))return;for(const a of document.querySelectorAll('link[rel="modulepreload"]'))n(a);new MutationObserver(a=>{for(const i of a)if(i.type==="childList")for(const s of i.addedNodes)s.tagName==="LINK"&&s.rel==="modulepreload"&&n(s)}).observe(document,{childList:!0,subtree:!0});function r(a){const i={};return a.integrity&&(i.integrity=a.integrity),a.referrerPolicy&&(i.referrerPolicy=a.referrerPolicy),a.crossOrigin==="use-credentials"?i.credentials="include":a.crossOrigin==="anonymous"?i.credentials="omit":i.credentials="same-origin",i}function n(a){if(a.ep)return;a.ep=!0;const i=r(a);fetch(a.href,i)}})();const Ze={},n_=Object.freeze(Object.defineProperty({__proto__:null,default:Ze},Symbol.toStringTag,{value:"Module"})),a_="3.0.0-alpha.0",Ki=typeof self<"u",i_=Ki&&self.constructor.name==="DedicatedWorkerGlobalScope",Sp=Ki&&"caches"in self,s_=typeof navigator<"u"&&"gpu"in navigator,kp=typeof process<"u",o_=kp&&process?.release?.name==="node",Pi=!Tp(Ze),Ep=!Tp(Ze),Jr=Object.freeze({IS_BROWSER_ENV:Ki,IS_WEBWORKER_ENV:i_,IS_WEB_CACHE_AVAILABLE:Sp,IS_WEBGPU_AVAILABLE:s_,IS_PROCESS_AVAILABLE:kp,IS_NODE_ENV:o_,IS_FS_AVAILABLE:Pi,IS_PATH_AVAILABLE:Ep}),zo=Pi&&Ep,Cp=zo?Ze.dirname(Ze.dirname(Ze.fileURLToPath(import.meta.url))):"./",u_=zo?Ze.join(Cp,"/.cache/"):null,ju="/models/",l_=zo?Ze.join(Cp,ju):ju,wt={version:a_,backends:{onnx:{},tfjs:{}},allowRemoteModels:!0,remoteHost:"https://huggingface.co/",remotePathTemplate:"{model}/resolve/{revision}/",allowLocalModels:!Ki,localModelPath:l_,useFS:Pi,useBrowserCache:Sp,useFSCache:Pi,cacheDir:u_,useCustomCache:!1,customCache:null};function Tp(t){return Object.keys(t).length===0}const $t=class{constructor(){let t=function(...e){return t._call(...e)};return Object.setPrototypeOf(t,new.target.prototype)}_call(...t){throw Error("Must implement _call method in subclass")}};function Xn(t,e){t&&t(e)}function d_(t){return Object.fromEntries(Object.entries(t).map(([e,r])=>[r,e]))}function Ip(t){return t.replace(/[.*+?^${}()|[\]\\]/g,"\\$&")}function c_(t){return Number.isInteger(t)||typeof t=="bigint"}function qu(t){const e=[];let r=t;for(;Array.isArray(r);)e.push(r.length),r=r[0];return e}function ct(...t){return Array.prototype.concat.apply([],t)}function Ri(t,e){return Math.abs((t+e)%(2*e)-e)}function Xr(t,e){return Object.assign({},...e.map(r=>{if(t[r]!==void 0)return{[r]:t[r]}}))}var xs={};class Yi{_CONTENT_TYPE_MAP={txt:"text/plain",html:"text/html",css:"text/css",js:"text/javascript",json:"application/json",png:"image/png",jpg:"image/jpeg",jpeg:"image/jpeg",gif:"image/gif"};constructor(e){if(this.filePath=e,this.headers=new Headers,this.exists=Ze.existsSync(e),this.exists){this.status=200,this.statusText="OK";let r=Ze.statSync(e);this.headers.set("content-length",r.size.toString()),this.updateContentType();let n=this;this.body=new ReadableStream({start(a){n.arrayBuffer().then(i=>{a.enqueue(new Uint8Array(i)),a.close()})}})}else this.status=404,this.statusText="Not Found",this.body=null}updateContentType(){const e=this.filePath.toString().split(".").pop().toLowerCase();this.headers.set("content-type",this._CONTENT_TYPE_MAP[e]??"application/octet-stream")}clone(){let e=new Yi(this.filePath);return e.exists=this.exists,e.status=this.status,e.statusText=this.statusText,e.headers=new Headers(this.headers),e}async arrayBuffer(){return(await Ze.promises.readFile(this.filePath)).buffer}async blob(){const e=await Ze.promises.readFile(this.filePath);return new Blob([e],{type:this.headers.get("content-type")})}async text(){return await Ze.promises.readFile(this.filePath,"utf8")}async json(){return JSON.parse(await this.text())}}function ho(t,e=null,r=null){let n;try{n=new URL(t)}catch{return!1}return!(e&&!e.includes(n.protocol)||r&&!r.includes(n.hostname))}async function fo(t){if(wt.useFS&&!ho(t,["http:","https:","blob:"]))return new Yi(t);if(typeof process<"u"&&process?.release?.name==="node"){const e=!!xs?.TESTING_REMOTELY,r=wt.version,n=new Headers;if(n.set("User-Agent",`transformers.js/${r}; is_ci/${e};`),ho(t,["http:","https:"],["huggingface.co","hf.co"])){const i=xs?.HF_TOKEN??xs?.HF_ACCESS_TOKEN;i&&n.set("Authorization",`Bearer ${i}`)}return fetch(t,{headers:n})}else return fetch(t)}const p_={400:"Bad request error occurred while trying to load file",401:"Unauthorized access to file",403:"Forbidden access to file",404:"Could not locate file",408:"Request timeout error occurred while trying to load file",500:"Internal server error error occurred while trying to load file",502:"Bad gateway error occurred while trying to load file",503:"Service unavailable error occurred while trying to load file",504:"Gateway timeout error occurred while trying to load file"};function h_(t,e,r){if(!r)return null;const n=p_[t]??`Error (${t}) occurred while trying to load file`;throw Error(`${n}: "${e}".`)}class Ku{constructor(e){this.path=e}async match(e){let r=Ze.join(this.path,e),n=new Yi(r);if(n.exists)return n}async put(e,r){const n=Buffer.from(await r.arrayBuffer());let a=Ze.join(this.path,e);try{await Ze.promises.mkdir(Ze.dirname(a),{recursive:!0}),await Ze.promises.writeFile(a,n)}catch(i){console.warn("An error occurred while writing the file to cache:",i)}}}async function f_(t,...e){for(let r of e)try{let n=await t.match(r);if(n)return n}catch{continue}}async function ki(t,e,r=!0,n={}){if(!wt.allowLocalModels&&n.local_files_only)throw Error("Invalid configuration detected: local models are disabled (`env.allowLocalModels=false`) but you have requested to only use local models (`local_files_only=true`).");Xn(n.progress_callback,{status:"initiate",name:t,file:e});let a;if(!a&&wt.useBrowserCache){if(typeof caches>"u")throw Error("Browser cache is not available in this environment.");try{a=await caches.open("transformers-cache")}catch(k){console.warn("An error occurred while opening the browser cache:",k)}}if(!a&&wt.useFSCache&&(a=new Ku(n.cache_dir??wt.cacheDir)),!a&&wt.useCustomCache)throw Error("`env.useCustomCache=true`, but `env.customCache` is not defined.");const i=n.revision??"main";let s=ui(t,e),o=ui(wt.localModelPath,s),u=ui(wt.remoteHost,wt.remotePathTemplate.replaceAll("{model}",t).replaceAll("{revision}",encodeURIComponent(i)),e),l=i==="main"?s:ui(t,i,e),p,f=a instanceof Ku?l:u,m=!1,c;a&&(c=await f_(a,o,f));const y=c!==void 0;if(c===void 0){if(wt.allowLocalModels)if(ho(s,["http:","https:"])){if(n.local_files_only)throw new Error(`\`local_files_only=true\`, but attempted to load a remote file from: ${s}.`)}else try{c=await fo(o),p=o}catch($){console.warn(`Unable to load from local path "${o}": "${$}"`)}if(c===void 0||c.status===404){if(n.local_files_only||!wt.allowRemoteModels){if(r)throw Error(`\`local_files_only=true\` or \`env.allowRemoteModels=false\` and file was not found locally at "${o}".`);return null}if(c=await fo(u),c.status!==200)return h_(c.status,u,r);p=f}m=a&&typeof Response<"u"&&c instanceof Response&&c.status===200}Xn(n.progress_callback,{status:"download",name:t,file:e});const w={status:"progress",name:t,file:e};let v;return n.progress_callback?y&&typeof navigator<"u"&&/firefox/i.test(navigator.userAgent)?(v=new Uint8Array(await c.arrayBuffer()),Xn(n.progress_callback,{...w,progress:100,loaded:v.length,total:v.length})):v=await m_(c,k=>{Xn(n.progress_callback,{...w,...k})}):v=new Uint8Array(await c.arrayBuffer()),m&&p&&await a.match(p)===void 0&&await a.put(p,new Response(v,{headers:c.headers})).catch(k=>{console.warn(`Unable to add response to browser cache: ${k}.`)}),Xn(n.progress_callback,{status:"done",name:t,file:e}),v}async function Ar(t,e,r=!0,n={}){let a=await ki(t,e,r,n);if(a===null)return{};let s=new TextDecoder("utf-8").decode(a);return JSON.parse(s)}async function m_(t,e){const r=t.headers.get("Content-Length");r===null&&console.warn("Unable to determine content-length from response headers. Will expand buffer when needed.");let n=parseInt(r??"0"),a=new Uint8Array(n),i=0;const s=t.body.getReader();async function o(){const{done:u,value:l}=await s.read();if(u)return;let p=i+l.length;if(p>n){n=p;let m=new Uint8Array(n);m.set(a),a=m}a.set(l,i),i=p;const f=i/n*100;return e({progress:f,loaded:i,total:n}),o()}return await o(),a}function ui(...t){return t=t.map((e,r)=>(r&&(e=e.replace(new RegExp("^/"),"")),r!==t.length-1&&(e=e.replace(new RegExp("/$"),"")),e)),t.join("/")}function g_(t,[e,r,n],[a,i],s="bilinear",o=!1){const u=i/n,l=a/r,p=new t.constructor(a*i*e),f=r*n,m=a*i;for(let c=0;c<a;++c)for(let y=0;y<i;++y){const w=c*i+y,v=(y+.5)/u-.5,k=(c+.5)/l-.5;let $=Math.floor(v),C=Math.floor(k);const T=Math.min($+1,n-1),A=Math.min(C+1,r-1);$=Math.max($,0),C=Math.max(C,0);const B=v-$,R=k-C,D=(1-B)*(1-R),K=B*(1-R),j=(1-B)*R,ie=B*R,te=C*n,oe=A*n,re=te+$,M=te+T,P=oe+$,H=oe+T;for(let le=0;le<e;++le){const G=le*f;p[le*m+w]=D*t[G+re]+K*t[G+M]+j*t[G+P]+ie*t[G+H]}}return p}function __(t,e,r){const n=new Array(r.length),a=new Array(r.length);for(let o=r.length-1,u=1;o>=0;--o)a[o]=u,n[o]=e[r[o]],u*=n[o];const i=r.map((o,u)=>a[r.indexOf(u)]),s=new t.constructor(t.length);for(let o=0;o<t.length;++o){let u=0;for(let l=e.length-1,p=o;l>=0;--l)u+=p%e[l]*i[l],p=Math.floor(p/e[l]);s[u]=t[o]}return[s,n]}function Tn(t){const e=vr(t)[0],r=t.map(i=>Math.exp(i-e)),n=r.reduce((i,s)=>i+s,0);return r.map(i=>i/n)}function y_(t){return Tn(t).map(n=>Math.log(n))}function w_(t,e){let r=0;for(let n=0;n<t.length;++n)r+=t[n]*e[n];return r}function Ap(t,e=0){return t=Array.from(t).map((r,n)=>[n,r]).sort((r,n)=>n[1]-r[1]),e!==null&&e>0&&(t=t.slice(0,e)),t}function Mp(t){if(t.length===0)throw Error("Array must not be empty");let e=t[0],r=0;for(let n=1;n<t.length;++n)t[n]<e&&(e=t[n],r=n);return[e,r]}function vr(t){if(t.length===0)throw Error("Array must not be empty");let e=t[0],r=0;for(let n=1;n<t.length;++n)t[n]>e&&(e=t[n],r=n);return[Number(e),r]}function Op(t){return t>0&&(t&t-1)===0}class zp{constructor(e){if(this.size=e|0,this.size<=1||!Op(this.size))throw new Error("FFT size must be a power of two larger than 1");this._csize=e<<1,this.table=new Float64Array(this.size*2);for(let n=0;n<this.table.length;n+=2){const a=Math.PI*n/this.size;this.table[n]=Math.cos(a),this.table[n+1]=-Math.sin(a)}let r=0;for(let n=1;this.size>n;n<<=1)++r;this._width=r%2===0?r-1:r,this._bitrev=new Int32Array(1<<this._width);for(let n=0;n<this._bitrev.length;++n){this._bitrev[n]=0;for(let a=0;a<this._width;a+=2){const i=this._width-a-2;this._bitrev[n]|=(n>>>a&3)<<i}}}createComplexArray(){return new Float64Array(this._csize)}fromComplexArray(e,r){const n=r||new Array(e.length>>>1);for(let a=0;a<e.length;a+=2)n[a>>>1]=e[a];return n}toComplexArray(e,r){const n=r||this.createComplexArray();for(let a=0;a<n.length;a+=2)n[a]=e[a>>>1],n[a+1]=0;return n}transform(e,r){if(e===r)throw new Error("Input and output buffers must be different");this._transform4(e,r,1)}realTransform(e,r){if(e===r)throw new Error("Input and output buffers must be different");this._realTransform4(e,r,1)}inverseTransform(e,r){if(e===r)throw new Error("Input and output buffers must be different");this._transform4(e,r,-1);for(let n=0;n<e.length;++n)e[n]/=this.size}_transform4(e,r,n){const a=this._csize;let s=1<<this._width,o=a/s<<1,u,l;const p=this._bitrev;if(o===4)for(u=0,l=0;u<a;u+=o,++l){const m=p[l];this._singleTransform2(r,e,u,m,s)}else for(u=0,l=0;u<a;u+=o,++l){const m=p[l];this._singleTransform4(r,e,u,m,s,n)}const f=this.table;for(s>>=2;s>=2;s>>=2){o=a/s<<1;const m=o>>>2;for(u=0;u<a;u+=o){const c=u+m-1;for(let y=u,w=0;y<c;y+=2,w+=s){const v=y,k=v+m,$=k+m,C=$+m,T=e[v],A=e[v+1],B=e[k],R=e[k+1],D=e[$],K=e[$+1],j=e[C],ie=e[C+1],te=f[w],oe=n*f[w+1],re=B*te-R*oe,M=B*oe+R*te,P=f[2*w],H=n*f[2*w+1],le=D*P-K*H,G=D*H+K*P,ne=f[3*w],N=n*f[3*w+1],ae=j*ne-ie*N,fe=j*N+ie*ne,Ce=T+le,Be=A+G,Ke=T-le,Qe=A-G,We=re+ae,dt=M+fe,Ve=n*(re-ae),Ne=n*(M-fe);e[v]=Ce+We,e[v+1]=Be+dt,e[k]=Ke+Ne,e[k+1]=Qe-Ve,e[$]=Ce-We,e[$+1]=Be-dt,e[C]=Ke-Ne,e[C+1]=Qe+Ve}}}}_singleTransform2(e,r,n,a,i){const s=e[a],o=e[a+1],u=e[a+i],l=e[a+i+1];r[n]=s+u,r[n+1]=o+l,r[n+2]=s-u,r[n+3]=o-l}_singleTransform4(e,r,n,a,i,s){const o=i*2,u=i*3,l=e[a],p=e[a+1],f=e[a+i],m=e[a+i+1],c=e[a+o],y=e[a+o+1],w=e[a+u],v=e[a+u+1],k=l+c,$=p+y,C=l-c,T=p-y,A=f+w,B=m+v,R=s*(f-w),D=s*(m-v);r[n]=k+A,r[n+1]=$+B,r[n+2]=C+D,r[n+3]=T-R,r[n+4]=k-A,r[n+5]=$-B,r[n+6]=C-D,r[n+7]=T+R}_realTransform4(e,r,n){const a=this._csize;let s=1<<this._width,o=a/s<<1,u,l;const p=this._bitrev;if(o===4)for(u=0,l=0;u<a;u+=o,++l){const c=p[l];this._singleRealTransform2(r,e,u,c>>>1,s>>>1)}else for(u=0,l=0;u<a;u+=o,++l){const c=p[l];this._singleRealTransform4(r,e,u,c>>>1,s>>>1,n)}const f=this.table;for(s>>=2;s>=2;s>>=2){o=a/s<<1;const c=o>>>1,y=c>>>1,w=y>>>1;for(u=0;u<a;u+=o)for(let v=0,k=0;v<=w;v+=2,k+=s){const $=u+v,C=$+y,T=C+y,A=T+y,B=e[$],R=e[$+1],D=e[C],K=e[C+1],j=e[T],ie=e[T+1],te=e[A],oe=e[A+1],re=B,M=R,P=f[k],H=n*f[k+1],le=D*P-K*H,G=D*H+K*P,ne=f[2*k],N=n*f[2*k+1],ae=j*ne-ie*N,fe=j*N+ie*ne,Ce=f[3*k],Be=n*f[3*k+1],Ke=te*Ce-oe*Be,Qe=te*Be+oe*Ce,We=re+ae,dt=M+fe,Ve=re-ae,Ne=M-fe,ht=le+Ke,Vt=G+Qe,st=n*(le-Ke),zr=n*(G-Qe);if(e[$]=We+ht,e[$+1]=dt+Vt,e[C]=Ve+zr,e[C+1]=Ne-st,v===0){e[T]=We-ht,e[T+1]=dt-Vt;continue}if(v===w)continue;const hr=u+y-v,ir=u+c-v;e[hr]=Ve-n*zr,e[hr+1]=-Ne-n*st,e[ir]=We-n*ht,e[ir+1]=-dt+n*Vt}}const m=a>>>1;for(let c=2;c<m;c+=2)e[a-c]=e[c],e[a-c+1]=-e[c+1]}_singleRealTransform2(e,r,n,a,i){const s=e[a],o=e[a+i];r[n]=s+o,r[n+1]=0,r[n+2]=s-o,r[n+3]=0}_singleRealTransform4(e,r,n,a,i,s){const o=i*2,u=i*3,l=e[a],p=e[a+i],f=e[a+o],m=e[a+u],c=l+f,y=l-f,w=p+m,v=s*(p-m);r[n]=c+w,r[n+1]=0,r[n+2]=y,r[n+3]=-v,r[n+4]=c-w,r[n+5]=0,r[n+6]=y,r[n+7]=v}}class b_{constructor(e){const r=2*(e-1),n=2*(2*e-1),a=2**Math.ceil(Math.log2(n));this.bufferSize=a,this._a=r;const i=new Float64Array(n),s=new Float64Array(a);this._chirpBuffer=new Float64Array(a),this._buffer1=new Float64Array(a),this._buffer2=new Float64Array(a),this._outBuffer1=new Float64Array(a),this._outBuffer2=new Float64Array(a);const o=-2*Math.PI/e,u=Math.cos(o),l=Math.sin(o);for(let p=0;p<n>>1;++p){const f=(p+1-e)**2/2,m=Math.sqrt(u**2+l**2)**f,c=f*Math.atan2(l,u),y=2*p;i[y]=m*Math.cos(c),i[y+1]=m*Math.sin(c),s[y]=i[y],s[y+1]=-i[y+1]}this._slicedChirpBuffer=i.subarray(r,n),this._f=new zp(a>>1),this._f.transform(this._chirpBuffer,s)}_transform(e,r,n){const a=this._buffer1,i=this._buffer2,s=this._outBuffer1,o=this._outBuffer2,u=this._chirpBuffer,l=this._slicedChirpBuffer,p=this._a;if(n)for(let f=0;f<l.length;f+=2){const m=f+1,c=f>>1,y=r[c];a[f]=y*l[f],a[m]=y*l[m]}else for(let f=0;f<l.length;f+=2){const m=f+1;a[f]=r[f]*l[f]-r[m]*l[m],a[m]=r[f]*l[m]+r[m]*l[f]}this._f.transform(s,a);for(let f=0;f<u.length;f+=2){const m=f+1;i[f]=s[f]*u[f]-s[m]*u[m],i[m]=s[f]*u[m]+s[m]*u[f]}this._f.inverseTransform(o,i);for(let f=0;f<o.length;f+=2){const m=o[f+p],c=o[f+p+1],y=l[f],w=l[f+1];e[f]=m*y-c*w,e[f+1]=m*w+c*y}}transform(e,r){this._transform(e,r,!1)}realTransform(e,r){this._transform(e,r,!0)}}class v_{constructor(e){this.fft_length=e,this.isPowerOfTwo=Op(e),this.isPowerOfTwo?(this.fft=new zp(e),this.outputBufferSize=2*e):(this.fft=new b_(e),this.outputBufferSize=this.fft.bufferSize)}realTransform(e,r){this.fft.realTransform(e,r)}transform(e,r){this.fft.transform(e,r)}}function $_(t,e){if(e%2===0||e<=0)throw new Error("Window size must be a positive odd number");const r=new t.constructor(t.length),n=new t.constructor(e),a=Math.floor(e/2);for(let i=0;i<t.length;++i){let s=0;for(let o=-a;o<=a;++o){let u=i+o;u<0?u=Math.abs(u):u>=t.length&&(u=2*(t.length-1)-u),n[s++]=t[u]}n.sort(),r[i]=n[a]}return r}function Ss(t,e){const r=Math.pow(10,e);return Math.round(t*r)/r}function x_(t){const e=Math.round(t);return Math.abs(t)%1===.5?e%2===0?e:e-1:e}/*!
 * ONNX Runtime Web v1.18.0
 * Copyright (c) Microsoft Corporation. All rights reserved.
 * Licensed under the MIT License.
 */var Po=Object.defineProperty,S_=Object.getOwnPropertyDescriptor,k_=Object.getOwnPropertyNames,E_=Object.prototype.hasOwnProperty,Z=(t,e)=>()=>(t&&(e=t(t=0)),e),In=(t,e)=>()=>(e||t((e={exports:{}}).exports,e),e.exports),An=(t,e)=>{for(var r in e)Po(t,r,{get:e[r],enumerable:!0})},C_=(t,e,r,n)=>{if(e&&typeof e=="object"||typeof e=="function")for(let a of k_(e))!E_.call(t,a)&&a!==r&&Po(t,a,{get:()=>e[a],enumerable:!(n=S_(e,a))||n.enumerable});return t},Qr=t=>C_(Po({},"__esModule",{value:!0}),t),Qn,Er,Yr,Yu,Ro,Bo=Z(()=>{Qn=new Map,Er=[],Yr=(t,e,r)=>{if(e&&typeof e.init=="function"&&typeof e.createInferenceSessionHandler=="function"){let n=Qn.get(t);if(n===void 0)Qn.set(t,{backend:e,priority:r});else{if(n.priority>r)return;if(n.priority===r&&n.backend!==e)throw new Error(`cannot register backend "${t}" using priority ${r}`)}if(r>=0){let a=Er.indexOf(t);a!==-1&&Er.splice(a,1);for(let i=0;i<Er.length;i++)if(Qn.get(Er[i]).priority<=r){Er.splice(i,0,t);return}Er.push(t)}return}throw new TypeError("not a valid backend")},Yu=async t=>{let e=Qn.get(t);if(!e)return"backend not found.";if(e.initialized)return e.backend;if(e.aborted)return e.error;{let r=!!e.initPromise;try{return r||(e.initPromise=e.backend.init(t)),await e.initPromise,e.initialized=!0,e.backend}catch(n){return r||(e.error=`${n}`,e.aborted=!0),e.error}finally{delete e.initPromise}}},Ro=async t=>{let e=t.executionProviders||[],r=e.map(u=>typeof u=="string"?u:u.name),n=r.length===0?Er:r,a,i=[],s=new Set;for(let u of n){let l=await Yu(u);typeof l=="string"?i.push({name:u,err:l}):(a||(a=l),a===l&&s.add(u))}if(!a)throw new Error(`no available backend found. ERR: ${i.map(u=>`[${u.name}] ${u.err}`).join(", ")}`);for(let{name:u,err:l}of i)r.includes(u)&&console.warn(`removing requested execution provider "${u}" from session options because it is not available: ${l}`);let o=e.filter(u=>s.has(typeof u=="string"?u:u.name));return[a,new Proxy(t,{get:(u,l)=>l==="executionProviders"?o:Reflect.get(u,l)})]}}),T_=Z(()=>{Bo()}),Pp,I_=Z(()=>{Pp="1.18.0"}),ks,Ft,Rp=Z(()=>{I_(),ks="warning",Ft={wasm:{},webgl:{},webgpu:{},versions:{common:Pp},set logLevel(t){if(t!==void 0){if(typeof t!="string"||["verbose","info","warning","error","fatal"].indexOf(t)===-1)throw new Error(`Unsupported logging level: ${t}`);ks=t}},get logLevel(){return ks}},Object.defineProperty(Ft,"logLevel",{enumerable:!0})}),Ue,A_=Z(()=>{Rp(),Ue=Ft}),Bp,Dp,M_=Z(()=>{Bp=(t,e)=>{let r=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);r.width=t.dims[3],r.height=t.dims[2];let n=r.getContext("2d");if(n!=null){let a,i;e?.tensorLayout!==void 0&&e.tensorLayout==="NHWC"?(a=t.dims[2],i=t.dims[3]):(a=t.dims[3],i=t.dims[2]);let s=e?.format!==void 0?e.format:"RGB",o=e?.norm,u,l;o===void 0||o.mean===void 0?u=[255,255,255,255]:typeof o.mean=="number"?u=[o.mean,o.mean,o.mean,o.mean]:(u=[o.mean[0],o.mean[1],o.mean[2],0],o.mean[3]!==void 0&&(u[3]=o.mean[3])),o===void 0||o.bias===void 0?l=[0,0,0,0]:typeof o.bias=="number"?l=[o.bias,o.bias,o.bias,o.bias]:(l=[o.bias[0],o.bias[1],o.bias[2],0],o.bias[3]!==void 0&&(l[3]=o.bias[3]));let p=i*a,f=0,m=p,c=p*2,y=-1;s==="RGBA"?(f=0,m=p,c=p*2,y=p*3):s==="RGB"?(f=0,m=p,c=p*2):s==="RBG"&&(f=0,c=p,m=p*2);for(let w=0;w<i;w++)for(let v=0;v<a;v++){let k=(t.data[f++]-l[0])*u[0],$=(t.data[m++]-l[1])*u[1],C=(t.data[c++]-l[2])*u[2],T=y===-1?255:(t.data[y++]-l[3])*u[3];n.fillStyle="rgba("+k+","+$+","+C+","+T+")",n.fillRect(v,w,1,1)}if("toDataURL"in r)return r.toDataURL();throw new Error("toDataURL is not supported")}else throw new Error("Can not access image data")},Dp=(t,e)=>{let r=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d"),n;if(r!=null){let a,i,s;e?.tensorLayout!==void 0&&e.tensorLayout==="NHWC"?(a=t.dims[2],i=t.dims[1],s=t.dims[3]):(a=t.dims[3],i=t.dims[2],s=t.dims[1]);let o=e!==void 0&&e.format!==void 0?e.format:"RGB",u=e?.norm,l,p;u===void 0||u.mean===void 0?l=[255,255,255,255]:typeof u.mean=="number"?l=[u.mean,u.mean,u.mean,u.mean]:(l=[u.mean[0],u.mean[1],u.mean[2],255],u.mean[3]!==void 0&&(l[3]=u.mean[3])),u===void 0||u.bias===void 0?p=[0,0,0,0]:typeof u.bias=="number"?p=[u.bias,u.bias,u.bias,u.bias]:(p=[u.bias[0],u.bias[1],u.bias[2],0],u.bias[3]!==void 0&&(p[3]=u.bias[3]));let f=i*a;if(e!==void 0&&(e.format!==void 0&&s===4&&e.format!=="RGBA"||s===3&&e.format!=="RGB"&&e.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");let m=4,c=0,y=1,w=2,v=3,k=0,$=f,C=f*2,T=-1;o==="RGBA"?(k=0,$=f,C=f*2,T=f*3):o==="RGB"?(k=0,$=f,C=f*2):o==="RBG"&&(k=0,C=f,$=f*2),n=r.createImageData(a,i);for(let A=0;A<i*a;c+=m,y+=m,w+=m,v+=m,A++)n.data[c]=(t.data[k++]-p[0])*l[0],n.data[y]=(t.data[$++]-p[1])*l[1],n.data[w]=(t.data[C++]-p[2])*l[2],n.data[v]=T===-1?255:(t.data[T++]-p[3])*l[3]}else throw new Error("Can not access image data");return n}}),li,Np,Fp,Lp,Up,O_=Z(()=>{Do(),li=(t,e)=>{if(t===void 0)throw new Error("Image buffer must be defined");if(e.height===void 0||e.width===void 0)throw new Error("Image height and width must be defined");if(e.tensorLayout==="NHWC")throw new Error("NHWC Tensor layout is not supported yet");let{height:r,width:n}=e,a=e.norm??{mean:255,bias:0},i,s;typeof a.mean=="number"?i=[a.mean,a.mean,a.mean,a.mean]:i=[a.mean[0],a.mean[1],a.mean[2],a.mean[3]??255],typeof a.bias=="number"?s=[a.bias,a.bias,a.bias,a.bias]:s=[a.bias[0],a.bias[1],a.bias[2],a.bias[3]??0];let o=e.format!==void 0?e.format:"RGBA",u=e.tensorFormat!==void 0&&e.tensorFormat!==void 0?e.tensorFormat:"RGB",l=r*n,p=u==="RGBA"?new Float32Array(l*4):new Float32Array(l*3),f=4,m=0,c=1,y=2,w=3,v=0,k=l,$=l*2,C=-1;o==="RGB"&&(f=3,m=0,c=1,y=2,w=-1),u==="RGBA"?C=l*3:u==="RBG"?(v=0,$=l,k=l*2):u==="BGR"&&($=0,k=l,v=l*2);for(let T=0;T<l;T++,m+=f,y+=f,c+=f,w+=f)p[v++]=(t[m]+s[0])/i[0],p[k++]=(t[c]+s[1])/i[1],p[$++]=(t[y]+s[2])/i[2],C!==-1&&w!==-1&&(p[C++]=(t[w]+s[3])/i[3]);return u==="RGBA"?new rr("float32",p,[1,4,r,n]):new rr("float32",p,[1,3,r,n])},Np=async(t,e)=>{let r=typeof HTMLImageElement<"u"&&t instanceof HTMLImageElement,n=typeof ImageData<"u"&&t instanceof ImageData,a=typeof ImageBitmap<"u"&&t instanceof ImageBitmap,i=typeof t=="string",s,o=e??{},u=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},l=p=>p instanceof HTMLCanvasElement||p instanceof OffscreenCanvas?p.getContext("2d"):null;if(r){let p=u();p.width=t.width,p.height=t.height;let f=l(p);if(f!=null){let m=t.height,c=t.width;if(e!==void 0&&e.resizedHeight!==void 0&&e.resizedWidth!==void 0&&(m=e.resizedHeight,c=e.resizedWidth),e!==void 0){if(o=e,e.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");o.tensorFormat="RGBA",o.height=m,o.width=c}else o.tensorFormat="RGBA",o.height=m,o.width=c;f.drawImage(t,0,0),s=f.getImageData(0,0,c,m).data}else throw new Error("Can not access image data")}else if(n){let p,f;if(e!==void 0&&e.resizedWidth!==void 0&&e.resizedHeight!==void 0?(p=e.resizedHeight,f=e.resizedWidth):(p=t.height,f=t.width),e!==void 0&&(o=e),o.format="RGBA",o.height=p,o.width=f,e!==void 0){let m=u();m.width=f,m.height=p;let c=l(m);if(c!=null)c.putImageData(t,0,0),s=c.getImageData(0,0,f,p).data;else throw new Error("Can not access image data")}else s=t.data}else if(a){if(e===void 0)throw new Error("Please provide image config with format for Imagebitmap");let p=u();p.width=t.width,p.height=t.height;let f=l(p);if(f!=null){let m=t.height,c=t.width;return f.drawImage(t,0,0,c,m),s=f.getImageData(0,0,c,m).data,o.height=m,o.width=c,li(s,o)}else throw new Error("Can not access image data")}else{if(i)return new Promise((p,f)=>{let m=u(),c=l(m);if(!t||!c)return f();let y=new Image;y.crossOrigin="Anonymous",y.src=t,y.onload=()=>{m.width=y.width,m.height=y.height,c.drawImage(y,0,0,m.width,m.height);let w=c.getImageData(0,0,m.width,m.height);o.height=m.height,o.width=m.width,p(li(w.data,o))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(s!==void 0)return li(s,o);throw new Error("Input data provided is not supported - aborted tensor creation")},Fp=(t,e)=>{let{width:r,height:n,download:a,dispose:i}=e,s=[1,n,r,4];return new rr({location:"texture",type:"float32",texture:t,dims:s,download:a,dispose:i})},Lp=(t,e)=>{let{dataType:r,dims:n,download:a,dispose:i}=e;return new rr({location:"gpu-buffer",type:r??"float32",gpuBuffer:t,dims:n,download:a,dispose:i})},Up=(t,e,r)=>new rr({location:"cpu-pinned",type:t,data:e,dims:r??[e.length]})}),jr,sa,Es,Wp,z_=Z(()=>{jr=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array]]),sa=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),Es=!1,Wp=()=>{if(!Es){Es=!0;let t=typeof BigInt64Array<"u"&&BigInt64Array.from,e=typeof BigUint64Array<"u"&&BigUint64Array.from,r=typeof Float16Array<"u"&&Float16Array.from;t&&(jr.set("int64",BigInt64Array),sa.set(BigInt64Array,"int64")),e&&(jr.set("uint64",BigUint64Array),sa.set(BigUint64Array,"uint64")),r?(jr.set("float16",Float16Array),sa.set(Float16Array,"float16")):jr.set("float16",Uint16Array)}}}),Vp,Gp,P_=Z(()=>{Do(),Vp=t=>{let e=1;for(let r=0;r<t.length;r++){let n=t[r];if(typeof n!="number"||!Number.isSafeInteger(n))throw new TypeError(`dims[${r}] must be an integer, got: ${n}`);if(n<0)throw new RangeError(`dims[${r}] must be a non-negative integer, got: ${n}`);e*=n}return e},Gp=(t,e)=>{switch(t.location){case"cpu":return new rr(t.type,t.data,e);case"cpu-pinned":return new rr({location:"cpu-pinned",data:t.data,type:t.type,dims:e});case"texture":return new rr({location:"texture",texture:t.texture,type:t.type,dims:e});case"gpu-buffer":return new rr({location:"gpu-buffer",gpuBuffer:t.gpuBuffer,type:t.type,dims:e});default:throw new Error(`tensorReshape: tensor location ${t.location} is not supported`)}}}),rr,Do=Z(()=>{M_(),O_(),z_(),P_(),rr=class{constructor(t,e,r){Wp();let n,a;if(typeof t=="object"&&"location"in t)switch(this.dataLocation=t.location,n=t.type,a=t.dims,t.location){case"cpu-pinned":{let s=jr.get(n);if(!s)throw new TypeError(`unsupported type "${n}" to create tensor from pinned buffer`);if(!(t.data instanceof s))throw new TypeError(`buffer should be of type ${s.name}`);this.cpuData=t.data;break}case"texture":{if(n!=="float32")throw new TypeError(`unsupported type "${n}" to create tensor from texture`);this.gpuTextureData=t.texture,this.downloader=t.download,this.disposer=t.dispose;break}case"gpu-buffer":{if(n!=="float32"&&n!=="float16"&&n!=="int32"&&n!=="int64"&&n!=="uint32"&&n!=="uint8"&&n!=="bool")throw new TypeError(`unsupported type "${n}" to create tensor from gpu buffer`);this.gpuBufferData=t.gpuBuffer,this.downloader=t.download,this.disposer=t.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let s,o;if(typeof t=="string")if(n=t,o=r,t==="string"){if(!Array.isArray(e))throw new TypeError("A string tensor's data must be a string array.");s=e}else{let u=jr.get(t);if(u===void 0)throw new TypeError(`Unsupported tensor type: ${t}.`);if(Array.isArray(e)){if(t==="float16"&&u===Uint16Array)throw new TypeError("Creating a float16 tensor from number array is not supported. Please use Uint16Array as data.");t==="uint64"||t==="int64"?s=u.from(e,BigInt):s=u.from(e)}else if(e instanceof u)s=e;else throw new TypeError(`A ${n} tensor's data must be type of ${u}`)}else if(o=e,Array.isArray(t)){if(t.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");let u=typeof t[0];if(u==="string")n="string",s=t;else if(u==="boolean")n="bool",s=Uint8Array.from(t);else throw new TypeError(`Invalid element type of data array: ${u}.`)}else{let u=sa.get(t.constructor);if(u===void 0)throw new TypeError(`Unsupported type for tensor data: ${t.constructor}.`);n=u,s=t}if(o===void 0)o=[s.length];else if(!Array.isArray(o))throw new TypeError("A tensor's dims must be a number array");a=o,this.cpuData=s,this.dataLocation="cpu"}let i=Vp(a);if(this.cpuData&&i!==this.cpuData.length)throw new Error(`Tensor's size(${i}) does not match data length(${this.cpuData.length}).`);this.type=n,this.dims=a,this.size=i}static async fromImage(t,e){return Np(t,e)}static fromTexture(t,e){return Fp(t,e)}static fromGpuBuffer(t,e){return Lp(t,e)}static fromPinnedBuffer(t,e,r){return Up(t,e,r)}toDataURL(t){return Bp(this,t)}toImageData(t){return Dp(this,t)}get data(){if(this.ensureValid(),!this.cpuData)throw new Error("The data is not on CPU. Use `getData()` to download GPU data to CPU, or use `texture` or `gpuBuffer` property to access the GPU data directly.");return this.cpuData}get location(){return this.dataLocation}get texture(){if(this.ensureValid(),!this.gpuTextureData)throw new Error("The data is not stored as a WebGL texture.");return this.gpuTextureData}get gpuBuffer(){if(this.ensureValid(),!this.gpuBufferData)throw new Error("The data is not stored as a WebGPU buffer.");return this.gpuBufferData}async getData(t){switch(this.ensureValid(),this.dataLocation){case"cpu":case"cpu-pinned":return this.data;case"texture":case"gpu-buffer":{if(!this.downloader)throw new Error("The current tensor is not created with a specified data downloader.");if(this.isDownloading)throw new Error("The current tensor is being downloaded.");try{this.isDownloading=!0;let e=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=e,t&&this.disposer&&(this.disposer(),this.disposer=void 0),e}finally{this.isDownloading=!1}}default:throw new Error(`cannot get data from location: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(t){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return Gp(this,t)}}}),vt,No=Z(()=>{Do(),vt=rr}),ha,Cs,nr,Lt,Hp=Z(()=>{Rp(),ha=(t,e)=>{(typeof Ft.trace>"u"?!Ft.wasm.trace:!Ft.trace)||console.timeStamp(`${t}::ORT::${e}`)},Cs=(t,e)=>{let r=new Error().stack?.split(/\r\n|\r|\n/g)||[],n=!1;for(let a=0;a<r.length;a++){if(n&&!r[a].includes("TRACE_FUNC")){let i=`FUNC_${t}::${r[a].trim().split(" ")[1]}`;e&&(i+=`::${e}`),ha("CPU",i);return}r[a].includes("TRACE_FUNC")&&(n=!0)}},nr=t=>{(typeof Ft.trace>"u"?!Ft.wasm.trace:!Ft.trace)||Cs("BEGIN",t)},Lt=t=>{(typeof Ft.trace>"u"?!Ft.wasm.trace:!Ft.trace)||Cs("END",t)}}),jp,R_=Z(()=>{Bo(),No(),Hp(),jp=class qp{constructor(e){this.handler=e}async run(e,r,n){nr();let a={},i={};if(typeof e!="object"||e===null||e instanceof vt||Array.isArray(e))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let s=!0;if(typeof r=="object"){if(r===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(r instanceof vt)throw new TypeError("'fetches' cannot be a Tensor");if(Array.isArray(r)){if(r.length===0)throw new TypeError("'fetches' cannot be an empty array.");s=!1;for(let l of r){if(typeof l!="string")throw new TypeError("'fetches' must be a string array or an object.");if(this.outputNames.indexOf(l)===-1)throw new RangeError(`'fetches' contains invalid output name: ${l}.`);a[l]=null}if(typeof n=="object"&&n!==null)i=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else{let l=!1,p=Object.getOwnPropertyNames(r);for(let f of this.outputNames)if(p.indexOf(f)!==-1){let m=r[f];(m===null||m instanceof vt)&&(l=!0,s=!1,a[f]=m)}if(l){if(typeof n=="object"&&n!==null)i=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else i=r}}else if(typeof r<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let l of this.inputNames)if(typeof e[l]>"u")throw new Error(`input '${l}' is missing in 'feeds'.`);if(s)for(let l of this.outputNames)a[l]=null;let o=await this.handler.run(e,a,i),u={};for(let l in o)if(Object.hasOwnProperty.call(o,l)){let p=o[l];p instanceof vt?u[l]=p:u[l]=new vt(p.type,p.data,p.dims)}return Lt(),u}async release(){return this.handler.dispose()}static async create(e,r,n,a){nr();let i,s={};if(typeof e=="string"){if(i=e,typeof r=="object"&&r!==null)s=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(e instanceof Uint8Array){if(i=e,typeof r=="object"&&r!==null)s=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(e instanceof ArrayBuffer||typeof SharedArrayBuffer<"u"&&e instanceof SharedArrayBuffer){let p=e,f=0,m=e.byteLength;if(typeof r=="object"&&r!==null)s=r;else if(typeof r=="number"){if(f=r,!Number.isSafeInteger(f))throw new RangeError("'byteOffset' must be an integer.");if(f<0||f>=p.byteLength)throw new RangeError(`'byteOffset' is out of range [0, ${p.byteLength}).`);if(m=e.byteLength-f,typeof n=="number"){if(m=n,!Number.isSafeInteger(m))throw new RangeError("'byteLength' must be an integer.");if(m<=0||f+m>p.byteLength)throw new RangeError(`'byteLength' is out of range (0, ${p.byteLength-f}].`);if(typeof a=="object"&&a!==null)s=a;else if(typeof a<"u")throw new TypeError("'options' must be an object.")}else if(typeof n<"u")throw new TypeError("'byteLength' must be a number.")}else if(typeof r<"u")throw new TypeError("'options' must be an object.");i=new Uint8Array(p,f,m)}else throw new TypeError("Unexpected argument[0]: must be 'path' or 'buffer'.");let[o,u]=await Ro(s),l=await o.createInferenceSessionHandler(i,u);return Lt(),new qp(l)}startProfiling(){this.handler.startProfiling()}endProfiling(){this.handler.endProfiling()}get inputNames(){return this.handler.inputNames}get outputNames(){return this.handler.outputNames}}}),Fo,B_=Z(()=>{R_(),Fo=jp}),D_=Z(()=>{}),N_=Z(()=>{}),F_=Z(()=>{}),L_=Z(()=>{}),Xu,Kp,U_=Z(()=>{Bo(),No(),Xu="Training backend could not be resolved. Make sure you're using the correct configuration & WebAssembly files.",Kp=class Yp{constructor(e,r,n){this.handler=e,this.hasOptimizerModel=r,this.hasEvalModel=n}get trainingInputNames(){return this.handler.inputNames}get trainingOutputNames(){return this.handler.outputNames}get evalInputNames(){if(this.hasEvalModel)return this.handler.evalInputNames;throw new Error("This training session has no evalModel loaded.")}get evalOutputNames(){if(this.hasEvalModel)return this.handler.evalOutputNames;throw new Error("This training session has no evalModel loaded.")}static async create(e,r){let n=e.evalModel||"",a=e.optimizerModel||"",i=r||{},[s,o]=await Ro(i);if(s.createTrainingSessionHandler){let u=await s.createTrainingSessionHandler(e.checkpointState,e.trainModel,n,a,o);return new Yp(u,!!e.optimizerModel,!!e.evalModel)}else throw new Error(Xu)}typeNarrowingForRunStep(e,r,n,a,i){let s={},o={};if(typeof n!="object"||n===null||n instanceof vt||Array.isArray(n))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let u=!0;if(typeof a=="object"){if(a===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(a instanceof vt)throw new TypeError("'fetches' cannot be a Tensor");if(Array.isArray(a)){if(a.length===0)throw new TypeError("'fetches' cannot be an empty array.");u=!1;for(let l of a){if(typeof l!="string")throw new TypeError("'fetches' must be a string array or an object.");if(r.indexOf(l)===-1)throw new RangeError(`'fetches' contains invalid output name: ${l}.`);s[l]=null}if(typeof i=="object"&&i!==null)o=i;else if(typeof i<"u")throw new TypeError("'options' must be an object.")}else{let l=!1,p=Object.getOwnPropertyNames(a);for(let f of r)if(p.indexOf(f)!==-1){let m=a[f];(m===null||m instanceof vt)&&(l=!0,u=!1,s[f]=m)}if(l){if(typeof i=="object"&&i!==null)o=i;else if(typeof i<"u")throw new TypeError("'options' must be an object.")}else o=a}}else if(typeof a<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let l of e)if(typeof n[l]>"u")throw new Error(`input '${l}' is missing in 'feeds'.`);if(u)for(let l of r)s[l]=null;return[s,o]}convertHandlerReturnTypeToMapOfTensors(e){let r={};for(let n in e)if(Object.hasOwnProperty.call(e,n)){let a=e[n];a instanceof vt?r[n]=a:r[n]=new vt(a.type,a.data,a.dims)}return r}async lazyResetGrad(){await this.handler.lazyResetGrad()}async runTrainStep(e,r,n){let[a,i]=this.typeNarrowingForRunStep(this.trainingInputNames,this.trainingOutputNames,e,r,n),s=await this.handler.runTrainStep(e,a,i);return this.convertHandlerReturnTypeToMapOfTensors(s)}async runOptimizerStep(e){if(this.hasOptimizerModel)await this.handler.runOptimizerStep(e||{});else throw new Error("This TrainingSession has no OptimizerModel loaded.")}async runEvalStep(e,r,n){if(this.hasEvalModel){let[a,i]=this.typeNarrowingForRunStep(this.evalInputNames,this.evalOutputNames,e,r,n),s=await this.handler.runEvalStep(e,a,i);return this.convertHandlerReturnTypeToMapOfTensors(s)}else throw new Error("This TrainingSession has no EvalModel loaded.")}async getParametersSize(e=!0){return this.handler.getParametersSize(e)}async loadParametersBuffer(e,r=!0){let n=await this.getParametersSize(r);if(e.length!==4*n)throw new Error("Size of the buffer passed into loadParametersBuffer must match the number of parameters in the model. Please use getParametersSize method to check.");return this.handler.loadParametersBuffer(e,r)}async getContiguousParameters(e=!0){return this.handler.getContiguousParameters(e)}async release(){return this.handler.dispose()}}}),Lo,W_=Z(()=>{U_(),Lo=Kp}),Xp={};An(Xp,{InferenceSession:()=>Fo,TRACE:()=>ha,TRACE_FUNC_BEGIN:()=>nr,TRACE_FUNC_END:()=>Lt,Tensor:()=>vt,TrainingSession:()=>Lo,env:()=>Ue,registerBackend:()=>Yr});var ar=Z(()=>{T_(),A_(),B_(),No(),D_(),N_(),Hp(),F_(),L_(),W_()}),Uo={};An(Uo,{createReadStream:()=>Zp,readFile:()=>Qp,readFileSync:()=>Jp});var Qp,Jp,Zp,eh=Z(()=>{Qp=void 0,Jp=void 0,Zp=void 0}),Wo={};An(Wo,{join:()=>th});var th,rh=Z(()=>{th=void 0}),V_=In((t,e)=>{var r=(()=>{var n=typeof document<"u"?document.currentScript?.src:void 0;return typeof __filename<"u"&&(n||=__filename),function(a={}){var i=a,s,o,u=new Promise((h,_)=>{s=h,o=_});i.mountExternalData=(h,_)=>{(i.eb||(i.eb=new Map)).set(h,_)},i.unmountExternalData=()=>{delete i.eb};let l=()=>{let h=(x,E,O)=>(...F)=>{let q=Et,J=E?.();F=x(...F);let de=E?.();return J!==de&&(x=de,O(J),E=O=null),Et!=q?Ka():F},_=x=>async(...E)=>{try{if(i.cb)throw Error("Session already started");let O=i.cb={xb:E[0],errors:[]},F=await x(...E);if(i.cb!==O)throw Error("Session mismatch");i.kb?.flush();let q=O.errors;if(0<q.length){let J=await Promise.all(q);if(J=J.filter(de=>de),0<J.length)throw Error(J.join(`
`))}return F}finally{i.cb=null}};i._OrtCreateSession=h(i._OrtCreateSession,()=>i._OrtCreateSession,x=>i._OrtCreateSession=x),i._OrtRun=_(h(i._OrtRun,()=>i._OrtRun,x=>i._OrtRun=x)),i._OrtRunWithBinding=_(h(i._OrtRunWithBinding,()=>i._OrtRunWithBinding,x=>i._OrtRunWithBinding=x)),i._OrtBindInput=h(i._OrtBindInput,()=>i._OrtBindInput,x=>i._OrtBindInput=x),l=void 0};i.jsepInit=(h,_)=>{if(l?.(),h==="webgpu"){[i.kb,i.pb,i.tb,i.lb,i.sb,i.Ra,i.ub,i.wb,i.qb,i.rb,i.vb]=_;let x=i.kb;i.jsepRegisterBuffer=(E,O,F,q)=>x.registerBuffer(E,O,F,q),i.jsepGetBuffer=E=>x.getBuffer(E),i.jsepCreateDownloader=(E,O,F)=>x.createDownloader(E,O,F),i.jsepOnReleaseSession=E=>{x.onReleaseSession(E)},i.jsepOnRunStart=E=>x.onRunStart(E)}};var p=Object.assign({},i),f="./this.program",m=(h,_)=>{throw _},c=typeof window=="object",y=typeof importScripts=="function",w=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",v="",k,$,C;if(w){var T=(eh(),Qr(Uo)),A=(rh(),Qr(Wo));v=y?A.dirname(v)+"/":__dirname+"/",k=(h,_)=>(h=dt(h)?new URL(h):A.normalize(h),T.readFileSync(h,_?void 0:"utf8")),C=h=>(h=k(h,!0),h.buffer||(h=new Uint8Array(h)),h),$=(h,_,x,E=!0)=>{h=dt(h)?new URL(h):A.normalize(h),T.readFile(h,E?void 0:"utf8",(O,F)=>{O?x(O):_(E?F.buffer:F)})},!i.thisProgram&&1<process.argv.length&&(f=process.argv[1].replace(/\\/g,"/")),process.argv.slice(2),m=(h,_)=>{throw process.exitCode=h,_}}else(c||y)&&(y?v=self.location.href:typeof document<"u"&&document.currentScript&&(v=document.currentScript.src),n&&(v=n),v.startsWith("blob:")?v="":v=v.substr(0,v.replace(/[?#].*/,"").lastIndexOf("/")+1),k=h=>{var _=new XMLHttpRequest;return _.open("GET",h,!1),_.send(null),_.responseText},y&&(C=h=>{var _=new XMLHttpRequest;return _.open("GET",h,!1),_.responseType="arraybuffer",_.send(null),new Uint8Array(_.response)}),$=(h,_,x)=>{var E=new XMLHttpRequest;E.open("GET",h,!0),E.responseType="arraybuffer",E.onload=()=>{E.status==200||E.status==0&&E.response?_(E.response):x()},E.onerror=x,E.send(null)});var B=console.log.bind(console),R=console.error.bind(console);Object.assign(i,p),p=null;var D,K=!1,j,ie,te,oe,re,M,P,H,le,G,ne;function N(){var h=D.buffer;i.HEAP8=ie=new Int8Array(h),i.HEAP16=oe=new Int16Array(h),i.HEAPU8=te=new Uint8Array(h),i.HEAPU16=re=new Uint16Array(h),i.HEAP32=M=new Int32Array(h),i.HEAPU32=P=new Uint32Array(h),i.HEAPF32=H=new Float32Array(h),i.HEAPF64=ne=new Float64Array(h),i.HEAP64=le=new BigInt64Array(h),i.HEAPU64=G=new BigUint64Array(h)}var ae=[],fe=[],Ce=[],Be=0,Ke=null;function Qe(h){throw h="Aborted("+h+")",R(h),K=!0,j=1,h=new WebAssembly.RuntimeError(h+". Build with -sASSERTIONS for more info."),o(h),h}var We=h=>h.startsWith("data:application/octet-stream;base64,"),dt=h=>h.startsWith("file://"),Ve;if(Ve="ort-wasm-simd.wasm",!We(Ve)){var Ne=Ve;Ve=i.locateFile?i.locateFile(Ne,v):v+Ne}function ht(h){if(C)return C(h);throw"both async and sync fetching of the wasm failed"}function Vt(h){if(c||y){if(typeof fetch=="function"&&!dt(h))return fetch(h,{credentials:"same-origin"}).then(_=>{if(!_.ok)throw`failed to load wasm binary file at '${h}'`;return _.arrayBuffer()}).catch(()=>ht(h));if($)return new Promise((_,x)=>{$(h,E=>_(new Uint8Array(E)),x)})}return Promise.resolve().then(()=>ht(h))}function st(h,_,x){return Vt(h).then(E=>WebAssembly.instantiate(E,_)).then(x,E=>{R(`failed to asynchronously prepare wasm: ${E}`),Qe(E)})}function zr(h,_){var x=Ve;return typeof WebAssembly.instantiateStreaming!="function"||We(x)||dt(x)||w||typeof fetch!="function"?st(x,h,_):fetch(x,{credentials:"same-origin"}).then(E=>WebAssembly.instantiateStreaming(E,h).then(_,function(O){return R(`wasm streaming compile failed: ${O}`),R("falling back to ArrayBuffer instantiation"),st(x,h,_)}))}var hr={824920:(h,_,x,E)=>{if(typeof i>"u"||!i.eb)return 1;if(h=Ye(h>>>0),h.startsWith("./")&&(h=h.substring(2)),h=i.eb.get(h),!h)return 2;if(_>>>=0,x>>>=0,_+x>h.byteLength)return 3;try{return te.set(h.subarray(_,_+x),E>>>0>>>0),0}catch{return 4}},825421:()=>{i.qb()},825452:()=>{i.rb()},825481:()=>{i.vb()},825506:h=>i.pb(h),825539:h=>i.tb(h),825571:(h,_,x)=>{i.lb(h,_,x,!0)},825610:(h,_,x)=>{i.lb(h,_,x)},825643:h=>{i.Ra("Abs",h,void 0)},825694:h=>{i.Ra("Neg",h,void 0)},825745:h=>{i.Ra("Floor",h,void 0)},825798:h=>{i.Ra("Ceil",h,void 0)},825850:h=>{i.Ra("Reciprocal",h,void 0)},825908:h=>{i.Ra("Sqrt",h,void 0)},825960:h=>{i.Ra("Exp",h,void 0)},826011:h=>{i.Ra("Erf",h,void 0)},826062:h=>{i.Ra("Sigmoid",h,void 0)},826117:(h,_,x)=>{i.Ra("HardSigmoid",h,{alpha:_,beta:x})},826196:h=>{i.Ra("Log",h,void 0)},826247:h=>{i.Ra("Sin",h,void 0)},826298:h=>{i.Ra("Cos",h,void 0)},826349:h=>{i.Ra("Tan",h,void 0)},826400:h=>{i.Ra("Asin",h,void 0)},826452:h=>{i.Ra("Acos",h,void 0)},826504:h=>{i.Ra("Atan",h,void 0)},826556:h=>{i.Ra("Sinh",h,void 0)},826608:h=>{i.Ra("Cosh",h,void 0)},826660:h=>{i.Ra("Asinh",h,void 0)},826713:h=>{i.Ra("Acosh",h,void 0)},826766:h=>{i.Ra("Atanh",h,void 0)},826819:h=>{i.Ra("Tanh",h,void 0)},826871:h=>{i.Ra("Not",h,void 0)},826922:(h,_,x)=>{i.Ra("Clip",h,{min:_,max:x})},826991:h=>{i.Ra("Clip",h,void 0)},827043:(h,_)=>{i.Ra("Elu",h,{alpha:_})},827101:h=>{i.Ra("Relu",h,void 0)},827153:(h,_)=>{i.Ra("LeakyRelu",h,{alpha:_})},827217:(h,_)=>{i.Ra("ThresholdedRelu",h,{alpha:_})},827287:(h,_)=>{i.Ra("Cast",h,{to:_})},827345:h=>{i.Ra("Add",h,void 0)},827396:h=>{i.Ra("Sub",h,void 0)},827447:h=>{i.Ra("Mul",h,void 0)},827498:h=>{i.Ra("Div",h,void 0)},827549:h=>{i.Ra("Pow",h,void 0)},827600:h=>{i.Ra("Equal",h,void 0)},827653:h=>{i.Ra("Greater",h,void 0)},827708:h=>{i.Ra("GreaterOrEqual",h,void 0)},827770:h=>{i.Ra("Less",h,void 0)},827822:h=>{i.Ra("LessOrEqual",h,void 0)},827881:(h,_,x,E,O)=>{i.Ra("ReduceMean",h,{keepDims:!!_,noopWithEmptyAxes:!!x,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},828040:(h,_,x,E,O)=>{i.Ra("ReduceMax",h,{keepDims:!!_,noopWithEmptyAxes:!!x,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},828198:(h,_,x,E,O)=>{i.Ra("ReduceMin",h,{keepDims:!!_,noopWithEmptyAxes:!!x,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},828356:(h,_,x,E,O)=>{i.Ra("ReduceProd",h,{keepDims:!!_,noopWithEmptyAxes:!!x,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},828515:(h,_,x,E,O)=>{i.Ra("ReduceSum",h,{keepDims:!!_,noopWithEmptyAxes:!!x,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},828673:(h,_,x,E,O)=>{i.Ra("ReduceL1",h,{keepDims:!!_,noopWithEmptyAxes:!!x,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},828830:(h,_,x,E,O)=>{i.Ra("ReduceL2",h,{keepDims:!!_,noopWithEmptyAxes:!!x,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},828987:(h,_,x,E,O)=>{i.Ra("ReduceLogSum",h,{keepDims:!!_,noopWithEmptyAxes:!!x,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},829148:(h,_,x,E,O)=>{i.Ra("ReduceSumSquare",h,{keepDims:!!_,noopWithEmptyAxes:!!x,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},829312:(h,_,x,E,O)=>{i.Ra("ReduceLogSumExp",h,{keepDims:!!_,noopWithEmptyAxes:!!x,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},829476:h=>{i.Ra("Where",h,void 0)},829529:(h,_,x)=>{i.Ra("Transpose",h,{perm:_?Array.from(M.subarray(_>>>0,x>>>0)):[]})},829637:(h,_,x,E)=>{i.Ra("DepthToSpace",h,{blocksize:_,mode:Ye(x),format:E?"NHWC":"NCHW"})},829770:(h,_,x,E)=>{i.Ra("DepthToSpace",h,{blocksize:_,mode:Ye(x),format:E?"NHWC":"NCHW"})},829903:(h,_,x,E,O,F,q,J,de,se,me,ze,Fe,L,xe)=>{i.Ra("ConvTranspose",h,{format:de?"NHWC":"NCHW",autoPad:_,dilations:[x],group:E,kernelShape:[O],pads:[F,q],strides:[J],wIsConst:()=>!!ie[se>>>0],outputPadding:me?Array.from(M.subarray(me>>>0,ze>>>0)):[],outputShape:Fe?Array.from(M.subarray(Fe>>>0,L>>>0)):[],activation:Ye(xe)})},830304:(h,_,x,E,O,F,q,J,de,se,me,ze,Fe,L)=>{i.Ra("ConvTranspose",h,{format:J?"NHWC":"NCHW",autoPad:_,dilations:Array.from(M.subarray(x>>>0,(x>>>0)+2>>>0)),group:E,kernelShape:Array.from(M.subarray(O>>>0,(O>>>0)+2>>>0)),pads:Array.from(M.subarray(F>>>0,(F>>>0)+4>>>0)),strides:Array.from(M.subarray(q>>>0,(q>>>0)+2>>>0)),wIsConst:()=>!!ie[de>>>0],outputPadding:se?Array.from(M.subarray(se>>>0,me>>>0)):[],outputShape:ze?Array.from(M.subarray(ze>>>0,Fe>>>0)):[],activation:Ye(L)})},830869:(h,_,x,E,O,F,q,J,de,se,me,ze,Fe,L,xe)=>{i.Ra("ConvTranspose",h,{format:de?"NHWC":"NCHW",autoPad:_,dilations:[x],group:E,kernelShape:[O],pads:[F,q],strides:[J],wIsConst:()=>!!ie[se>>>0],outputPadding:me?Array.from(M.subarray(me>>>0,ze>>>0)):[],outputShape:Fe?Array.from(M.subarray(Fe>>>0,L>>>0)):[],activation:Ye(xe)})},831270:(h,_,x,E,O,F,q,J,de,se,me,ze,Fe,L)=>{i.Ra("ConvTranspose",h,{format:J?"NHWC":"NCHW",autoPad:_,dilations:Array.from(M.subarray(x>>>0,(x>>>0)+2>>>0)),group:E,kernelShape:Array.from(M.subarray(O>>>0,(O>>>0)+2>>>0)),pads:Array.from(M.subarray(F>>>0,(F>>>0)+4>>>0)),strides:Array.from(M.subarray(q>>>0,(q>>>0)+2>>>0)),wIsConst:()=>!!ie[de>>>0],outputPadding:se?Array.from(M.subarray(se>>>0,me>>>0)):[],outputShape:ze?Array.from(M.subarray(ze>>>0,Fe>>>0)):[],activation:Ye(L)})},831835:(h,_)=>{i.Ra("GlobalAveragePool",h,{format:_?"NHWC":"NCHW"})},831926:(h,_,x,E,O,F,q,J,de,se,me,ze,Fe,L,xe,Oe)=>{i.Ra("AveragePool",h,{format:Oe?"NHWC":"NCHW",auto_pad:_,ceil_mode:x,count_include_pad:E,storage_order:O,dilations:[F,q],kernel_shape:[J,de],pads:[se,me,ze,Fe],strides:[L,xe]})},832210:(h,_)=>{i.Ra("GlobalAveragePool",h,{format:_?"NHWC":"NCHW"})},832301:(h,_,x,E,O,F,q,J,de,se,me,ze,Fe,L,xe,Oe)=>{i.Ra("AveragePool",h,{format:Oe?"NHWC":"NCHW",auto_pad:_,ceil_mode:x,count_include_pad:E,storage_order:O,dilations:[F,q],kernel_shape:[J,de],pads:[se,me,ze,Fe],strides:[L,xe]})},832585:(h,_)=>{i.Ra("GlobalMaxPool",h,{format:_?"NHWC":"NCHW"})},832672:(h,_,x,E,O,F,q,J,de,se,me,ze,Fe,L,xe,Oe)=>{i.Ra("MaxPool",h,{format:Oe?"NHWC":"NCHW",auto_pad:_,ceil_mode:x,count_include_pad:E,storage_order:O,dilations:[F,q],kernel_shape:[J,de],pads:[se,me,ze,Fe],strides:[L,xe]})},832952:(h,_)=>{i.Ra("GlobalMaxPool",h,{format:_?"NHWC":"NCHW"})},833039:(h,_,x,E,O,F,q,J,de,se,me,ze,Fe,L,xe,Oe)=>{i.Ra("MaxPool",h,{format:Oe?"NHWC":"NCHW",auto_pad:_,ceil_mode:x,count_include_pad:E,storage_order:O,dilations:[F,q],kernel_shape:[J,de],pads:[se,me,ze,Fe],strides:[L,xe]})},833319:(h,_,x,E,O)=>{i.Ra("Gemm",h,{alpha:_,beta:x,transA:E,transB:O})},833423:h=>{i.Ra("MatMul",h,void 0)},833477:(h,_,x,E)=>{i.Ra("ArgMax",h,{keepDims:!!_,selectLastIndex:!!x,axis:E})},833585:(h,_,x,E)=>{i.Ra("ArgMin",h,{keepDims:!!_,selectLastIndex:!!x,axis:E})},833693:(h,_)=>{i.Ra("Softmax",h,{axis:_})},833756:(h,_)=>{i.Ra("Concat",h,{axis:_})},833816:(h,_,x,E,O)=>{i.Ra("Split",h,{axis:_,numOutputs:x,splitSizes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},833956:h=>{i.Ra("Expand",h,void 0)},834010:(h,_)=>{i.Ra("Gather",h,{axis:Number(_)})},834081:(h,_)=>{i.Ra("GatherElements",h,{axis:Number(_)})},834160:(h,_,x,E,O,F,q,J,de,se,me)=>{i.Ra("Resize",h,{antialias:_,axes:x?Array.from(M.subarray(x>>>0,E>>>0)):[],coordinateTransformMode:Ye(O),cubicCoeffA:F,excludeOutside:q,extrapolationValue:J,keepAspectRatioPolicy:Ye(de),mode:Ye(se),nearestMode:Ye(me)})},834506:(h,_,x,E,O,F,q)=>{i.Ra("Slice",h,{starts:_?Array.from(M.subarray(_>>>0,x>>>0)):[],ends:E?Array.from(M.subarray(E>>>0,O>>>0)):[],axes:F?Array.from(M.subarray(F>>>0,q>>>0)):[]})},834722:h=>{i.Ra("Tile",h,void 0)},834774:(h,_,x,E)=>{i.Ra("LayerNormalization",h,{axis:_,epsilon:x,simplified:!!E})},834885:(h,_,x)=>{i.Ra("InstanceNormalization",h,{epsilon:_,format:x?"NHWC":"NCHW"})},834999:(h,_,x)=>{i.Ra("InstanceNormalization",h,{epsilon:_,format:x?"NHWC":"NCHW"})},835113:h=>{i.Ra("Range",h,void 0)},835166:(h,_)=>{i.Ra("Einsum",h,{equation:Ye(_)})},835247:(h,_,x,E,O)=>{i.Ra("Pad",h,{mode:_,value:x,pads:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},835374:(h,_,x,E,O,F)=>{i.Ra("BatchNormalization",h,{epsilon:_,momentum:x,spatial:!!O,trainingMode:!!E,format:F?"NHWC":"NCHW"})},835543:(h,_,x,E,O,F)=>{i.Ra("BatchNormalization",h,{epsilon:_,momentum:x,spatial:!!O,trainingMode:!!E,format:F?"NHWC":"NCHW"})},835712:(h,_,x)=>{i.Ra("CumSum",h,{exclusive:Number(_),reverse:Number(x)})},835809:(h,_,x,E,O,F,q,J,de)=>{i.Ra("Attention",h,{numHeads:_,isUnidirectional:x,maskFilterValue:E,scale:O,doRotary:F,qkvHiddenSizes:q?Array.from(M.subarray(Number(J)>>>0,Number(J)+q>>>0)):[],pastPresentShareBuffer:!!de})},836081:h=>{i.Ra("BiasAdd",h,void 0)},836136:h=>{i.Ra("BiasSplitGelu",h,void 0)},836197:h=>{i.Ra("FastGelu",h,void 0)},836253:(h,_,x,E,O,F,q,J,de,se,me,ze,Fe)=>{i.Ra("Conv",h,{format:de?"NHWC":"NCHW",auto_pad:_,dilations:[x],group:E,kernel_shape:[O],pads:F?Array.from(M.subarray(F>>>0,q>>>0)):[],strides:[J],w_is_const:()=>!!ie[se>>>0],activation:Ye(me),activation_params:ze?Array.from(H.subarray(ze>>>0,Fe>>>0)):[]})},836623:(h,_,x,E,O,F,q,J,de,se,me,ze,Fe,L,xe,Oe)=>{i.Ra("Conv",h,{format:ze?"NHWC":"NCHW",auto_pad:_,dilations:[x,E],group:O,kernel_shape:[F,q],pads:J?Array.from(M.subarray(J>>>0,de>>>0)):[],strides:[se,me],w_is_const:()=>!!ie[Fe>>>0],activation:Ye(L),activation_params:xe?Array.from(H.subarray(xe>>>0,Oe>>>0)):[]})},837014:h=>{i.Ra("Gelu",h,void 0)},837066:(h,_,x,E,O,F)=>{i.Ra("MatMulNBits",h,{k:_,n:x,accuracyLevel:E,bits:O,blockSize:F})},837193:(h,_,x,E,O,F)=>{i.Ra("MultiHeadAttention",h,{numHeads:_,isUnidirectional:x,maskFilterValue:E,scale:O,doRotary:F})},837352:(h,_,x,E,O)=>{i.Ra("RotaryEmbedding",h,{interleaved:!!_,numHeads:x,rotaryEmbeddingDim:E,scale:O})},837491:(h,_,x)=>{i.Ra("SkipLayerNormalization",h,{epsilon:_,simplified:!!x})},837593:(h,_,x)=>{i.Ra("SkipLayerNormalization",h,{epsilon:_,simplified:!!x})},837695:(h,_,x,E)=>{i.Ra("LayerNormalization",h,{axis:_,epsilon:x,simplified:!!E})},837806:h=>{i.ub(h)},837840:(h,_)=>i.wb(h,_,i.cb.xb,i.cb.errors)};function ir(h){this.name="ExitStatus",this.message=`Program terminated with exit(${h})`,this.status=h}class ls{constructor(_){this.hb=_-24}}var fr=0,xr=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,Pn=(h,_,x)=>{_>>>=0;var E=_+x;for(x=_;h[x]&&!(x>=E);)++x;if(16<x-_&&h.buffer&&xr)return xr.decode(h.subarray(_,x));for(E="";_<x;){var O=h[_++];if(O&128){var F=h[_++]&63;if((O&224)==192)E+=String.fromCharCode((O&31)<<6|F);else{var q=h[_++]&63;O=(O&240)==224?(O&15)<<12|F<<6|q:(O&7)<<18|F<<12|q<<6|h[_++]&63,65536>O?E+=String.fromCharCode(O):(O-=65536,E+=String.fromCharCode(55296|O>>10,56320|O&1023))}}else E+=String.fromCharCode(O)}return E},Ye=(h,_)=>(h>>>=0)?Pn(te,h,_):"",on=h=>{for(var _=0,x=0;x<h.length;++x){var E=h.charCodeAt(x);127>=E?_++:2047>=E?_+=2:55296<=E&&57343>=E?(_+=4,++x):_+=3}return _},Bt=(h,_,x,E)=>{if(x>>>=0,!(0<E))return 0;var O=x;E=x+E-1;for(var F=0;F<h.length;++F){var q=h.charCodeAt(F);if(55296<=q&&57343>=q){var J=h.charCodeAt(++F);q=65536+((q&1023)<<10)|J&1023}if(127>=q){if(x>=E)break;_[x++>>>0]=q}else{if(2047>=q){if(x+1>=E)break;_[x++>>>0]=192|q>>6}else{if(65535>=q){if(x+2>=E)break;_[x++>>>0]=224|q>>12}else{if(x+3>=E)break;_[x++>>>0]=240|q>>18,_[x++>>>0]=128|q>>12&63}_[x++>>>0]=128|q>>6&63}_[x++>>>0]=128|q&63}}return _[x>>>0]=0,x-O},Sr,At=h=>{for(var _="";te[h>>>0];)_+=Sr[te[h++>>>0]];return _},Rn={},un={},sr;function Bn(h,_,x={}){var E=_.name;if(!h)throw new sr(`type "${E}" must have a positive integer typeid pointer`);if(un.hasOwnProperty(h)){if(x.nb)return;throw new sr(`Cannot register type '${E}' twice`)}un[h]=_,Rn.hasOwnProperty(h)&&(_=Rn[h],delete Rn[h],_.forEach(O=>O()))}function St(h,_,x={}){if(!("argPackAdvance"in _))throw new TypeError("registerType registeredInstance requires argPackAdvance");return Bn(h,_,x)}var ln=(h,_,x)=>{switch(_){case 1:return x?E=>ie[E>>>0]:E=>te[E>>>0];case 2:return x?E=>oe[E>>>1>>>0]:E=>re[E>>>1>>>0];case 4:return x?E=>M[E>>>2>>>0]:E=>P[E>>>2>>>0];case 8:return x?E=>le[E>>>3]:E=>G[E>>>3];default:throw new TypeError(`invalid integer width (${_}): ${h}`)}},dn=[],ot=[];function Dn(h){h>>>=0,9<h&&--ot[h+1]===0&&(ot[h]=void 0,dn.push(h))}var Ie=h=>{if(!h)throw new sr("Cannot use deleted val. handle = "+h);return ot[h]},kt=h=>{switch(h){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let _=dn.pop()||ot.length;return ot[_]=h,ot[_+1]=1,_}};function Pr(h){return this.fromWireType(P[h>>>2>>>0])}var ds={name:"emscripten::val",fromWireType:h=>{var _=Ie(h);return Dn(h),_},toWireType:(h,_)=>kt(_),argPackAdvance:8,readValueFromPointer:Pr,bb:null},La=(h,_)=>{switch(_){case 4:return function(x){return this.fromWireType(H[x>>>2>>>0])};case 8:return function(x){return this.fromWireType(ne[x>>>3>>>0])};default:throw new TypeError(`invalid float width (${_}): ${h}`)}},Ua=typeof TextDecoder<"u"?new TextDecoder("utf-16le"):void 0,be=(h,_)=>{for(var x=h>>1,E=x+_/2;!(x>=E)&&re[x>>>0];)++x;if(x<<=1,32<x-h&&Ua)return Ua.decode(te.subarray(h>>>0,x>>>0));for(x="",E=0;!(E>=_/2);++E){var O=oe[h+2*E>>>1>>>0];if(O==0)break;x+=String.fromCharCode(O)}return x},cn=(h,_,x)=>{if(x??=2147483647,2>x)return 0;x-=2;var E=_;x=x<2*h.length?x/2:h.length;for(var O=0;O<x;++O)oe[_>>>1>>>0]=h.charCodeAt(O),_+=2;return oe[_>>>1>>>0]=0,_-E},Wa=h=>2*h.length,cs=(h,_)=>{for(var x=0,E="";!(x>=_/4);){var O=M[h+4*x>>>2>>>0];if(O==0)break;++x,65536<=O?(O-=65536,E+=String.fromCharCode(55296|O>>10,56320|O&1023)):E+=String.fromCharCode(O)}return E},Va=(h,_,x)=>{if(_>>>=0,x??=2147483647,4>x)return 0;var E=_;x=E+x-4;for(var O=0;O<h.length;++O){var F=h.charCodeAt(O);if(55296<=F&&57343>=F){var q=h.charCodeAt(++O);F=65536+((F&1023)<<10)|q&1023}if(M[_>>>2>>>0]=F,_+=4,_+4>x)break}return M[_>>>2>>>0]=0,_-E},Ga=h=>{for(var _=0,x=0;x<h.length;++x){var E=h.charCodeAt(x);55296<=E&&57343>=E&&++x,_+=4}return _},pn=(h,_)=>{var x=un[h];if(x===void 0)throw h=qn(h),x=At(h),_t(h),new sr(`${_} has unknown type ${x}`);return x},Nn=(h,_,x)=>{var E=[];return h=h.toWireType(E,x),E.length&&(P[_>>>2>>>0]=kt(E)),h},Rr=h=>{try{h()}catch(_){Qe(_)}},nt=h=>{if(!K)try{h();try{j=j=h=j,i.onExit?.(h),K=!0,m(h,new ir(h))}catch(_){_ instanceof ir||_=="unwind"||m(1,_)}}catch(_){_ instanceof ir||_=="unwind"||m(1,_)}};function Ha(){var h=we,_={};for(let[x,E]of Object.entries(h))_[x]=typeof E=="function"?(...O)=>{Ht.push(x);try{return E(...O)}finally{K||(Ht.pop(),Et&&Gt===1&&Ht.length===0&&(Gt=0,Rr(ti),typeof Fibers<"u"&&Fibers.Db()))}}:E;return _}var Gt=0,Et=null,Fn=0,Ht=[],Ln={},Un={},ja=0,hn=null,qa=[];function Ka(){return new Promise((h,_)=>{hn={resolve:h,reject:_}})}function Ya(){var h=kr(65548),_=h+12;P[h>>>2>>>0]=_,P[h+4>>>2>>>0]=_+65536,_=Ht[0];var x=Ln[_];return x===void 0&&(x=ja++,Ln[_]=x,Un[x]=_),M[h+8>>>2>>>0]=x,h}function Xa(h){if(!K){if(Gt===0){var _=!1,x=!1;h((E=0)=>{if(!K&&(Fn=E,_=!0,x)){Gt=2,Rr(()=>qt(Et)),typeof Browser<"u"&&Browser.ib.mb&&Browser.ib.resume(),E=!1;try{var O=(0,we[Un[M[Et+8>>>2>>>0]]])()}catch(J){O=J,E=!0}var F=!1;if(!Et){var q=hn;q&&(hn=null,(E?q.reject:q.resolve)(O),F=!0)}if(E&&!F)throw O}}),x=!0,_||(Gt=1,Et=Ya(),typeof Browser<"u"&&Browser.ib.mb&&Browser.ib.pause(),Rr(()=>Nr(Et)))}else Gt===2?(Gt=0,Rr(zt),_t(Et),Et=null,qa.forEach(nt)):Qe(`invalid state: ${Gt}`);return Fn}}function Wn(h){return Xa(_=>{h().then(_)})}var Br=[],Qa={},fn=h=>{var _=Qa[h];return _===void 0?At(h):_},Mt=()=>typeof globalThis=="object"?globalThis:Function("return this")(),Vn=h=>{var _=Br.length;return Br.push(h),_},Gn=(h,_)=>{for(var x=Array(h),E=0;E<h;++E)x[E]=pn(P[_+4*E>>>2>>>0],"parameter "+E);return x},jt=(h,_)=>Object.defineProperty(_,"name",{value:h});function ps(h){var _=Function;if(!(_ instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof _} which is not a function`);var x=jt(_.name||"unknownFunctionName",function(){});return x.prototype=_.prototype,x=new x,h=_.apply(x,h),h instanceof Object?h:x}var ft=h=>h%4===0&&(h%100!==0||h%400===0),Hn=[0,31,60,91,121,152,182,213,244,274,305,335],mn=[0,31,59,90,120,151,181,212,243,273,304,334],Ot=[],gn=(h,_)=>{Ot.length=0;for(var x;x=te[h++>>>0];){var E=x!=105;E&=x!=112,_+=E&&_%8?4:0,Ot.push(x==112?P[_>>>2>>>0]:x==106?le[_>>>3]:x==105?M[_>>>2>>>0]:ne[_>>>3>>>0]),_+=E?8:4}return Ot},ut={},Ct=()=>{if(!Dr){var h={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(typeof navigator=="object"&&navigator.languages&&navigator.languages[0]||"C").replace("-","_")+".UTF-8",_:f||"./this.program"},_;for(_ in ut)ut[_]===void 0?delete h[_]:h[_]=ut[_];var x=[];for(_ in h)x.push(`${_}=${h[_]}`);Dr=x}return Dr},Dr,hs=[null,[],[]],Ja=[31,29,31,30,31,30,31,31,30,31,30,31],jn=[31,28,31,30,31,30,31,31,30,31,30,31];function fs(h){var _=Array(on(h)+1);return Bt(h,_,0,_.length),_}function Za(h,_,x,E){function O(L,xe,Oe){for(L=typeof L=="number"?L.toString():L||"";L.length<xe;)L=Oe[0]+L;return L}function F(L,xe){return O(L,xe,"0")}function q(L,xe){function Oe(_r){return 0>_r?-1:0<_r?1:0}var gr;return(gr=Oe(L.getFullYear()-xe.getFullYear()))===0&&(gr=Oe(L.getMonth()-xe.getMonth()))===0&&(gr=Oe(L.getDate()-xe.getDate())),gr}function J(L){switch(L.getDay()){case 0:return new Date(L.getFullYear()-1,11,29);case 1:return L;case 2:return new Date(L.getFullYear(),0,3);case 3:return new Date(L.getFullYear(),0,2);case 4:return new Date(L.getFullYear(),0,1);case 5:return new Date(L.getFullYear()-1,11,31);case 6:return new Date(L.getFullYear()-1,11,30)}}function de(L){var xe=L.$a;for(L=new Date(new Date(L.ab+1900,0,1).getTime());0<xe;){var Oe=L.getMonth(),gr=(ft(L.getFullYear())?Ja:jn)[Oe];if(xe>gr-L.getDate())xe-=gr-L.getDate()+1,L.setDate(1),11>Oe?L.setMonth(Oe+1):(L.setMonth(0),L.setFullYear(L.getFullYear()+1));else{L.setDate(L.getDate()+xe);break}}return Oe=new Date(L.getFullYear()+1,0,4),xe=J(new Date(L.getFullYear(),0,4)),Oe=J(Oe),0>=q(xe,L)?0>=q(Oe,L)?L.getFullYear()+1:L.getFullYear():L.getFullYear()-1}h>>>=0,_>>>=0,x>>>=0,E>>>=0;var se=P[E+40>>>2>>>0];E={Ab:M[E>>>2>>>0],zb:M[E+4>>>2>>>0],fb:M[E+8>>>2>>>0],jb:M[E+12>>>2>>>0],gb:M[E+16>>>2>>>0],ab:M[E+20>>>2>>>0],Va:M[E+24>>>2>>>0],$a:M[E+28>>>2>>>0],Cb:M[E+32>>>2>>>0],yb:M[E+36>>>2>>>0],Bb:se?Ye(se):""},x=Ye(x),se={"%c":"%a %b %d %H:%M:%S %Y","%D":"%m/%d/%y","%F":"%Y-%m-%d","%h":"%b","%r":"%I:%M:%S %p","%R":"%H:%M","%T":"%H:%M:%S","%x":"%m/%d/%y","%X":"%H:%M:%S","%Ec":"%c","%EC":"%C","%Ex":"%m/%d/%y","%EX":"%H:%M:%S","%Ey":"%y","%EY":"%Y","%Od":"%d","%Oe":"%e","%OH":"%H","%OI":"%I","%Om":"%m","%OM":"%M","%OS":"%S","%Ou":"%u","%OU":"%U","%OV":"%V","%Ow":"%w","%OW":"%W","%Oy":"%y"};for(var me in se)x=x.replace(new RegExp(me,"g"),se[me]);var ze="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),Fe="January February March April May June July August September October November December".split(" ");se={"%a":L=>ze[L.Va].substring(0,3),"%A":L=>ze[L.Va],"%b":L=>Fe[L.gb].substring(0,3),"%B":L=>Fe[L.gb],"%C":L=>F((L.ab+1900)/100|0,2),"%d":L=>F(L.jb,2),"%e":L=>O(L.jb,2," "),"%g":L=>de(L).toString().substring(2),"%G":de,"%H":L=>F(L.fb,2),"%I":L=>(L=L.fb,L==0?L=12:12<L&&(L-=12),F(L,2)),"%j":L=>{for(var xe=0,Oe=0;Oe<=L.gb-1;xe+=(ft(L.ab+1900)?Ja:jn)[Oe++]);return F(L.jb+xe,3)},"%m":L=>F(L.gb+1,2),"%M":L=>F(L.zb,2),"%n":()=>`
`,"%p":L=>0<=L.fb&&12>L.fb?"AM":"PM","%S":L=>F(L.Ab,2),"%t":()=>"	","%u":L=>L.Va||7,"%U":L=>F(Math.floor((L.$a+7-L.Va)/7),2),"%V":L=>{var xe=Math.floor((L.$a+7-(L.Va+6)%7)/7);if(2>=(L.Va+371-L.$a-2)%7&&xe++,xe)xe==53&&(Oe=(L.Va+371-L.$a)%7,Oe==4||Oe==3&&ft(L.ab)||(xe=1));else{xe=52;var Oe=(L.Va+7-L.$a-1)%7;(Oe==4||Oe==5&&ft(L.ab%400-1))&&xe++}return F(xe,2)},"%w":L=>L.Va,"%W":L=>F(Math.floor((L.$a+7-(L.Va+6)%7)/7),2),"%y":L=>(L.ab+1900).toString().substring(2),"%Y":L=>L.ab+1900,"%z":L=>{L=L.yb;var xe=0<=L;return L=Math.abs(L)/60,(xe?"+":"-")+("0000"+(L/60*100+L%60)).slice(-4)},"%Z":L=>L.Bb,"%%":()=>"%"},x=x.replace(/%%/g,"\0\0");for(me in se)x.includes(me)&&(x=x.replace(new RegExp(me,"g"),se[me](E)));return x=x.replace(/\0\0/g,"%"),me=fs(x),me.length>_?0:(ie.set(me,h>>>0),me.length-1)}for(var ei=Array(256),_n=0;256>_n;++_n)ei[_n]=String.fromCharCode(_n);Sr=ei,sr=i.BindingError=class extends Error{constructor(h){super(h),this.name="BindingError"}},i.InternalError=class extends Error{constructor(h){super(h),this.name="InternalError"}},ot.push(0,1,void 0,1,null,1,!0,1,!1,1),i.count_emval_handles=()=>ot.length/2-5-dn.length;var ms={ia:function(h,_,x){return Wn(async()=>{await i.sb(h,_,x)})},a:function(h,_,x){h>>>=0;var E=new ls(h);throw P[E.hb+16>>>2>>>0]=0,P[E.hb+4>>>2>>>0]=_>>>0,P[E.hb+8>>>2>>>0]=x>>>0,fr=h,fr},y:function(){return 0},ea:function(){},R:function(){},T:function(){},ga:function(){return 0},ca:function(){},Z:function(){},ba:function(){},G:function(){},S:function(){},P:function(){},da:function(){},Q:function(){},C:function(h,_,x){_=At(_>>>0),St(h>>>0,{name:_,fromWireType:E=>E,toWireType:function(E,O){if(typeof O!="bigint"&&typeof O!="number")throw O===null?O="null":(E=typeof O,O=E==="object"||E==="array"||E==="function"?O.toString():""+O),new TypeError(`Cannot convert "${O}" to ${this.name}`);return typeof O=="number"&&(O=BigInt(O)),O},argPackAdvance:8,readValueFromPointer:ln(_,x>>>0,_.indexOf("u")==-1),bb:null})},K:function(h,_,x,E){_=At(_>>>0),St(h>>>0,{name:_,fromWireType:function(O){return!!O},toWireType:function(O,F){return F?x:E},argPackAdvance:8,readValueFromPointer:function(O){return this.fromWireType(te[O>>>0])},bb:null})},J:function(h){return St(h>>>0,ds)},B:function(h,_,x){_=At(_>>>0),St(h>>>0,{name:_,fromWireType:E=>E,toWireType:(E,O)=>O,argPackAdvance:8,readValueFromPointer:La(_,x>>>0),bb:null})},s:function(h,_,x,E,O){if(h>>>=0,x>>>=0,_=At(_>>>0),O===-1&&(O=4294967295),O=J=>J,E===0){var F=32-8*x;O=J=>J<<F>>>F}var q=_.includes("unsigned")?function(J,de){return de>>>0}:function(J,de){return de};St(h,{name:_,fromWireType:O,toWireType:q,argPackAdvance:8,readValueFromPointer:ln(_,x,E!==0),bb:null})},o:function(h,_,x){function E(F){return new O(ie.buffer,P[F+4>>>2>>>0],P[F>>>2>>>0])}var O=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][_];x=At(x>>>0),St(h>>>0,{name:x,fromWireType:E,argPackAdvance:8,readValueFromPointer:E},{nb:!0})},D:function(h,_){_=At(_>>>0);var x=_==="std::string";St(h>>>0,{name:_,fromWireType:function(E){var O=P[E>>>2>>>0],F=E+4;if(x)for(var q=F,J=0;J<=O;++J){var de=F+J;if(J==O||te[de>>>0]==0){if(q=Ye(q,de-q),se===void 0)var se=q;else se+="\0",se+=q;q=de+1}}else{for(se=Array(O),J=0;J<O;++J)se[J]=String.fromCharCode(te[F+J>>>0]);se=se.join("")}return _t(E),se},toWireType:function(E,O){O instanceof ArrayBuffer&&(O=new Uint8Array(O));var F=typeof O=="string";if(!(F||O instanceof Uint8Array||O instanceof Uint8ClampedArray||O instanceof Int8Array))throw new sr("Cannot pass non-string to std::string");var q=x&&F?on(O):O.length,J=kr(4+q+1),de=J+4;if(P[J>>>2>>>0]=q,x&&F)Bt(O,te,de,q+1);else if(F)for(F=0;F<q;++F){var se=O.charCodeAt(F);if(255<se)throw _t(de),new sr("String has UTF-16 code units that do not fit in 8 bits");te[de+F>>>0]=se}else for(F=0;F<q;++F)te[de+F>>>0]=O[F];return E!==null&&E.push(_t,J),J},argPackAdvance:8,readValueFromPointer:Pr,bb(E){_t(E)}})},x:function(h,_,x){if(_>>>=0,x>>>=0,x=At(x),_===2)var E=be,O=cn,F=Wa,q=J=>re[J>>>1>>>0];else _===4&&(E=cs,O=Va,F=Ga,q=J=>P[J>>>2>>>0]);St(h>>>0,{name:x,fromWireType:J=>{for(var de=P[J>>>2>>>0],se,me=J+4,ze=0;ze<=de;++ze){var Fe=J+4+ze*_;(ze==de||q(Fe)==0)&&(me=E(me,Fe-me),se===void 0?se=me:(se+="\0",se+=me),me=Fe+_)}return _t(J),se},toWireType:(J,de)=>{if(typeof de!="string")throw new sr(`Cannot pass non-string to C++ string type ${x}`);var se=F(de),me=kr(4+se+_);return P[me>>>2>>>0]=se/_,O(de,me+4,se+_),J!==null&&J.push(_t,me),me},argPackAdvance:8,readValueFromPointer:Pr,bb(J){_t(J)}})},L:function(h,_){_=At(_>>>0),St(h>>>0,{ob:!0,name:_,argPackAdvance:0,fromWireType:()=>{},toWireType:()=>{}})},ha:()=>1,u:function(h,_,x){return _>>>=0,x>>>=0,h=Ie(h>>>0),_=pn(_,"emval::as"),Nn(_,x,h)},w:function(h){return h>>>=0,Wn(()=>(h=Ie(h),h.then(kt)))},n:function(h,_,x,E){return x>>>=0,E>>>=0,h=Br[h>>>0],_=Ie(_>>>0),h(null,_,x,E)},j:function(h,_,x,E,O){return x>>>=0,E>>>=0,O>>>=0,h=Br[h>>>0],_=Ie(_>>>0),x=fn(x),h(_,_[x],E,O)},b:Dn,A:function(h,_){return _>>>=0,h=Ie(h>>>0),_=Ie(_),h==_},m:function(h){return h>>>=0,h===0?kt(Mt()):(h=fn(h),kt(Mt()[h]))},i:function(h,_,x){_=Gn(h,_>>>0);var E=_.shift();h--;var O=`return function (obj, func, destructorsRef, args) {
`,F=0,q=[];x===0&&q.push("obj");for(var J=["retType"],de=[E],se=0;se<h;++se)q.push("arg"+se),J.push("argType"+se),de.push(_[se]),O+=`  var arg${se} = argType${se}.readValueFromPointer(args${F?"+"+F:""});
`,F+=_[se].argPackAdvance;return O+=`  var rv = ${x===1?"new func":"func.call"}(${q.join(", ")});
`,E.ob||(J.push("emval_returnValue"),de.push(Nn),O+=`  return emval_returnValue(retType, destructorsRef, rv);
`),J.push(O+`};
`),h=ps(J)(...de),x=`methodCaller<(${_.map(me=>me.name).join(", ")}) => ${E.name}>`,Vn(jt(x,h))},r:function(h,_){return _>>>=0,h=Ie(h>>>0),_=Ie(_),kt(h[_])},e:function(h){h>>>=0,9<h&&(ot[h+1]+=1)},t:function(){return kt([])},l:function(h){h=Ie(h>>>0);for(var _=Array(h.length),x=0;x<h.length;x++)_[x]=h[x];return kt(_)},f:function(h){return kt(fn(h>>>0))},k:function(){return kt({})},h:function(h){h>>>=0;for(var _=Ie(h);_.length;){var x=_.pop();_.pop()(x)}Dn(h)},g:function(h,_,x){_>>>=0,x>>>=0,h=Ie(h>>>0),_=Ie(_),x=Ie(x),h[_]=x},c:function(h,_){return _>>>=0,h=pn(h>>>0,"_emval_take_value"),h=h.readValueFromPointer(_),kt(h)},W:function(h,_){h=-9007199254740992>h||9007199254740992<h?NaN:Number(h),_>>>=0,h=new Date(1e3*h),M[_>>>2>>>0]=h.getUTCSeconds(),M[_+4>>>2>>>0]=h.getUTCMinutes(),M[_+8>>>2>>>0]=h.getUTCHours(),M[_+12>>>2>>>0]=h.getUTCDate(),M[_+16>>>2>>>0]=h.getUTCMonth(),M[_+20>>>2>>>0]=h.getUTCFullYear()-1900,M[_+24>>>2>>>0]=h.getUTCDay(),M[_+28>>>2>>>0]=(h.getTime()-Date.UTC(h.getUTCFullYear(),0,1,0,0,0,0))/864e5|0},X:function(h,_){h=-9007199254740992>h||9007199254740992<h?NaN:Number(h),_>>>=0,h=new Date(1e3*h),M[_>>>2>>>0]=h.getSeconds(),M[_+4>>>2>>>0]=h.getMinutes(),M[_+8>>>2>>>0]=h.getHours(),M[_+12>>>2>>>0]=h.getDate(),M[_+16>>>2>>>0]=h.getMonth(),M[_+20>>>2>>>0]=h.getFullYear()-1900,M[_+24>>>2>>>0]=h.getDay(),M[_+28>>>2>>>0]=(ft(h.getFullYear())?Hn:mn)[h.getMonth()]+h.getDate()-1|0,M[_+36>>>2>>>0]=-(60*h.getTimezoneOffset());var x=new Date(h.getFullYear(),6,1).getTimezoneOffset(),E=new Date(h.getFullYear(),0,1).getTimezoneOffset();M[_+32>>>2>>>0]=(x!=E&&h.getTimezoneOffset()==Math.min(E,x))|0},Y:function(h){h>>>=0;var _=new Date(M[h+20>>>2>>>0]+1900,M[h+16>>>2>>>0],M[h+12>>>2>>>0],M[h+8>>>2>>>0],M[h+4>>>2>>>0],M[h>>>2>>>0],0),x=M[h+32>>>2>>>0],E=_.getTimezoneOffset(),O=new Date(_.getFullYear(),6,1).getTimezoneOffset(),F=new Date(_.getFullYear(),0,1).getTimezoneOffset(),q=Math.min(F,O);return 0>x?M[h+32>>>2>>>0]=+(O!=F&&q==E):0<x!=(q==E)&&(O=Math.max(F,O),_.setTime(_.getTime()+6e4*((0<x?q:O)-E))),M[h+24>>>2>>>0]=_.getDay(),M[h+28>>>2>>>0]=(ft(_.getFullYear())?Hn:mn)[_.getMonth()]+_.getDate()-1|0,M[h>>>2>>>0]=_.getSeconds(),M[h+4>>>2>>>0]=_.getMinutes(),M[h+8>>>2>>>0]=_.getHours(),M[h+12>>>2>>>0]=_.getDate(),M[h+16>>>2>>>0]=_.getMonth(),M[h+20>>>2>>>0]=_.getYear(),h=_.getTime(),BigInt(isNaN(h)?-1:h/1e3)},U:function(){return-52},V:function(){},N:function(h,_,x,E){x>>>=0,E>>>=0;var O=new Date().getFullYear(),F=new Date(O,0,1),q=new Date(O,6,1);O=F.getTimezoneOffset();var J=q.getTimezoneOffset();P[h>>>0>>>2>>>0]=60*Math.max(O,J),M[_>>>0>>>2>>>0]=+(O!=J),h=de=>de.toLocaleTimeString(void 0,{hour12:!1,timeZoneName:"short"}).split(" ")[1],F=h(F),q=h(q),J<O?(Bt(F,te,x,17),Bt(q,te,E,17)):(Bt(F,te,E,17),Bt(q,te,x,17))},v:()=>{Qe("")},d:function(h,_,x){return h>>>=0,_=gn(_>>>0,x>>>0),hr[h](..._)},I:function(h,_,x){return h>>>=0,_=gn(_>>>0,x>>>0),hr[h](..._)},H:()=>Date.now(),O:function(){return 4294901760},q:()=>performance.now(),M:function(h){h>>>=0;var _=te.length;if(4294901760<h)return!1;for(var x=1;4>=x;x*=2){var E=_*(1+.2/x);E=Math.min(E,h+100663296);var O=Math;E=Math.max(h,E);e:{O=(O.min.call(O,4294901760,E+(65536-E%65536)%65536)-D.buffer.byteLength+65535)/65536;try{D.grow(O),N();var F=1;break e}catch{}F=void 0}if(F)return!0}return!1},$:function(h,_){h>>>=0,_>>>=0;var x=0;return Ct().forEach((E,O)=>{var F=_+x;for(O=P[h+4*O>>>2>>>0]=F,F=0;F<E.length;++F)ie[O++>>>0]=E.charCodeAt(F);ie[O>>>0]=0,x+=E.length+1}),0},aa:function(h,_){h>>>=0,_>>>=0;var x=Ct();P[h>>>2>>>0]=x.length;var E=0;return x.forEach(O=>E+=O.length+1),P[_>>>2>>>0]=E,0},z:()=>52,F:function(){return 52},_:function(){return 70},E:function(h,_,x,E){_>>>=0,x>>>=0,E>>>=0;for(var O=0,F=0;F<x;F++){var q=P[_>>>2>>>0],J=P[_+4>>>2>>>0];_+=8;for(var de=0;de<J;de++){var se=te[q+de>>>0],me=hs[h];se===0||se===10?((h===1?B:R)(Pn(me,0)),me.length=0):me.push(se)}O+=J}return P[E>>>2>>>0]=O,0},fa:Za,p:function(h,_,x,E){return Za(h>>>0,_>>>0,x>>>0,E>>>0)}},we=function(){function h(x){return we=x.exports,we=Ha(),we=ri(),D=we.ja,N(),fe.unshift(we.ka),Be--,Be==0&&Ke&&(x=Ke,Ke=null,x()),we}var _={a:ms};if(Be++,i.instantiateWasm)try{return i.instantiateWasm(_,h)}catch(x){R(`Module.instantiateWasm callback failed with error: ${x}`),o(x)}return zr(_,function(x){h(x.instance)}).catch(o),{}}(),qn=h=>(qn=we.la)(h);i._OrtInit=(h,_)=>(i._OrtInit=we.ma)(h,_),i._OrtGetLastError=(h,_)=>(i._OrtGetLastError=we.na)(h,_),i._OrtCreateSessionOptions=(h,_,x,E,O,F,q,J,de,se)=>(i._OrtCreateSessionOptions=we.oa)(h,_,x,E,O,F,q,J,de,se),i._OrtAppendExecutionProvider=(h,_)=>(i._OrtAppendExecutionProvider=we.pa)(h,_),i._OrtAddFreeDimensionOverride=(h,_,x)=>(i._OrtAddFreeDimensionOverride=we.qa)(h,_,x),i._OrtAddSessionConfigEntry=(h,_,x)=>(i._OrtAddSessionConfigEntry=we.ra)(h,_,x),i._OrtReleaseSessionOptions=h=>(i._OrtReleaseSessionOptions=we.sa)(h),i._OrtCreateSession=(h,_,x)=>(i._OrtCreateSession=we.ta)(h,_,x),i._OrtReleaseSession=h=>(i._OrtReleaseSession=we.ua)(h),i._OrtGetInputOutputCount=(h,_,x)=>(i._OrtGetInputOutputCount=we.va)(h,_,x),i._OrtGetInputName=(h,_)=>(i._OrtGetInputName=we.wa)(h,_),i._OrtGetOutputName=(h,_)=>(i._OrtGetOutputName=we.xa)(h,_),i._OrtFree=h=>(i._OrtFree=we.ya)(h),i._OrtCreateTensor=(h,_,x,E,O,F)=>(i._OrtCreateTensor=we.za)(h,_,x,E,O,F),i._OrtGetTensorData=(h,_,x,E,O)=>(i._OrtGetTensorData=we.Aa)(h,_,x,E,O),i._OrtReleaseTensor=h=>(i._OrtReleaseTensor=we.Ba)(h),i._OrtCreateRunOptions=(h,_,x,E)=>(i._OrtCreateRunOptions=we.Ca)(h,_,x,E),i._OrtAddRunConfigEntry=(h,_,x)=>(i._OrtAddRunConfigEntry=we.Da)(h,_,x),i._OrtReleaseRunOptions=h=>(i._OrtReleaseRunOptions=we.Ea)(h),i._OrtCreateBinding=h=>(i._OrtCreateBinding=we.Fa)(h),i._OrtBindInput=(h,_,x)=>(i._OrtBindInput=we.Ga)(h,_,x),i._OrtBindOutput=(h,_,x,E)=>(i._OrtBindOutput=we.Ha)(h,_,x,E),i._OrtClearBoundOutputs=h=>(i._OrtClearBoundOutputs=we.Ia)(h),i._OrtReleaseBinding=h=>(i._OrtReleaseBinding=we.Ja)(h),i._OrtRunWithBinding=(h,_,x,E,O)=>(i._OrtRunWithBinding=we.Ka)(h,_,x,E,O),i._OrtRun=(h,_,x,E,O,F,q,J)=>(i._OrtRun=we.La)(h,_,x,E,O,F,q,J),i._OrtEndProfiling=h=>(i._OrtEndProfiling=we.Ma)(h),i._JsepOutput=(h,_,x)=>(i._JsepOutput=we.Na)(h,_,x),i._JsepGetNodeName=h=>(i._JsepGetNodeName=we.Oa)(h);var kr=i._malloc=h=>(kr=i._malloc=we.Pa)(h),_t=i._free=h=>(_t=i._free=we.Qa)(h),yn=h=>(yn=we.Sa)(h),wn=h=>(wn=we.Ta)(h),Kn=()=>(Kn=we.Ua)(),Nr=h=>(Nr=we.Wa)(h),ti=()=>(ti=we.Xa)(),qt=h=>(qt=we.Ya)(h),zt=()=>(zt=we.Za)();i.___start_em_js=837952,i.___stop_em_js=838113;function ri(){var h=we;h=Object.assign({},h);var _=x=>E=>x(E)>>>0;return h.la=_(h.la),h.Pa=_(h.Pa),h.Ta=_(h.Ta),h.Ua=(x=>()=>x()>>>0)(h.Ua),h}i.stackSave=()=>Kn(),i.stackRestore=h=>yn(h),i.stackAlloc=h=>wn(h),i.UTF8ToString=Ye,i.stringToUTF8=(h,_,x)=>Bt(h,te,_,x),i.lengthBytesUTF8=on;var mr;Ke=function h(){mr||Yn(),mr||(Ke=h)};function Yn(){if(!(0<Be)){if(i.preRun)for(typeof i.preRun=="function"&&(i.preRun=[i.preRun]);i.preRun.length;){var h=i.preRun.shift();ae.unshift(h)}for(;0<ae.length;)ae.shift()(i);if(!(0<Be||mr||(mr=!0,i.calledRun=!0,K))){for(;0<fe.length;)fe.shift()(i);for(s(i);0<Ce.length;)Ce.shift()(i)}}}return Yn(),u}})();typeof t=="object"&&typeof e=="object"?e.exports=r:typeof define=="function"&&define.amd&&define([],()=>r)}),G_=In(()=>{}),H_=In(()=>{}),nh={};An(nh,{cpus:()=>ah});var ah,j_=Z(()=>{ah=void 0}),q_=In((t,e)=>{var r=(()=>{var n=typeof document<"u"?document.currentScript?.src:void 0;return typeof __filename<"u"&&(n||=__filename),function(a={}){function i(){return ne.buffer!=Ce.buffer&&st(),Ce}function s(){return ne.buffer!=Ce.buffer&&st(),Be}function o(){return ne.buffer!=Ce.buffer&&st(),Ke}function u(){return ne.buffer!=Ce.buffer&&st(),Qe}function l(){return ne.buffer!=Ce.buffer&&st(),We}function p(){return ne.buffer!=Ce.buffer&&st(),dt}function f(){return ne.buffer!=Ce.buffer&&st(),Ve}function m(){return ne.buffer!=Ce.buffer&&st(),Vt}var c=a,y,w,v=new Promise((d,g)=>{y=d,w=g});c.mountExternalData=(d,g)=>{(c.Db||(c.Db=new Map)).set(d,g)},c.unmountExternalData=()=>{delete c.Db};let k=()=>{let d=(b,S,I)=>(...z)=>{let V=zt,X=S?.();z=b(...z);let ce=S?.();return X!==ce&&(b=ce,I(X),S=I=null),zt!=V?O():z},g=b=>async(...S)=>{try{if(c.Cb)throw Error("Session already started");let I=c.Cb={ec:S[0],errors:[]},z=await b(...S);if(c.Cb!==I)throw Error("Session mismatch");c.Kb?.flush();let V=I.errors;if(0<V.length){let X=await Promise.all(V);if(X=X.filter(ce=>ce),0<X.length)throw Error(X.join(`
`))}return z}finally{c.Cb=null}};c._OrtCreateSession=d(c._OrtCreateSession,()=>c._OrtCreateSession,b=>c._OrtCreateSession=b),c._OrtRun=g(d(c._OrtRun,()=>c._OrtRun,b=>c._OrtRun=b)),c._OrtRunWithBinding=g(d(c._OrtRunWithBinding,()=>c._OrtRunWithBinding,b=>c._OrtRunWithBinding=b)),c._OrtBindInput=d(c._OrtBindInput,()=>c._OrtBindInput,b=>c._OrtBindInput=b),k=void 0};c.jsepInit=(d,g)=>{if(k?.(),d==="webgpu"){[c.Kb,c.Wb,c.$b,c.Lb,c.Zb,c.ob,c.ac,c.cc,c.Xb,c.Yb,c.bc]=g;let b=c.Kb;c.jsepRegisterBuffer=(S,I,z,V)=>b.registerBuffer(S,I,z,V),c.jsepGetBuffer=S=>b.getBuffer(S),c.jsepCreateDownloader=(S,I,z)=>b.createDownloader(S,I,z),c.jsepOnReleaseSession=S=>{b.onReleaseSession(S)},c.jsepOnRunStart=S=>b.onRunStart(S)}};var $=Object.assign({},c),C="./this.program",T=(d,g)=>{throw g},A=typeof window=="object",B=typeof importScripts=="function",R=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",D=c.ENVIRONMENT_IS_PTHREAD||!1,K="";function j(d){return c.locateFile?c.locateFile(d,K):K+d}var ie,te,oe;if(R){var re=(eh(),Qr(Uo)),M=(rh(),Qr(Wo));K=B?M.dirname(K)+"/":__dirname+"/",ie=(d,g)=>(d=Bt(d)?new URL(d):M.normalize(d),re.readFileSync(d,g?void 0:"utf8")),oe=d=>(d=ie(d,!0),d.buffer||(d=new Uint8Array(d)),d),te=(d,g,b,S=!0)=>{d=Bt(d)?new URL(d):M.normalize(d),re.readFile(d,S?void 0:"utf8",(I,z)=>{I?b(I):g(S?z.buffer:z)})},!c.thisProgram&&1<process.argv.length&&(C=process.argv[1].replace(/\\/g,"/")),process.argv.slice(2),T=(d,g)=>{throw process.exitCode=d,g},global.Worker=G_().Worker}else(A||B)&&(B?K=self.location.href:typeof document<"u"&&document.currentScript&&(K=document.currentScript.src),typeof n<"u"&&n&&(K=n),K.startsWith("blob:")?K="":K=K.substr(0,K.replace(/[?#].*/,"").lastIndexOf("/")+1),R||(ie=d=>{var g=new XMLHttpRequest;return g.open("GET",d,!1),g.send(null),g.responseText},B&&(oe=d=>{var g=new XMLHttpRequest;return g.open("GET",d,!1),g.responseType="arraybuffer",g.send(null),new Uint8Array(g.response)}),te=(d,g,b)=>{var S=new XMLHttpRequest;S.open("GET",d,!0),S.responseType="arraybuffer",S.onload=()=>{S.status==200||S.status==0&&S.response?g(S.response):b()},S.onerror=b,S.send(null)}));R&&typeof performance>"u"&&(global.performance=H_().performance);var P=console.log.bind(console),H=console.error.bind(console);R&&(P=(...d)=>re.writeSync(1,d.join(" ")+`
`),H=(...d)=>re.writeSync(2,d.join(" ")+`
`));var le=P,G=H;Object.assign(c,$),$=null;var ne,N,ae=!1,fe,Ce,Be,Ke,Qe,We,dt,Ve,Ne,ht,Vt;function st(){var d=ne.buffer;c.HEAP8=Ce=new Int8Array(d),c.HEAP16=Ke=new Int16Array(d),c.HEAPU8=Be=new Uint8Array(d),c.HEAPU16=Qe=new Uint16Array(d),c.HEAP32=We=new Int32Array(d),c.HEAPU32=dt=new Uint32Array(d),c.HEAPF32=Ve=new Float32Array(d),c.HEAPF64=Vt=new Float64Array(d),c.HEAP64=Ne=new BigInt64Array(d),c.HEAPU64=ht=new BigUint64Array(d)}var zr=16777216;if(D)ne=c.wasmMemory;else if(c.wasmMemory)ne=c.wasmMemory;else if(ne=new WebAssembly.Memory({initial:zr/65536,maximum:65536,shared:!0}),!(ne.buffer instanceof SharedArrayBuffer))throw G("requested a shared WebAssembly.Memory but the returned buffer is not a SharedArrayBuffer, indicating that while the browser has SharedArrayBuffer it does not have WebAssembly threads support - you may need to set a flag"),R&&G("(on node you may need: --experimental-wasm-threads --experimental-wasm-bulk-memory and/or recent version)"),Error("bad memory");st(),zr=ne.buffer.byteLength;var hr=[],ir=[],ls=[],fr=0,xr=null;function Pn(){if(fr--,fr==0&&xr){var d=xr;xr=null,d()}}function Ye(d){throw d="Aborted("+d+")",G(d),ae=!0,fe=1,d=new WebAssembly.RuntimeError(d+". Build with -sASSERTIONS for more info."),w(d),d}var on=d=>d.startsWith("data:application/octet-stream;base64,"),Bt=d=>d.startsWith("file://"),Sr;Sr="ort-wasm-simd-threaded.wasm",on(Sr)||(Sr=j(Sr));function At(d){if(oe)return oe(d);throw"both async and sync fetching of the wasm failed"}function Rn(d){if(A||B){if(typeof fetch=="function"&&!Bt(d))return fetch(d,{credentials:"same-origin"}).then(g=>{if(!g.ok)throw`failed to load wasm binary file at '${d}'`;return g.arrayBuffer()}).catch(()=>At(d));if(te)return new Promise((g,b)=>{te(d,S=>g(new Uint8Array(S)),b)})}return Promise.resolve().then(()=>At(d))}function un(d,g,b){return Rn(d).then(S=>WebAssembly.instantiate(S,g)).then(b,S=>{G(`failed to asynchronously prepare wasm: ${S}`),Ye(S)})}function sr(d,g){var b=Sr;return typeof WebAssembly.instantiateStreaming!="function"||on(b)||Bt(b)||R||typeof fetch!="function"?un(b,d,g):fetch(b,{credentials:"same-origin"}).then(S=>WebAssembly.instantiateStreaming(S,d).then(g,function(I){return G(`wasm streaming compile failed: ${I}`),G("falling back to ArrayBuffer instantiation"),un(b,d,g)}))}var Bn={826468:(d,g,b,S)=>{if(typeof c>"u"||!c.Db)return 1;if(d=nt(d>>>0),d.startsWith("./")&&(d=d.substring(2)),d=c.Db.get(d),!d)return 2;if(g>>>=0,b>>>=0,S>>>=0,g+b>d.byteLength)return 3;try{return s().set(d.subarray(g,g+b),S>>>0),0}catch{return 4}},826969:()=>{c.Xb()},827e3:()=>{c.Yb()},827029:()=>{c.bc()},827054:d=>c.Wb(d),827087:d=>c.$b(d),827119:(d,g,b)=>{c.Lb(d,g,b,!0)},827158:(d,g,b)=>{c.Lb(d,g,b)},827191:d=>{c.ob("Abs",d,void 0)},827242:d=>{c.ob("Neg",d,void 0)},827293:d=>{c.ob("Floor",d,void 0)},827346:d=>{c.ob("Ceil",d,void 0)},827398:d=>{c.ob("Reciprocal",d,void 0)},827456:d=>{c.ob("Sqrt",d,void 0)},827508:d=>{c.ob("Exp",d,void 0)},827559:d=>{c.ob("Erf",d,void 0)},827610:d=>{c.ob("Sigmoid",d,void 0)},827665:(d,g,b)=>{c.ob("HardSigmoid",d,{alpha:g,beta:b})},827744:d=>{c.ob("Log",d,void 0)},827795:d=>{c.ob("Sin",d,void 0)},827846:d=>{c.ob("Cos",d,void 0)},827897:d=>{c.ob("Tan",d,void 0)},827948:d=>{c.ob("Asin",d,void 0)},828e3:d=>{c.ob("Acos",d,void 0)},828052:d=>{c.ob("Atan",d,void 0)},828104:d=>{c.ob("Sinh",d,void 0)},828156:d=>{c.ob("Cosh",d,void 0)},828208:d=>{c.ob("Asinh",d,void 0)},828261:d=>{c.ob("Acosh",d,void 0)},828314:d=>{c.ob("Atanh",d,void 0)},828367:d=>{c.ob("Tanh",d,void 0)},828419:d=>{c.ob("Not",d,void 0)},828470:(d,g,b)=>{c.ob("Clip",d,{min:g,max:b})},828539:d=>{c.ob("Clip",d,void 0)},828591:(d,g)=>{c.ob("Elu",d,{alpha:g})},828649:d=>{c.ob("Relu",d,void 0)},828701:(d,g)=>{c.ob("LeakyRelu",d,{alpha:g})},828765:(d,g)=>{c.ob("ThresholdedRelu",d,{alpha:g})},828835:(d,g)=>{c.ob("Cast",d,{to:g})},828893:d=>{c.ob("Add",d,void 0)},828944:d=>{c.ob("Sub",d,void 0)},828995:d=>{c.ob("Mul",d,void 0)},829046:d=>{c.ob("Div",d,void 0)},829097:d=>{c.ob("Pow",d,void 0)},829148:d=>{c.ob("Equal",d,void 0)},829201:d=>{c.ob("Greater",d,void 0)},829256:d=>{c.ob("GreaterOrEqual",d,void 0)},829318:d=>{c.ob("Less",d,void 0)},829370:d=>{c.ob("LessOrEqual",d,void 0)},829429:(d,g,b,S,I)=>{c.ob("ReduceMean",d,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:S?Array.from(l().subarray(S>>>0,I>>>0)):[]})},829588:(d,g,b,S,I)=>{c.ob("ReduceMax",d,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:S?Array.from(l().subarray(S>>>0,I>>>0)):[]})},829746:(d,g,b,S,I)=>{c.ob("ReduceMin",d,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:S?Array.from(l().subarray(S>>>0,I>>>0)):[]})},829904:(d,g,b,S,I)=>{c.ob("ReduceProd",d,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:S?Array.from(l().subarray(S>>>0,I>>>0)):[]})},830063:(d,g,b,S,I)=>{c.ob("ReduceSum",d,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:S?Array.from(l().subarray(S>>>0,I>>>0)):[]})},830221:(d,g,b,S,I)=>{c.ob("ReduceL1",d,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:S?Array.from(l().subarray(S>>>0,I>>>0)):[]})},830378:(d,g,b,S,I)=>{c.ob("ReduceL2",d,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:S?Array.from(l().subarray(S>>>0,I>>>0)):[]})},830535:(d,g,b,S,I)=>{c.ob("ReduceLogSum",d,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:S?Array.from(l().subarray(S>>>0,I>>>0)):[]})},830696:(d,g,b,S,I)=>{c.ob("ReduceSumSquare",d,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:S?Array.from(l().subarray(S>>>0,I>>>0)):[]})},830860:(d,g,b,S,I)=>{c.ob("ReduceLogSumExp",d,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:S?Array.from(l().subarray(S>>>0,I>>>0)):[]})},831024:d=>{c.ob("Where",d,void 0)},831077:(d,g,b)=>{c.ob("Transpose",d,{perm:g?Array.from(l().subarray(g>>>0,b>>>0)):[]})},831185:(d,g,b,S)=>{c.ob("DepthToSpace",d,{blocksize:g,mode:nt(b),format:S?"NHWC":"NCHW"})},831318:(d,g,b,S)=>{c.ob("DepthToSpace",d,{blocksize:g,mode:nt(b),format:S?"NHWC":"NCHW"})},831451:(d,g,b,S,I,z,V,X,ce,ue,_e,Le,Ge,U,ke)=>{c.ob("ConvTranspose",d,{format:ce?"NHWC":"NCHW",autoPad:g,dilations:[b],group:S,kernelShape:[I],pads:[z,V],strides:[X],wIsConst:()=>!!i()[ue>>>0],outputPadding:_e?Array.from(l().subarray(_e>>>0,Le>>>0)):[],outputShape:Ge?Array.from(l().subarray(Ge>>>0,U>>>0)):[],activation:nt(ke)})},831852:(d,g,b,S,I,z,V,X,ce,ue,_e,Le,Ge,U)=>{c.ob("ConvTranspose",d,{format:X?"NHWC":"NCHW",autoPad:g,dilations:Array.from(l().subarray(b>>>0,(b>>>0)+2>>>0)),group:S,kernelShape:Array.from(l().subarray(I>>>0,(I>>>0)+2>>>0)),pads:Array.from(l().subarray(z>>>0,(z>>>0)+4>>>0)),strides:Array.from(l().subarray(V>>>0,(V>>>0)+2>>>0)),wIsConst:()=>!!i()[ce>>>0],outputPadding:ue?Array.from(l().subarray(ue>>>0,_e>>>0)):[],outputShape:Le?Array.from(l().subarray(Le>>>0,Ge>>>0)):[],activation:nt(U)})},832417:(d,g,b,S,I,z,V,X,ce,ue,_e,Le,Ge,U,ke)=>{c.ob("ConvTranspose",d,{format:ce?"NHWC":"NCHW",autoPad:g,dilations:[b],group:S,kernelShape:[I],pads:[z,V],strides:[X],wIsConst:()=>!!i()[ue>>>0],outputPadding:_e?Array.from(l().subarray(_e>>>0,Le>>>0)):[],outputShape:Ge?Array.from(l().subarray(Ge>>>0,U>>>0)):[],activation:nt(ke)})},832818:(d,g,b,S,I,z,V,X,ce,ue,_e,Le,Ge,U)=>{c.ob("ConvTranspose",d,{format:X?"NHWC":"NCHW",autoPad:g,dilations:Array.from(l().subarray(b>>>0,(b>>>0)+2>>>0)),group:S,kernelShape:Array.from(l().subarray(I>>>0,(I>>>0)+2>>>0)),pads:Array.from(l().subarray(z>>>0,(z>>>0)+4>>>0)),strides:Array.from(l().subarray(V>>>0,(V>>>0)+2>>>0)),wIsConst:()=>!!i()[ce>>>0],outputPadding:ue?Array.from(l().subarray(ue>>>0,_e>>>0)):[],outputShape:Le?Array.from(l().subarray(Le>>>0,Ge>>>0)):[],activation:nt(U)})},833383:(d,g)=>{c.ob("GlobalAveragePool",d,{format:g?"NHWC":"NCHW"})},833474:(d,g,b,S,I,z,V,X,ce,ue,_e,Le,Ge,U,ke,De)=>{c.ob("AveragePool",d,{format:De?"NHWC":"NCHW",auto_pad:g,ceil_mode:b,count_include_pad:S,storage_order:I,dilations:[z,V],kernel_shape:[X,ce],pads:[ue,_e,Le,Ge],strides:[U,ke]})},833758:(d,g)=>{c.ob("GlobalAveragePool",d,{format:g?"NHWC":"NCHW"})},833849:(d,g,b,S,I,z,V,X,ce,ue,_e,Le,Ge,U,ke,De)=>{c.ob("AveragePool",d,{format:De?"NHWC":"NCHW",auto_pad:g,ceil_mode:b,count_include_pad:S,storage_order:I,dilations:[z,V],kernel_shape:[X,ce],pads:[ue,_e,Le,Ge],strides:[U,ke]})},834133:(d,g)=>{c.ob("GlobalMaxPool",d,{format:g?"NHWC":"NCHW"})},834220:(d,g,b,S,I,z,V,X,ce,ue,_e,Le,Ge,U,ke,De)=>{c.ob("MaxPool",d,{format:De?"NHWC":"NCHW",auto_pad:g,ceil_mode:b,count_include_pad:S,storage_order:I,dilations:[z,V],kernel_shape:[X,ce],pads:[ue,_e,Le,Ge],strides:[U,ke]})},834500:(d,g)=>{c.ob("GlobalMaxPool",d,{format:g?"NHWC":"NCHW"})},834587:(d,g,b,S,I,z,V,X,ce,ue,_e,Le,Ge,U,ke,De)=>{c.ob("MaxPool",d,{format:De?"NHWC":"NCHW",auto_pad:g,ceil_mode:b,count_include_pad:S,storage_order:I,dilations:[z,V],kernel_shape:[X,ce],pads:[ue,_e,Le,Ge],strides:[U,ke]})},834867:(d,g,b,S,I)=>{c.ob("Gemm",d,{alpha:g,beta:b,transA:S,transB:I})},834971:d=>{c.ob("MatMul",d,void 0)},835025:(d,g,b,S)=>{c.ob("ArgMax",d,{keepDims:!!g,selectLastIndex:!!b,axis:S})},835133:(d,g,b,S)=>{c.ob("ArgMin",d,{keepDims:!!g,selectLastIndex:!!b,axis:S})},835241:(d,g)=>{c.ob("Softmax",d,{axis:g})},835304:(d,g)=>{c.ob("Concat",d,{axis:g})},835364:(d,g,b,S,I)=>{c.ob("Split",d,{axis:g,numOutputs:b,splitSizes:S?Array.from(l().subarray(S>>>0,I>>>0)):[]})},835504:d=>{c.ob("Expand",d,void 0)},835558:(d,g)=>{c.ob("Gather",d,{axis:Number(g)})},835629:(d,g)=>{c.ob("GatherElements",d,{axis:Number(g)})},835708:(d,g,b,S,I,z,V,X,ce,ue,_e)=>{c.ob("Resize",d,{antialias:g,axes:b?Array.from(l().subarray(b>>>0,S>>>0)):[],coordinateTransformMode:nt(I),cubicCoeffA:z,excludeOutside:V,extrapolationValue:X,keepAspectRatioPolicy:nt(ce),mode:nt(ue),nearestMode:nt(_e)})},836054:(d,g,b,S,I,z,V)=>{c.ob("Slice",d,{starts:g?Array.from(l().subarray(g>>>0,b>>>0)):[],ends:S?Array.from(l().subarray(S>>>0,I>>>0)):[],axes:z?Array.from(l().subarray(z>>>0,V>>>0)):[]})},836270:d=>{c.ob("Tile",d,void 0)},836322:(d,g,b,S)=>{c.ob("LayerNormalization",d,{axis:g,epsilon:b,simplified:!!S})},836433:(d,g,b)=>{c.ob("InstanceNormalization",d,{epsilon:g,format:b?"NHWC":"NCHW"})},836547:(d,g,b)=>{c.ob("InstanceNormalization",d,{epsilon:g,format:b?"NHWC":"NCHW"})},836661:d=>{c.ob("Range",d,void 0)},836714:(d,g)=>{c.ob("Einsum",d,{equation:nt(g)})},836795:(d,g,b,S,I)=>{c.ob("Pad",d,{mode:g,value:b,pads:S?Array.from(l().subarray(S>>>0,I>>>0)):[]})},836922:(d,g,b,S,I,z)=>{c.ob("BatchNormalization",d,{epsilon:g,momentum:b,spatial:!!I,trainingMode:!!S,format:z?"NHWC":"NCHW"})},837091:(d,g,b,S,I,z)=>{c.ob("BatchNormalization",d,{epsilon:g,momentum:b,spatial:!!I,trainingMode:!!S,format:z?"NHWC":"NCHW"})},837260:(d,g,b)=>{c.ob("CumSum",d,{exclusive:Number(g),reverse:Number(b)})},837357:(d,g,b,S,I,z,V,X,ce)=>{c.ob("Attention",d,{numHeads:g,isUnidirectional:b,maskFilterValue:S,scale:I,doRotary:z,qkvHiddenSizes:V?Array.from(l().subarray(Number(X)>>>0,Number(X)+V>>>0)):[],pastPresentShareBuffer:!!ce})},837629:d=>{c.ob("BiasAdd",d,void 0)},837684:d=>{c.ob("BiasSplitGelu",d,void 0)},837745:d=>{c.ob("FastGelu",d,void 0)},837801:(d,g,b,S,I,z,V,X,ce,ue,_e,Le,Ge)=>{c.ob("Conv",d,{format:ce?"NHWC":"NCHW",auto_pad:g,dilations:[b],group:S,kernel_shape:[I],pads:z?Array.from(l().subarray(z>>>0,V>>>0)):[],strides:[X],w_is_const:()=>!!i()[ue>>>0],activation:nt(_e),activation_params:Le?Array.from(f().subarray(Le>>>0,Ge>>>0)):[]})},838171:(d,g,b,S,I,z,V,X,ce,ue,_e,Le,Ge,U,ke,De)=>{c.ob("Conv",d,{format:Le?"NHWC":"NCHW",auto_pad:g,dilations:[b,S],group:I,kernel_shape:[z,V],pads:X?Array.from(l().subarray(X>>>0,ce>>>0)):[],strides:[ue,_e],w_is_const:()=>!!i()[Ge>>>0],activation:nt(U),activation_params:ke?Array.from(f().subarray(ke>>>0,De>>>0)):[]})},838562:d=>{c.ob("Gelu",d,void 0)},838614:(d,g,b,S,I,z)=>{c.ob("MatMulNBits",d,{k:g,n:b,accuracyLevel:S,bits:I,blockSize:z})},838741:(d,g,b,S,I,z)=>{c.ob("MultiHeadAttention",d,{numHeads:g,isUnidirectional:b,maskFilterValue:S,scale:I,doRotary:z})},838900:(d,g,b,S,I)=>{c.ob("RotaryEmbedding",d,{interleaved:!!g,numHeads:b,rotaryEmbeddingDim:S,scale:I})},839039:(d,g,b)=>{c.ob("SkipLayerNormalization",d,{epsilon:g,simplified:!!b})},839141:(d,g,b)=>{c.ob("SkipLayerNormalization",d,{epsilon:g,simplified:!!b})},839243:(d,g,b,S)=>{c.ob("LayerNormalization",d,{axis:g,epsilon:b,simplified:!!S})},839354:d=>{c.ac(d)},839388:(d,g)=>c.cc(d,g,c.Cb.ec,c.Cb.errors)};function St(d){this.name="ExitStatus",this.message=`Program terminated with exit(${d})`,this.status=d}var ln=d=>{d.terminate(),d.onmessage=()=>{}},dn=d=>{be.xb.length==0&&(La(),be.Mb(be.xb[0]));var g=be.xb.pop();if(!g)return 6;be.yb.push(g),be.ub[d.wb]=g,g.wb=d.wb;var b={cmd:"run",start_routine:d.fc,arg:d.Rb,pthread_ptr:d.wb};return R&&g.unref(),g.postMessage(b,d.mc),0},ot=0,Dn=d=>{var g=$s();return d=d(),si(g),d},Ie=(d,g,...b)=>Dn(()=>{for(var S=2*b.length,I=vs(8*S),z=I>>>3,V=0;V<b.length;V++){var X=b[V];typeof X=="bigint"?(Ne[z+2*V]=1n,Ne[z+2*V+1]=X):(Ne[z+2*V]=0n,m()[z+2*V+1>>>0]=X)}return Bu(d,0,S,I,g)});function kt(d){if(D)return Ie(0,1,d);fe=d,0<ot||(be.hc(),c.onExit?.(d),ae=!0),T(d,new St(d))}var Pr=d=>{if(fe=d,D)throw Wa(d),"unwind";kt(d)};function ds(){for(var d=c.numThreads;d--;)La();hr.unshift(()=>{fr++,Ua(()=>Pn())})}function La(){var d=j("ort-wasm-simd-threaded.worker.js");d=new Worker(d),be.xb.push(d)}function Ua(d){D?d():Promise.all(be.xb.map(be.Mb)).then(d)}var be={xb:[],yb:[],Qb:[],ub:{},Gb(){D?(be.receiveObjectTransfer=be.dc,be.threadInitTLS=be.Pb,be.setExitStatus=be.Ob):ds()},Ob:d=>fe=d,pc:["$terminateWorker"],hc:()=>{for(var d of be.yb)ln(d);for(d of be.xb)ln(d);be.xb=[],be.yb=[],be.ub=[]},Nb:d=>{var g=d.wb;delete be.ub[g],be.xb.push(d),be.yb.splice(be.yb.indexOf(d),1),d.wb=0,ws(g)},dc(){},Pb(){be.Qb.forEach(d=>d())},Mb:d=>new Promise(g=>{d.onmessage=z=>{z=z.data;var V=z.cmd;if(z.targetThread&&z.targetThread!=ai()){var X=be.ub[z.targetThread];X?X.postMessage(z,z.transferList):G(`Internal error! Worker sent a message "${V}" to target pthread ${z.targetThread}, but that thread no longer exists!`)}else V==="checkMailbox"?_t():V==="spawnThread"?dn(z):V==="cleanupThread"?be.Nb(be.ub[z.thread]):V==="killThread"?(z=z.thread,V=be.ub[z],delete be.ub[z],ln(V),ws(z),be.yb.splice(be.yb.indexOf(V),1),V.wb=0):V==="cancelThread"?be.ub[z.thread].postMessage({cmd:"cancel"}):V==="loaded"?(d.loaded=!0,R&&!d.wb&&d.unref(),g(d)):V==="alert"?alert(`Thread ${z.threadId}: ${z.text}`):z.target==="setimmediate"?d.postMessage(z):V==="callHandler"?c[z.handler](...z.args):V&&G(`worker sent an unknown command ${V}`)},d.onerror=z=>{throw G(`worker sent an error! ${z.filename}:${z.lineno}: ${z.message}`),z},R&&(d.on("message",z=>d.onmessage({data:z})),d.on("error",z=>d.onerror(z)));var b=[],S=["onExit"],I;for(I of S)c.hasOwnProperty(I)&&b.push(I);d.postMessage({cmd:"load",handlers:b,urlOrBlob:c.mainScriptUrlOrBlob||n,wasmMemory:ne,wasmModule:N})})};c.PThread=be;var cn=d=>{for(;0<d.length;)d.shift()(c)};c.establishStackSpace=()=>{var d=ai(),g=p()[d+52>>>2>>>0];d=p()[d+56>>>2>>>0],Nu(g,g-d),si(g)};function Wa(d){if(D)return Ie(1,0,d);Pr(d)}c.invokeEntryPoint=(d,g)=>{ot=0,d=Fu(d,g),0<ot?be.Ob(d):bs(d)};class cs{constructor(g){this.Ib=g-24}Gb(g,b){p()[this.Ib+16>>>2>>>0]=0,p()[this.Ib+4>>>2>>>0]=g,p()[this.Ib+8>>>2>>>0]=b}}var Va=0;function Ga(d,g,b,S){return D?Ie(2,1,d,g,b,S):pn(d,g,b,S)}function pn(d,g,b,S){if(d>>>=0,g>>>=0,b>>>=0,S>>>=0,typeof SharedArrayBuffer>"u")return G("Current environment does not support SharedArrayBuffer, pthreads are not available!"),6;var I=[];return D&&I.length===0?Ga(d,g,b,S):(d={fc:b,wb:d,Rb:S,mc:I},D?(d.oc="spawnThread",postMessage(d,I),0):dn(d))}var Nn=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,Rr=(d,g,b)=>{g>>>=0;var S=g+b;for(b=g;d[b]&&!(b>=S);)++b;if(16<b-g&&d.buffer&&Nn)return Nn.decode(d.buffer instanceof SharedArrayBuffer?d.slice(g,b):d.subarray(g,b));for(S="";g<b;){var I=d[g++];if(I&128){var z=d[g++]&63;if((I&224)==192)S+=String.fromCharCode((I&31)<<6|z);else{var V=d[g++]&63;I=(I&240)==224?(I&15)<<12|z<<6|V:(I&7)<<18|z<<12|V<<6|d[g++]&63,65536>I?S+=String.fromCharCode(I):(I-=65536,S+=String.fromCharCode(55296|I>>10,56320|I&1023))}}else S+=String.fromCharCode(I)}return S},nt=(d,g)=>(d>>>=0)?Rr(s(),d,g):"";function Ha(d,g,b){return D?Ie(3,1,d,g,b):0}function Gt(d,g){if(D)return Ie(4,1,d,g)}var Et=d=>{for(var g=0,b=0;b<d.length;++b){var S=d.charCodeAt(b);127>=S?g++:2047>=S?g+=2:55296<=S&&57343>=S?(g+=4,++b):g+=3}return g},Fn=(d,g,b,S)=>{if(b>>>=0,!(0<S))return 0;var I=b;S=b+S-1;for(var z=0;z<d.length;++z){var V=d.charCodeAt(z);if(55296<=V&&57343>=V){var X=d.charCodeAt(++z);V=65536+((V&1023)<<10)|X&1023}if(127>=V){if(b>=S)break;g[b++>>>0]=V}else{if(2047>=V){if(b+1>=S)break;g[b++>>>0]=192|V>>6}else{if(65535>=V){if(b+2>=S)break;g[b++>>>0]=224|V>>12}else{if(b+3>=S)break;g[b++>>>0]=240|V>>18,g[b++>>>0]=128|V>>12&63}g[b++>>>0]=128|V>>6&63}g[b++>>>0]=128|V&63}}return g[b>>>0]=0,b-I},Ht=(d,g,b)=>Fn(d,s(),g,b);function Ln(d,g){if(D)return Ie(5,1,d,g)}function Un(d,g,b){if(D)return Ie(6,1,d,g,b)}function ja(d,g,b){return D?Ie(7,1,d,g,b):0}function hn(d,g){if(D)return Ie(8,1,d,g)}function qa(d,g,b){if(D)return Ie(9,1,d,g,b)}function Ka(d,g,b,S){if(D)return Ie(10,1,d,g,b,S)}function Ya(d,g,b,S){if(D)return Ie(11,1,d,g,b,S)}function Xa(d,g,b,S){if(D)return Ie(12,1,d,g,b,S)}function Wn(d){if(D)return Ie(13,1,d)}function Br(d,g){if(D)return Ie(14,1,d,g)}function Qa(d,g,b){if(D)return Ie(15,1,d,g,b)}var fn,Mt=d=>{for(var g="";s()[d>>>0];)g+=fn[s()[d++>>>0]];return g},Vn={},Gn={},jt;function ps(d,g,b={}){var S=g.name;if(!d)throw new jt(`type "${S}" must have a positive integer typeid pointer`);if(Gn.hasOwnProperty(d)){if(b.Ub)return;throw new jt(`Cannot register type '${S}' twice`)}Gn[d]=g,Vn.hasOwnProperty(d)&&(g=Vn[d],delete Vn[d],g.forEach(I=>I()))}function ft(d,g,b={}){if(!("argPackAdvance"in g))throw new TypeError("registerType registeredInstance requires argPackAdvance");return ps(d,g,b)}var Hn=(d,g,b)=>{switch(g){case 1:return b?S=>i()[S>>>0]:S=>s()[S>>>0];case 2:return b?S=>o()[S>>>1>>>0]:S=>u()[S>>>1>>>0];case 4:return b?S=>l()[S>>>2>>>0]:S=>p()[S>>>2>>>0];case 8:return b?S=>Ne[S>>>3]:S=>ht[S>>>3];default:throw new TypeError(`invalid integer width (${g}): ${d}`)}},mn=[],Ot=[];function gn(d){d>>>=0,9<d&&--Ot[d+1]===0&&(Ot[d]=void 0,mn.push(d))}var ut=d=>{if(!d)throw new jt("Cannot use deleted val. handle = "+d);return Ot[d]},Ct=d=>{switch(d){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let g=mn.pop()||Ot.length;return Ot[g]=d,Ot[g+1]=1,g}};function Dr(d){return this.fromWireType(p()[d>>>2>>>0])}var hs={name:"emscripten::val",fromWireType:d=>{var g=ut(d);return gn(d),g},toWireType:(d,g)=>Ct(g),argPackAdvance:8,readValueFromPointer:Dr,Bb:null},Ja=(d,g)=>{switch(g){case 4:return function(b){return this.fromWireType(f()[b>>>2>>>0])};case 8:return function(b){return this.fromWireType(m()[b>>>3>>>0])};default:throw new TypeError(`invalid float width (${g}): ${d}`)}},jn=typeof TextDecoder<"u"?new TextDecoder("utf-16le"):void 0,fs=(d,g)=>{for(var b=d>>1,S=b+g/2;!(b>=S)&&u()[b>>>0];)++b;if(b<<=1,32<b-d&&jn)return jn.decode(s().slice(d,b));for(b="",S=0;!(S>=g/2);++S){var I=o()[d+2*S>>>1>>>0];if(I==0)break;b+=String.fromCharCode(I)}return b},Za=(d,g,b)=>{if(b??=2147483647,2>b)return 0;b-=2;var S=g;b=b<2*d.length?b/2:d.length;for(var I=0;I<b;++I){var z=d.charCodeAt(I);o()[g>>>1>>>0]=z,g+=2}return o()[g>>>1>>>0]=0,g-S},ei=d=>2*d.length,_n=(d,g)=>{for(var b=0,S="";!(b>=g/4);){var I=l()[d+4*b>>>2>>>0];if(I==0)break;++b,65536<=I?(I-=65536,S+=String.fromCharCode(55296|I>>10,56320|I&1023)):S+=String.fromCharCode(I)}return S},ms=(d,g,b)=>{if(g>>>=0,b??=2147483647,4>b)return 0;var S=g;b=S+b-4;for(var I=0;I<d.length;++I){var z=d.charCodeAt(I);if(55296<=z&&57343>=z){var V=d.charCodeAt(++I);z=65536+((z&1023)<<10)|V&1023}if(l()[g>>>2>>>0]=z,g+=4,g+4>b)break}return l()[g>>>2>>>0]=0,g-S},we=d=>{for(var g=0,b=0;b<d.length;++b){var S=d.charCodeAt(b);55296<=S&&57343>=S&&++b,g+=4}return g},qn=d=>{if(!ae)try{if(d(),!(0<ot))try{D?bs(fe):Pr(fe)}catch(g){g instanceof St||g=="unwind"||T(1,g)}}catch(g){g instanceof St||g=="unwind"||T(1,g)}};function kr(d){d>>>=0,typeof Atomics.nc=="function"&&(Atomics.nc(l(),d>>>2,d).value.then(_t),d+=128,Atomics.store(l(),d>>>2,1))}c.__emscripten_thread_mailbox_await=kr;var _t=()=>{var d=ai();d&&(kr(d),qn(Du))};c.checkMailbox=_t;var yn=[],wn=(d,g)=>{var b=Gn[d];if(b===void 0)throw d=Pu(d),b=Mt(d),or(d),new jt(`${g} has unknown type ${b}`);return b},Kn=(d,g,b)=>{var S=[];return d=d.toWireType(S,b),S.length&&(p()[g>>>2>>>0]=Ct(S)),d},Nr=d=>{try{d()}catch(g){Ye(g)}};function ti(){var d=he,g={};for(let[b,S]of Object.entries(d))g[b]=typeof S=="function"?(...I)=>{mr.push(b);try{return S(...I)}finally{ae||(mr.pop(),zt&&qt===1&&mr.length===0&&(qt=0,ot+=1,Nr(Uu),typeof Fibers<"u"&&Fibers.rc()))}}:S;return g}var qt=0,zt=null,ri=0,mr=[],Yn={},h={},_=0,x=null,E=[];function O(){return new Promise((d,g)=>{x={resolve:d,reject:g}})}function F(){var d=ii(65548),g=d+12;p()[d>>>2>>>0]=g,p()[d+4>>>2>>>0]=g+65536,g=mr[0];var b=Yn[g];return b===void 0&&(b=_++,Yn[g]=b,h[b]=g),g=b,l()[d+8>>>2>>>0]=g,d}function q(){var d=l()[zt+8>>>2>>>0];return d=he[h[d]],--ot,d()}function J(d){if(!ae){if(qt===0){var g=!1,b=!1;d((S=0)=>{if(!ae&&(ri=S,g=!0,b)){qt=2,Nr(()=>Wu(zt)),typeof Browser<"u"&&Browser.Hb.Tb&&Browser.Hb.resume(),S=!1;try{var I=q()}catch(X){I=X,S=!0}var z=!1;if(!zt){var V=x;V&&(x=null,(S?V.reject:V.resolve)(I),z=!0)}if(S&&!z)throw I}}),b=!0,g||(qt=1,zt=F(),typeof Browser<"u"&&Browser.Hb.Tb&&Browser.Hb.pause(),Nr(()=>Lu(zt)))}else qt===2?(qt=0,Nr(Vu),or(zt),zt=null,E.forEach(qn)):Ye(`invalid state: ${qt}`);return ri}}function de(d){return J(g=>{d().then(g)})}var se=[],me={},ze=d=>{var g=me[d];return g===void 0?Mt(d):g},Fe=()=>typeof globalThis=="object"?globalThis:Function("return this")(),L=d=>{var g=se.length;return se.push(d),g},xe=(d,g)=>{for(var b=Array(d),S=0;S<d;++S)b[S]=wn(p()[g+4*S>>>2>>>0],"parameter "+S);return b},Oe=(d,g)=>Object.defineProperty(g,"name",{value:d});function gr(d){var g=Function;if(!(g instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof g} which is not a function`);var b=Oe(g.name||"unknownFunctionName",function(){});return b.prototype=g.prototype,b=new b,d=g.apply(b,d),d instanceof Object?d:b}var _r=d=>d%4===0&&(d%100!==0||d%400===0),yu=[0,31,60,91,121,152,182,213,244,274,305,335],wu=[0,31,59,90,120,151,181,212,243,273,304,334];function bu(d,g,b,S,I,z,V){return D?Ie(16,1,d,g,b,S,I,z,V):-52}function vu(d,g,b,S,I,z){if(D)return Ie(17,1,d,g,b,S,I,z)}var gs=[],$u=(d,g)=>{gs.length=0;for(var b;b=s()[d++>>>0];){var S=b!=105;S&=b!=112,g+=S&&g%8?4:0,gs.push(b==112?p()[g>>>2>>>0]:b==106?Ne[g>>>3]:b==105?l()[g>>>2>>>0]:m()[g>>>3>>>0]),g+=S?8:4}return gs},_s={},xu=()=>{if(!ys){var d={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(typeof navigator=="object"&&navigator.languages&&navigator.languages[0]||"C").replace("-","_")+".UTF-8",_:C||"./this.program"},g;for(g in _s)_s[g]===void 0?delete d[g]:d[g]=_s[g];var b=[];for(g in d)b.push(`${g}=${d[g]}`);ys=b}return ys},ys;function Su(d,g){if(D)return Ie(18,1,d,g);d>>>=0,g>>>=0;var b=0;return xu().forEach((S,I)=>{var z=g+b;for(I=p()[d+4*I>>>2>>>0]=z,z=0;z<S.length;++z)i()[I++>>>0]=S.charCodeAt(z);i()[I>>>0]=0,b+=S.length+1}),0}function ku(d,g){if(D)return Ie(19,1,d,g);d>>>=0,g>>>=0;var b=xu();p()[d>>>2>>>0]=b.length;var S=0;return b.forEach(I=>S+=I.length+1),p()[g>>>2>>>0]=S,0}function Eu(d){return D?Ie(20,1,d):52}function Cu(d,g,b,S){return D?Ie(21,1,d,g,b,S):52}function Tu(d,g,b,S){return D?Ie(22,1,d,g,b,S):70}var Q0=[null,[],[]];function Iu(d,g,b,S){if(D)return Ie(23,1,d,g,b,S);g>>>=0,b>>>=0,S>>>=0;for(var I=0,z=0;z<b;z++){var V=p()[g>>>2>>>0],X=p()[g+4>>>2>>>0];g+=8;for(var ce=0;ce<X;ce++){var ue=s()[V+ce>>>0],_e=Q0[d];ue===0||ue===10?((d===1?le:G)(Rr(_e,0)),_e.length=0):_e.push(ue)}I+=X}return p()[S>>>2>>>0]=I,0}var Au=[31,29,31,30,31,30,31,31,30,31,30,31],Mu=[31,28,31,30,31,30,31,31,30,31,30,31];function J0(d){var g=Array(Et(d)+1);return Fn(d,g,0,g.length),g}var Z0=(d,g)=>{i().set(d,g>>>0)};function Ou(d,g,b,S){function I(U,ke,De){for(U=typeof U=="number"?U.toString():U||"";U.length<ke;)U=De[0]+U;return U}function z(U,ke){return I(U,ke,"0")}function V(U,ke){function De(Hu){return 0>Hu?-1:0<Hu?1:0}var Fr;return(Fr=De(U.getFullYear()-ke.getFullYear()))===0&&(Fr=De(U.getMonth()-ke.getMonth()))===0&&(Fr=De(U.getDate()-ke.getDate())),Fr}function X(U){switch(U.getDay()){case 0:return new Date(U.getFullYear()-1,11,29);case 1:return U;case 2:return new Date(U.getFullYear(),0,3);case 3:return new Date(U.getFullYear(),0,2);case 4:return new Date(U.getFullYear(),0,1);case 5:return new Date(U.getFullYear()-1,11,31);case 6:return new Date(U.getFullYear()-1,11,30)}}function ce(U){var ke=U.zb;for(U=new Date(new Date(U.Ab+1900,0,1).getTime());0<ke;){var De=U.getMonth(),Fr=(_r(U.getFullYear())?Au:Mu)[De];if(ke>Fr-U.getDate())ke-=Fr-U.getDate()+1,U.setDate(1),11>De?U.setMonth(De+1):(U.setMonth(0),U.setFullYear(U.getFullYear()+1));else{U.setDate(U.getDate()+ke);break}}return De=new Date(U.getFullYear()+1,0,4),ke=X(new Date(U.getFullYear(),0,4)),De=X(De),0>=V(ke,U)?0>=V(De,U)?U.getFullYear()+1:U.getFullYear():U.getFullYear()-1}d>>>=0,g>>>=0,b>>>=0,S>>>=0;var ue=p()[S+40>>>2>>>0];S={kc:l()[S>>>2>>>0],jc:l()[S+4>>>2>>>0],Eb:l()[S+8>>>2>>>0],Jb:l()[S+12>>>2>>>0],Fb:l()[S+16>>>2>>>0],Ab:l()[S+20>>>2>>>0],vb:l()[S+24>>>2>>>0],zb:l()[S+28>>>2>>>0],qc:l()[S+32>>>2>>>0],ic:l()[S+36>>>2>>>0],lc:ue?nt(ue):""},b=nt(b),ue={"%c":"%a %b %d %H:%M:%S %Y","%D":"%m/%d/%y","%F":"%Y-%m-%d","%h":"%b","%r":"%I:%M:%S %p","%R":"%H:%M","%T":"%H:%M:%S","%x":"%m/%d/%y","%X":"%H:%M:%S","%Ec":"%c","%EC":"%C","%Ex":"%m/%d/%y","%EX":"%H:%M:%S","%Ey":"%y","%EY":"%Y","%Od":"%d","%Oe":"%e","%OH":"%H","%OI":"%I","%Om":"%m","%OM":"%M","%OS":"%S","%Ou":"%u","%OU":"%U","%OV":"%V","%Ow":"%w","%OW":"%W","%Oy":"%y"};for(var _e in ue)b=b.replace(new RegExp(_e,"g"),ue[_e]);var Le="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),Ge="January February March April May June July August September October November December".split(" ");ue={"%a":U=>Le[U.vb].substring(0,3),"%A":U=>Le[U.vb],"%b":U=>Ge[U.Fb].substring(0,3),"%B":U=>Ge[U.Fb],"%C":U=>z((U.Ab+1900)/100|0,2),"%d":U=>z(U.Jb,2),"%e":U=>I(U.Jb,2," "),"%g":U=>ce(U).toString().substring(2),"%G":ce,"%H":U=>z(U.Eb,2),"%I":U=>(U=U.Eb,U==0?U=12:12<U&&(U-=12),z(U,2)),"%j":U=>{for(var ke=0,De=0;De<=U.Fb-1;ke+=(_r(U.Ab+1900)?Au:Mu)[De++]);return z(U.Jb+ke,3)},"%m":U=>z(U.Fb+1,2),"%M":U=>z(U.jc,2),"%n":()=>`
`,"%p":U=>0<=U.Eb&&12>U.Eb?"AM":"PM","%S":U=>z(U.kc,2),"%t":()=>"	","%u":U=>U.vb||7,"%U":U=>z(Math.floor((U.zb+7-U.vb)/7),2),"%V":U=>{var ke=Math.floor((U.zb+7-(U.vb+6)%7)/7);if(2>=(U.vb+371-U.zb-2)%7&&ke++,ke)ke==53&&(De=(U.vb+371-U.zb)%7,De==4||De==3&&_r(U.Ab)||(ke=1));else{ke=52;var De=(U.vb+7-U.zb-1)%7;(De==4||De==5&&_r(U.Ab%400-1))&&ke++}return z(ke,2)},"%w":U=>U.vb,"%W":U=>z(Math.floor((U.zb+7-(U.vb+6)%7)/7),2),"%y":U=>(U.Ab+1900).toString().substring(2),"%Y":U=>U.Ab+1900,"%z":U=>{U=U.ic;var ke=0<=U;return U=Math.abs(U)/60,(ke?"+":"-")+("0000"+(U/60*100+U%60)).slice(-4)},"%Z":U=>U.lc,"%%":()=>"%"},b=b.replace(/%%/g,"\0\0");for(_e in ue)b.includes(_e)&&(b=b.replace(new RegExp(_e,"g"),ue[_e](S)));return b=b.replace(/\0\0/g,"%"),_e=J0(b),_e.length>g?0:(Z0(_e,d),_e.length-1)}be.Gb();for(var zu=Array(256),ni=0;256>ni;++ni)zu[ni]=String.fromCharCode(ni);fn=zu,jt=c.BindingError=class extends Error{constructor(d){super(d),this.name="BindingError"}},c.InternalError=class extends Error{constructor(d){super(d),this.name="InternalError"}},Ot.push(0,1,void 0,1,null,1,!0,1,!1,1),c.count_emval_handles=()=>Ot.length/2-5-mn.length;var e_=[kt,Wa,Ga,Ha,Gt,Ln,Un,ja,hn,qa,Ka,Ya,Xa,Wn,Br,Qa,bu,vu,Su,ku,Eu,Cu,Tu,Iu],t_={ua:function(d,g,b){return de(async()=>{await c.Zb(d,g,b)})},b:function(d,g,b){throw d>>>=0,new cs(d).Gb(g>>>0,b>>>0),Va=d,Va},ia:function(d){Ru(d>>>0,!B,1,!A,131072,!1),be.Pb()},G:function(d){d>>>=0,D?postMessage({cmd:"cleanupThread",thread:d}):be.Nb(be.ub[d])},_:pn,A:Ha,pa:Gt,W:Ln,Y:Un,qa:ja,na:hn,fa:qa,ma:Ka,K:Ya,X:Xa,U:Wn,oa:Br,V:Qa,E:function(d,g,b){d>>>=0,b>>>=0,g=Mt(g>>>0),ft(d,{name:g,fromWireType:S=>S,toWireType:function(S,I){if(typeof I!="bigint"&&typeof I!="number")throw I===null?I="null":(S=typeof I,I=S==="object"||S==="array"||S==="function"?I.toString():""+I),new TypeError(`Cannot convert "${I}" to ${this.name}`);return typeof I=="number"&&(I=BigInt(I)),I},argPackAdvance:8,readValueFromPointer:Hn(g,b,g.indexOf("u")==-1),Bb:null})},O:function(d,g,b,S){d>>>=0,g=Mt(g>>>0),ft(d,{name:g,fromWireType:function(I){return!!I},toWireType:function(I,z){return z?b:S},argPackAdvance:8,readValueFromPointer:function(I){return this.fromWireType(s()[I>>>0])},Bb:null})},N:function(d){return ft(d>>>0,hs)},D:function(d,g,b){d>>>=0,b>>>=0,g=Mt(g>>>0),ft(d,{name:g,fromWireType:S=>S,toWireType:(S,I)=>I,argPackAdvance:8,readValueFromPointer:Ja(g,b),Bb:null})},t:function(d,g,b,S,I){if(d>>>=0,b>>>=0,g=Mt(g>>>0),I===-1&&(I=4294967295),I=X=>X,S===0){var z=32-8*b;I=X=>X<<z>>>z}var V=g.includes("unsigned")?function(X,ce){return ce>>>0}:function(X,ce){return ce};ft(d,{name:g,fromWireType:I,toWireType:V,argPackAdvance:8,readValueFromPointer:Hn(g,b,S!==0),Bb:null})},p:function(d,g,b){function S(z){var V=p()[z>>>2>>>0];return z=p()[z+4>>>2>>>0],new I(i().buffer,z,V)}d>>>=0;var I=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][g];b=Mt(b>>>0),ft(d,{name:b,fromWireType:S,argPackAdvance:8,readValueFromPointer:S},{Ub:!0})},F:function(d,g){d>>>=0,g=Mt(g>>>0);var b=g==="std::string";ft(d,{name:g,fromWireType:function(S){var I=p()[S>>>2>>>0],z=S+4;if(b)for(var V=z,X=0;X<=I;++X){var ce=z+X;if(X==I||s()[ce>>>0]==0){if(V=nt(V,ce-V),ue===void 0)var ue=V;else ue+="\0",ue+=V;V=ce+1}}else{for(ue=Array(I),X=0;X<I;++X)ue[X]=String.fromCharCode(s()[z+X>>>0]);ue=ue.join("")}return or(S),ue},toWireType:function(S,I){I instanceof ArrayBuffer&&(I=new Uint8Array(I));var z=typeof I=="string";if(!(z||I instanceof Uint8Array||I instanceof Uint8ClampedArray||I instanceof Int8Array))throw new jt("Cannot pass non-string to std::string");var V=b&&z?Et(I):I.length,X=ii(4+V+1),ce=X+4;if(p()[X>>>2>>>0]=V,b&&z)Ht(I,ce,V+1);else if(z)for(z=0;z<V;++z){var ue=I.charCodeAt(z);if(255<ue)throw or(ce),new jt("String has UTF-16 code units that do not fit in 8 bits");s()[ce+z>>>0]=ue}else for(z=0;z<V;++z)s()[ce+z>>>0]=I[z];return S!==null&&S.push(or,X),X},argPackAdvance:8,readValueFromPointer:Dr,Bb(S){or(S)}})},z:function(d,g,b){if(d>>>=0,g>>>=0,b>>>=0,b=Mt(b),g===2)var S=fs,I=Za,z=ei,V=X=>u()[X>>>1>>>0];else g===4&&(S=_n,I=ms,z=we,V=X=>p()[X>>>2>>>0]);ft(d,{name:b,fromWireType:X=>{for(var ce=p()[X>>>2>>>0],ue,_e=X+4,Le=0;Le<=ce;++Le){var Ge=X+4+Le*g;(Le==ce||V(Ge)==0)&&(_e=S(_e,Ge-_e),ue===void 0?ue=_e:(ue+="\0",ue+=_e),_e=Ge+g)}return or(X),ue},toWireType:(X,ce)=>{if(typeof ce!="string")throw new jt(`Cannot pass non-string to C++ string type ${b}`);var ue=z(ce),_e=ii(4+ue+g);return p()[_e>>>2>>>0]=ue/g,I(ce,_e+4,ue+g),X!==null&&X.push(or,_e),_e},argPackAdvance:8,readValueFromPointer:Dr,Bb(X){or(X)}})},P:function(d,g){d>>>=0,g=Mt(g>>>0),ft(d,{Vb:!0,name:g,argPackAdvance:0,fromWireType:()=>{},toWireType:()=>{}})},ta:()=>1,S:function(d,g){d>>>=0,d==g>>>0?setTimeout(_t):D?postMessage({targetThread:d,cmd:"checkMailbox"}):(d=be.ub[d])&&d.postMessage({cmd:"checkMailbox"})},$:function(d,g,b,S,I){g>>>=0,b>>>=0,S/=2,yn.length=S,I=I>>>0>>>3;for(var z=0;z<S;z++)yn[z]=Ne[I+2*z]?Ne[I+2*z+1]:m()[I+2*z+1>>>0];return d=g?Bn[g]:e_[d],be.Sb=b,b=d(...yn),be.Sb=0,b},ha:kr,sa:function(d){R&&be.ub[d>>>0].ref()},w:function(d,g,b){return g>>>=0,b>>>=0,d=ut(d>>>0),g=wn(g,"emval::as"),Kn(g,b,d)},y:function(d){return d>>>=0,de(()=>(d=ut(d),d.then(Ct)))},o:function(d,g,b,S){return b>>>=0,S>>>=0,d=se[d>>>0],g=ut(g>>>0),d(null,g,b,S)},k:function(d,g,b,S,I){return b>>>=0,S>>>=0,I>>>=0,d=se[d>>>0],g=ut(g>>>0),b=ze(b),d(g,g[b],S,I)},c:gn,C:function(d,g){return g>>>=0,d=ut(d>>>0),g=ut(g),d==g},n:function(d){return d>>>=0,d===0?Ct(Fe()):(d=ze(d),Ct(Fe()[d]))},j:function(d,g,b){g=xe(d,g>>>0);var S=g.shift();d--;var I=`return function (obj, func, destructorsRef, args) {
`,z=0,V=[];b===0&&V.push("obj");for(var X=["retType"],ce=[S],ue=0;ue<d;++ue)V.push("arg"+ue),X.push("argType"+ue),ce.push(g[ue]),I+=`  var arg${ue} = argType${ue}.readValueFromPointer(args${z?"+"+z:""});
`,z+=g[ue].argPackAdvance;return I+=`  var rv = ${b===1?"new func":"func.call"}(${V.join(", ")});
`,S.Vb||(X.push("emval_returnValue"),ce.push(Kn),I+=`  return emval_returnValue(retType, destructorsRef, rv);
`),X.push(I+`};
`),d=gr(X)(...ce),b=`methodCaller<(${g.map(_e=>_e.name).join(", ")}) => ${S.name}>`,L(Oe(b,d))},s:function(d,g){return g>>>=0,d=ut(d>>>0),g=ut(g),Ct(d[g])},d:function(d){d>>>=0,9<d&&(Ot[d+1]+=1)},v:function(){return Ct([])},m:function(d){d=ut(d>>>0);for(var g=Array(d.length),b=0;b<d.length;b++)g[b]=d[b];return Ct(g)},g:function(d){return Ct(ze(d>>>0))},l:function(){return Ct({})},i:function(d){d>>>=0;for(var g=ut(d);g.length;){var b=g.pop();g.pop()(b)}gn(d)},h:function(d,g,b){g>>>=0,b>>>=0,d=ut(d>>>0),g=ut(g),b=ut(b),d[g]=b},e:function(d,g){return g>>>=0,d=wn(d>>>0,"_emval_take_value"),d=d.readValueFromPointer(g),Ct(d)},ca:function(d,g){d=-9007199254740992>d||9007199254740992<d?NaN:Number(d),g>>>=0,d=new Date(1e3*d),l()[g>>>2>>>0]=d.getUTCSeconds(),l()[g+4>>>2>>>0]=d.getUTCMinutes(),l()[g+8>>>2>>>0]=d.getUTCHours(),l()[g+12>>>2>>>0]=d.getUTCDate(),l()[g+16>>>2>>>0]=d.getUTCMonth(),l()[g+20>>>2>>>0]=d.getUTCFullYear()-1900,l()[g+24>>>2>>>0]=d.getUTCDay(),d=(d.getTime()-Date.UTC(d.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,l()[g+28>>>2>>>0]=d},da:function(d,g){d=-9007199254740992>d||9007199254740992<d?NaN:Number(d),g>>>=0,d=new Date(1e3*d),l()[g>>>2>>>0]=d.getSeconds(),l()[g+4>>>2>>>0]=d.getMinutes(),l()[g+8>>>2>>>0]=d.getHours(),l()[g+12>>>2>>>0]=d.getDate(),l()[g+16>>>2>>>0]=d.getMonth(),l()[g+20>>>2>>>0]=d.getFullYear()-1900,l()[g+24>>>2>>>0]=d.getDay();var b=(_r(d.getFullYear())?yu:wu)[d.getMonth()]+d.getDate()-1|0;l()[g+28>>>2>>>0]=b,l()[g+36>>>2>>>0]=-(60*d.getTimezoneOffset()),b=new Date(d.getFullYear(),6,1).getTimezoneOffset();var S=new Date(d.getFullYear(),0,1).getTimezoneOffset();d=(b!=S&&d.getTimezoneOffset()==Math.min(S,b))|0,l()[g+32>>>2>>>0]=d},ea:function(d){d>>>=0;var g=new Date(l()[d+20>>>2>>>0]+1900,l()[d+16>>>2>>>0],l()[d+12>>>2>>>0],l()[d+8>>>2>>>0],l()[d+4>>>2>>>0],l()[d>>>2>>>0],0),b=l()[d+32>>>2>>>0],S=g.getTimezoneOffset(),I=new Date(g.getFullYear(),6,1).getTimezoneOffset(),z=new Date(g.getFullYear(),0,1).getTimezoneOffset(),V=Math.min(z,I);return 0>b?l()[d+32>>>2>>>0]=+(I!=z&&V==S):0<b!=(V==S)&&(I=Math.max(z,I),g.setTime(g.getTime()+6e4*((0<b?V:I)-S))),l()[d+24>>>2>>>0]=g.getDay(),b=(_r(g.getFullYear())?yu:wu)[g.getMonth()]+g.getDate()-1|0,l()[d+28>>>2>>>0]=b,l()[d>>>2>>>0]=g.getSeconds(),l()[d+4>>>2>>>0]=g.getMinutes(),l()[d+8>>>2>>>0]=g.getHours(),l()[d+12>>>2>>>0]=g.getDate(),l()[d+16>>>2>>>0]=g.getMonth(),l()[d+20>>>2>>>0]=g.getYear(),d=g.getTime(),BigInt(isNaN(d)?-1:d/1e3)},aa:bu,ba:vu,R:function(d,g,b,S){d>>>=0,g>>>=0,b>>>=0,S>>>=0;var I=new Date().getFullYear(),z=new Date(I,0,1),V=new Date(I,6,1);I=z.getTimezoneOffset();var X=V.getTimezoneOffset(),ce=Math.max(I,X);p()[d>>>2>>>0]=60*ce,l()[g>>>2>>>0]=+(I!=X),d=ue=>ue.toLocaleTimeString(void 0,{hour12:!1,timeZoneName:"short"}).split(" ")[1],z=d(z),V=d(V),X<I?(Ht(z,b,17),Ht(V,S,17)):(Ht(z,S,17),Ht(V,b,17))},q:()=>{Ye("")},f:function(d,g,b){return d>>>=0,g=$u(g>>>0,b>>>0),Bn[d](...g)},M:function(d,g,b){return d>>>=0,g=$u(g>>>0,b>>>0),Bn[d](...g)},H:()=>{},L:()=>Date.now(),ra:()=>{throw ot+=1,"unwind"},T:function(){return 4294901760},u:()=>performance.timeOrigin+performance.now(),x:()=>R?(j_(),Qr(nh)).cpus().length:navigator.hardwareConcurrency,Q:function(d){d>>>=0;var g=s().length;if(d<=g||4294901760<d)return!1;for(var b=1;4>=b;b*=2){var S=g*(1+.2/b);S=Math.min(S,d+100663296);var I=Math;S=Math.max(d,S);e:{I=(I.min.call(I,4294901760,S+(65536-S%65536)%65536)-ne.buffer.byteLength+65535)/65536;try{ne.grow(I),st();var z=1;break e}catch{}z=void 0}if(z)return!0}return!1},ja:Su,la:ku,Z:Pr,B:Eu,J:Cu,ga:Tu,I:Iu,a:ne||c.wasmMemory,ka:Ou,r:function(d,g,b,S){return Ou(d>>>0,g>>>0,b>>>0,S>>>0)}},he=function(){function d(b,S){return he=b.exports,he=ti(),he=r_(),be.Qb.push(he.cb),ir.unshift(he.va),N=S,Pn(),he}var g={a:t_};if(fr++,c.instantiateWasm)try{return c.instantiateWasm(g,d)}catch(b){G(`Module.instantiateWasm callback failed with error: ${b}`),w(b)}return sr(g,function(b){d(b.instance,b.module)}).catch(w),{}}(),Pu=d=>(Pu=he.wa)(d);c.__embind_initialize_bindings=()=>(c.__embind_initialize_bindings=he.xa)(),c._OrtInit=(d,g)=>(c._OrtInit=he.ya)(d,g),c._OrtGetLastError=(d,g)=>(c._OrtGetLastError=he.za)(d,g),c._OrtCreateSessionOptions=(d,g,b,S,I,z,V,X,ce,ue)=>(c._OrtCreateSessionOptions=he.Aa)(d,g,b,S,I,z,V,X,ce,ue),c._OrtAppendExecutionProvider=(d,g)=>(c._OrtAppendExecutionProvider=he.Ba)(d,g),c._OrtAddFreeDimensionOverride=(d,g,b)=>(c._OrtAddFreeDimensionOverride=he.Ca)(d,g,b),c._OrtAddSessionConfigEntry=(d,g,b)=>(c._OrtAddSessionConfigEntry=he.Da)(d,g,b),c._OrtReleaseSessionOptions=d=>(c._OrtReleaseSessionOptions=he.Ea)(d),c._OrtCreateSession=(d,g,b)=>(c._OrtCreateSession=he.Fa)(d,g,b),c._OrtReleaseSession=d=>(c._OrtReleaseSession=he.Ga)(d),c._OrtGetInputOutputCount=(d,g,b)=>(c._OrtGetInputOutputCount=he.Ha)(d,g,b),c._OrtGetInputName=(d,g)=>(c._OrtGetInputName=he.Ia)(d,g),c._OrtGetOutputName=(d,g)=>(c._OrtGetOutputName=he.Ja)(d,g),c._OrtFree=d=>(c._OrtFree=he.Ka)(d),c._OrtCreateTensor=(d,g,b,S,I,z)=>(c._OrtCreateTensor=he.La)(d,g,b,S,I,z),c._OrtGetTensorData=(d,g,b,S,I)=>(c._OrtGetTensorData=he.Ma)(d,g,b,S,I),c._OrtReleaseTensor=d=>(c._OrtReleaseTensor=he.Na)(d),c._OrtCreateRunOptions=(d,g,b,S)=>(c._OrtCreateRunOptions=he.Oa)(d,g,b,S),c._OrtAddRunConfigEntry=(d,g,b)=>(c._OrtAddRunConfigEntry=he.Pa)(d,g,b),c._OrtReleaseRunOptions=d=>(c._OrtReleaseRunOptions=he.Qa)(d),c._OrtCreateBinding=d=>(c._OrtCreateBinding=he.Ra)(d),c._OrtBindInput=(d,g,b)=>(c._OrtBindInput=he.Sa)(d,g,b),c._OrtBindOutput=(d,g,b,S)=>(c._OrtBindOutput=he.Ta)(d,g,b,S),c._OrtClearBoundOutputs=d=>(c._OrtClearBoundOutputs=he.Ua)(d),c._OrtReleaseBinding=d=>(c._OrtReleaseBinding=he.Va)(d),c._OrtRunWithBinding=(d,g,b,S,I)=>(c._OrtRunWithBinding=he.Wa)(d,g,b,S,I),c._OrtRun=(d,g,b,S,I,z,V,X)=>(c._OrtRun=he.Xa)(d,g,b,S,I,z,V,X),c._OrtEndProfiling=d=>(c._OrtEndProfiling=he.Ya)(d),c._JsepOutput=(d,g,b)=>(c._JsepOutput=he.Za)(d,g,b),c._JsepGetNodeName=d=>(c._JsepGetNodeName=he._a)(d);var ai=c._pthread_self=()=>(ai=c._pthread_self=he.$a)(),ii=c._malloc=d=>(ii=c._malloc=he.ab)(d),or=c._free=d=>(or=c._free=he.bb)(d);c.__emscripten_tls_init=()=>(c.__emscripten_tls_init=he.cb)();var Ru=c.__emscripten_thread_init=(d,g,b,S,I,z)=>(Ru=c.__emscripten_thread_init=he.eb)(d,g,b,S,I,z);c.__emscripten_thread_crashed=()=>(c.__emscripten_thread_crashed=he.fb)();var Bu=(d,g,b,S,I)=>(Bu=he.gb)(d,g,b,S,I),ws=d=>(ws=he.hb)(d),bs=c.__emscripten_thread_exit=d=>(bs=c.__emscripten_thread_exit=he.ib)(d),Du=()=>(Du=he.jb)(),Nu=(d,g)=>(Nu=he.kb)(d,g),si=d=>(si=he.lb)(d),vs=d=>(vs=he.mb)(d),$s=()=>($s=he.nb)(),Fu=c.dynCall_ii=(d,g)=>(Fu=c.dynCall_ii=he.pb)(d,g),Lu=d=>(Lu=he.qb)(d),Uu=()=>(Uu=he.rb)(),Wu=d=>(Wu=he.sb)(d),Vu=()=>(Vu=he.tb)();c.___start_em_js=839500,c.___stop_em_js=839661;function r_(){var d=he;d=Object.assign({},d);var g=S=>I=>S(I)>>>0,b=S=>()=>S()>>>0;return d.wa=g(d.wa),d.$a=b(d.$a),d.ab=g(d.ab),d.emscripten_main_runtime_thread_id=b(d.emscripten_main_runtime_thread_id),d.mb=g(d.mb),d.nb=b(d.nb),d}c.wasmMemory=ne,c.stackSave=()=>$s(),c.stackRestore=d=>si(d),c.stackAlloc=d=>vs(d),c.keepRuntimeAlive=()=>0<ot,c.UTF8ToString=nt,c.stringToUTF8=Ht,c.lengthBytesUTF8=Et,c.ExitStatus=St,c.PThread=be;var oi;xr=function d(){oi||Gu(),oi||(xr=d)};function Gu(){if(!(0<fr))if(D)y(c),D||cn(ir),startWorker(c);else{if(c.preRun)for(typeof c.preRun=="function"&&(c.preRun=[c.preRun]);c.preRun.length;)hr.unshift(c.preRun.shift());cn(hr),0<fr||oi||(oi=!0,c.calledRun=!0,ae||(D||cn(ir),y(c),D||cn(ls)))}}return Gu(),v}})();typeof t=="object"&&typeof e=="object"?e.exports=r:typeof define=="function"&&define.amd&&define([],()=>r)}),K_=In((t,e)=>{e.exports='"use strict";var Module={},ENVIRONMENT_IS_NODE=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string";if(ENVIRONMENT_IS_NODE){var nodeWorkerThreads=require("worker_threads"),parentPort=nodeWorkerThreads.parentPort;parentPort.on("message",e=>onmessage({data:e}));var fs=require("fs"),vm=require("vm");Object.assign(global,{self:global,require,Module,location:{href:__filename},Worker:nodeWorkerThreads.Worker,importScripts:e=>vm.runInThisContext(fs.readFileSync(e,"utf8"),{filename:e}),postMessage:e=>parentPort.postMessage(e),performance:global.performance||{now:Date.now}})}var initializedJS=!1;function threadPrintErr(...e){var a=e.join(" ");if(ENVIRONMENT_IS_NODE){fs.writeSync(2,a+`\n`);return}console.error(a)}function threadAlert(...e){var a=e.join(" ");postMessage({cmd:"alert",text:a,threadId:Module._pthread_self()})}var err=threadPrintErr;self.alert=threadAlert,Module.instantiateWasm=(e,a)=>{var r=Module.wasmModule;Module.wasmModule=null;var t=new WebAssembly.Instance(r,e);return a(t)},self.onunhandledrejection=e=>{throw e.reason||e};function handleMessage(e){try{if(e.data.cmd==="load"){let r=[];self.onmessage=t=>r.push(t),self.startWorker=t=>{Module=t,postMessage({cmd:"loaded"});for(let s of r)handleMessage(s);self.onmessage=handleMessage},Module.wasmModule=e.data.wasmModule;for(const t of e.data.handlers)Module[t]=(...s)=>{postMessage({cmd:"callHandler",handler:t,args:s})};if(Module.wasmMemory=e.data.wasmMemory,Module.buffer=Module.wasmMemory.buffer,Module.ENVIRONMENT_IS_PTHREAD=!0,typeof e.data.urlOrBlob=="string")importScripts(e.data.urlOrBlob);else{var a=URL.createObjectURL(e.data.urlOrBlob);importScripts(a),URL.revokeObjectURL(a)}ortWasmThreaded(Module)}else if(e.data.cmd==="run"){Module.__emscripten_thread_init(e.data.pthread_ptr,0,0,1),Module.__emscripten_thread_mailbox_await(e.data.pthread_ptr),Module.establishStackSpace(),Module.PThread.receiveObjectTransfer(e.data),Module.PThread.threadInitTLS(),initializedJS||(initializedJS=!0);try{Module.invokeEntryPoint(e.data.start_routine,e.data.arg)}catch(r){if(r!="unwind")throw r}}else e.data.cmd==="cancel"?Module._pthread_self()&&Module.__emscripten_thread_exit(-1):e.data.target==="setimmediate"||(e.data.cmd==="checkMailbox"?initializedJS&&Module.checkMailbox():e.data.cmd&&(err(`worker.js received unknown command ${e.data.cmd}`),err(e.data)))}catch(r){throw Module.__emscripten_thread_crashed?.(),r}}self.onmessage=handleMessage;\n'}),Qu,Ju,Ts,di,Jn,Is,Zu,el,tl,ih,at,va=Z(()=>{Qu=V_(),Ju=q_(),di=!1,Jn=!1,Is=!1,Zu=t=>{if(t===1)return!1;if(typeof SharedArrayBuffer>"u")return typeof self<"u"&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+t+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),!1;typeof process<"u"&&process.versions&&process.versions.node&&console.warn("env.wasm.numThreads is set to "+t+", however, currently onnxruntime-web does not support multi-threads in Node.js. Please consider using onnxruntime-node for performance critical scenarios.");try{return typeof MessageChannel<"u"&&new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)),WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11]))}catch{return!1}},el=()=>{try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,10,30,1,28,0,65,0,253,15,253,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,253,186,1,26,11]))}catch{return!1}},tl=(t,e)=>t?e?"ort-wasm-simd-threaded.wasm":"ort-wasm-simd.wasm":e?"ort-wasm-threaded.wasm":"ort-wasm.wasm",ih=async t=>{if(di)return Promise.resolve();if(Jn)throw new Error("multiple calls to 'initializeWebAssembly()' detected.");if(Is)throw new Error("previous call to 'initializeWebAssembly()' failed.");Jn=!0;let e=t.initTimeout,r=t.numThreads,n=t.simd,a=Zu(r),i=n&&el(),s=t.wasmPaths,o=typeof s=="string"?s:void 0,u=tl(i,a),l=typeof s=="object"?s[u]:void 0,p=!1,f=[];if(e>0&&f.push(new Promise(m=>{setTimeout(()=>{p=!0,m()},e)})),f.push(new Promise((m,c)=>{let y=a?Ju:Qu,w={locateFile:(v,k)=>{if(a&&v.endsWith(".worker.js")&&typeof Blob<"u")return URL.createObjectURL(new Blob([K_()],{type:"text/javascript"}));if(v.endsWith(".wasm")){if(l)return l;let $=o??k;return u==="ort-wasm-simd.wasm"?$+"ort-wasm-simd.jsep.wasm":u==="ort-wasm-simd-threaded.wasm"?$+"ort-wasm-simd-threaded.jsep.wasm":$+u}return k+v}};if(a)if(w.numThreads=r,typeof Blob>"u")w.mainScriptUrlOrBlob=(void 0)(__dirname,"ort-wasm-threaded.js");else{let v=`var ortWasmThreaded=${y.toString()};`;w.mainScriptUrlOrBlob=new Blob([v],{type:"text/javascript"})}y(w).then(v=>{Jn=!1,di=!0,Ts=v,m()},v=>{Jn=!1,Is=!0,c(v)})})),await Promise.race(f),p)throw new Error(`WebAssembly backend initializing failed due to timeout: ${e}ms`)},at=()=>{if(di&&Ts)return Ts;throw new Error("WebAssembly is not initialized yet.")}}),tt,Bi,je,Vo=Z(()=>{va(),tt=(t,e)=>{let r=at(),n=r.lengthBytesUTF8(t)+1,a=r._malloc(n);return r.stringToUTF8(t,a,n),e.push(a),a},Bi=(t,e,r,n)=>{if(typeof t=="object"&&t!==null){if(r.has(t))throw new Error("Circular reference in options");r.add(t)}Object.entries(t).forEach(([a,i])=>{let s=e?e+a:a;if(typeof i=="object")Bi(i,s+".",r,n);else if(typeof i=="string"||typeof i=="number")n(s,i.toString());else if(typeof i=="boolean")n(s,i?"1":"0");else throw new Error(`Can't handle extra config type: ${typeof i}`)})},je=t=>{let e=at(),r=e.stackSave();try{let n=e.stackAlloc(8);e._OrtGetLastError(n,n+4);let a=e.HEAP32[n/4],i=e.HEAPU32[n/4+1],s=i?e.UTF8ToString(i):"";throw new Error(`${t} ERROR_CODE: ${a}, ERROR_MESSAGE: ${s}`)}finally{e.stackRestore(r)}}}),sh,Y_=Z(()=>{va(),Vo(),sh=t=>{let e=at(),r=0,n=[],a=t||{};try{if(t?.logSeverityLevel===void 0)a.logSeverityLevel=2;else if(typeof t.logSeverityLevel!="number"||!Number.isInteger(t.logSeverityLevel)||t.logSeverityLevel<0||t.logSeverityLevel>4)throw new Error(`log serverity level is not valid: ${t.logSeverityLevel}`);if(t?.logVerbosityLevel===void 0)a.logVerbosityLevel=0;else if(typeof t.logVerbosityLevel!="number"||!Number.isInteger(t.logVerbosityLevel))throw new Error(`log verbosity level is not valid: ${t.logVerbosityLevel}`);t?.terminate===void 0&&(a.terminate=!1);let i=0;return t?.tag!==void 0&&(i=tt(t.tag,n)),r=e._OrtCreateRunOptions(a.logSeverityLevel,a.logVerbosityLevel,!!a.terminate,i),r===0&&je("Can't create run options."),t?.extra!==void 0&&Bi(t.extra,"",new WeakSet,(s,o)=>{let u=tt(s,n),l=tt(o,n);e._OrtAddRunConfigEntry(r,u,l)!==0&&je(`Can't set a run config entry: ${s} - ${o}.`)}),[r,n]}catch(i){throw r!==0&&e._OrtReleaseRunOptions(r),n.forEach(s=>e._free(s)),i}}}),rl,nl,al,il,oh,X_=Z(()=>{va(),Vo(),rl=t=>{switch(t){case"disabled":return 0;case"basic":return 1;case"extended":return 2;case"all":return 99;default:throw new Error(`unsupported graph optimization level: ${t}`)}},nl=t=>{switch(t){case"sequential":return 0;case"parallel":return 1;default:throw new Error(`unsupported execution mode: ${t}`)}},al=t=>{t.extra||(t.extra={}),t.extra.session||(t.extra.session={});let e=t.extra.session;e.use_ort_model_bytes_directly||(e.use_ort_model_bytes_directly="1"),t.executionProviders&&t.executionProviders.some(r=>(typeof r=="string"?r:r.name)==="webgpu")&&(t.enableMemPattern=!1)},il=(t,e,r)=>{for(let n of e){let a=typeof n=="string"?n:n.name;switch(a){case"webnn":if(a="WEBNN",typeof n!="string"){let s=n;if(s?.deviceType){let o=tt("deviceType",r),u=tt(s.deviceType,r);at()._OrtAddSessionConfigEntry(t,o,u)!==0&&je(`Can't set a session config entry: 'deviceType' - ${s.deviceType}.`)}if(s?.numThreads){let o=s.numThreads;(typeof o!="number"||!Number.isInteger(o)||o<0)&&(o=0);let u=tt("numThreads",r),l=tt(o.toString(),r);at()._OrtAddSessionConfigEntry(t,u,l)!==0&&je(`Can't set a session config entry: 'numThreads' - ${s.numThreads}.`)}if(s?.powerPreference){let o=tt("powerPreference",r),u=tt(s.powerPreference,r);at()._OrtAddSessionConfigEntry(t,o,u)!==0&&je(`Can't set a session config entry: 'powerPreference' - ${s.powerPreference}.`)}}break;case"webgpu":if(a="JS",typeof n!="string"){let s=n;if(s?.preferredLayout){if(s.preferredLayout!=="NCHW"&&s.preferredLayout!=="NHWC")throw new Error(`preferredLayout must be either 'NCHW' or 'NHWC': ${s.preferredLayout}`);let o=tt("preferredLayout",r),u=tt(s.preferredLayout,r);at()._OrtAddSessionConfigEntry(t,o,u)!==0&&je(`Can't set a session config entry: 'preferredLayout' - ${s.preferredLayout}.`)}}break;case"wasm":case"cpu":continue;default:throw new Error(`not supported execution provider: ${a}`)}let i=tt(a,r);at()._OrtAppendExecutionProvider(t,i)!==0&&je(`Can't append execution provider: ${a}.`)}},oh=t=>{let e=at(),r=0,n=[],a=t||{};al(a);try{let i=rl(a.graphOptimizationLevel??"all"),s=nl(a.executionMode??"sequential"),o=typeof a.logId=="string"?tt(a.logId,n):0,u=a.logSeverityLevel??2;if(!Number.isInteger(u)||u<0||u>4)throw new Error(`log serverity level is not valid: ${u}`);let l=a.logVerbosityLevel??0;if(!Number.isInteger(l)||l<0||l>4)throw new Error(`log verbosity level is not valid: ${l}`);let p=typeof a.optimizedModelFilePath=="string"?tt(a.optimizedModelFilePath,n):0;if(r=e._OrtCreateSessionOptions(i,!!a.enableCpuMemArena,!!a.enableMemPattern,s,!!a.enableProfiling,0,o,u,l,p),r===0&&je("Can't create session options."),a.executionProviders&&il(r,a.executionProviders,n),a.enableGraphCapture!==void 0){if(typeof a.enableGraphCapture!="boolean")throw new Error(`enableGraphCapture must be a boolean value: ${a.enableGraphCapture}`);let f=tt("enableGraphCapture",n),m=tt(a.enableGraphCapture.toString(),n);e._OrtAddSessionConfigEntry(r,f,m)!==0&&je(`Can't set a session config entry: 'enableGraphCapture' - ${a.enableGraphCapture}.`)}if(a.freeDimensionOverrides)for(let[f,m]of Object.entries(a.freeDimensionOverrides)){if(typeof f!="string")throw new Error(`free dimension override name must be a string: ${f}`);if(typeof m!="number"||!Number.isInteger(m)||m<0)throw new Error(`free dimension override value must be a non-negative integer: ${m}`);let c=tt(f,n);e._OrtAddFreeDimensionOverride(r,c,m)!==0&&je(`Can't set a free dimension override: ${f} - ${m}.`)}return a.extra!==void 0&&Bi(a.extra,"",new WeakSet,(f,m)=>{let c=tt(f,n),y=tt(m,n);e._OrtAddSessionConfigEntry(r,c,y)!==0&&je(`Can't set a session config entry: ${f} - ${m}.`)}),[r,n]}catch(i){throw r!==0&&e._OrtReleaseSessionOptions(r),n.forEach(s=>e._free(s)),i}}}),mo,qr,fa,Go,Di,Ho,go,$e=Z(()=>{mo=t=>{switch(t){case"int8":return 3;case"uint8":return 2;case"bool":return 9;case"int16":return 5;case"uint16":return 4;case"int32":return 6;case"uint32":return 12;case"float16":return 10;case"float32":return 1;case"float64":return 11;case"string":return 8;case"int64":return 7;case"uint64":return 13;default:throw new Error(`unsupported data type: ${t}`)}},qr=t=>{switch(t){case 3:return"int8";case 2:return"uint8";case 9:return"bool";case 5:return"int16";case 4:return"uint16";case 6:return"int32";case 12:return"uint32";case 10:return"float16";case 1:return"float32";case 11:return"float64";case 8:return"string";case 7:return"int64";case 13:return"uint64";default:throw new Error(`unsupported data type: ${t}`)}},fa=t=>[void 0,4,1,1,2,2,4,8,void 0,1,2,8,4,8,void 0,void 0,void 0][t],Go=t=>{switch(t){case"float16":return typeof Float16Array<"u"&&Float16Array.from?Float16Array:Uint16Array;case"float32":return Float32Array;case"uint8":return Uint8Array;case"int8":return Int8Array;case"uint16":return Uint16Array;case"int16":return Int16Array;case"int32":return Int32Array;case"bool":return Uint8Array;case"float64":return Float64Array;case"uint32":return Uint32Array;case"int64":return BigInt64Array;case"uint64":return BigUint64Array;default:throw new Error(`unsupported type: ${t}`)}},Di=t=>{switch(t){case"verbose":return 0;case"info":return 1;case"warning":return 2;case"error":return 3;case"fatal":return 4;default:throw new Error(`unsupported logging level: ${t}`)}},Ho=t=>t==="float32"||t==="float16"||t==="int32"||t==="int64"||t==="uint32"||t==="uint8"||t==="bool",go=t=>{switch(t){case"none":return 0;case"cpu":return 1;case"cpu-pinned":return 2;case"texture":return 3;case"gpu-buffer":return 4;default:throw new Error(`unsupported data location: ${t}`)}}}),Ni,uh=Z(()=>{Ni=async t=>{if(typeof t=="string")if(typeof process<"u"&&process.versions&&process.versions.node)try{return new Uint8Array(await(void 0)(t))}catch(e){if(e.code==="ERR_FS_FILE_TOO_LARGE"){let r=(void 0)(t),n=[];for await(let a of r)n.push(a);return new Uint8Array(Buffer.concat(n))}throw e}else{let e=await fetch(t);if(!e.ok)throw new Error(`failed to load external data file: ${t}`);let r=e.headers.get("Content-Length"),n=r?parseInt(r,10):0;if(n<1073741824)return new Uint8Array(await e.arrayBuffer());{if(!e.body)throw new Error(`failed to load external data file: ${t}, no response body.`);let a=e.body.getReader(),i;try{i=new ArrayBuffer(n)}catch(o){if(o instanceof RangeError){let u=Math.ceil(n/65536);i=new WebAssembly.Memory({initial:u,maximum:u}).buffer}else throw o}let s=0;for(;;){let{done:o,value:u}=await a.read();if(o)break;let l=u.byteLength;new Uint8Array(i,s,l).set(u),s+=l}return new Uint8Array(i,0,n)}}else return t instanceof Blob?new Uint8Array(await t.arrayBuffer()):t instanceof Uint8Array?t:new Uint8Array(t)}}),sl,ol,ul,ll,lh,dl,rt,rn=Z(()=>{$e(),sl=["V","I","W","E","F"],ol=(t,e)=>{console.log(`[${sl[t]},${new Date().toISOString()}]${e}`)},lh=(t,e)=>{ul=t,ll=e},dl=(t,e)=>{let r=Di(t),n=Di(ul);r>=n&&ol(r,typeof e=="function"?e():e)},rt=(...t)=>{ll&&dl(...t)}}),dh,Q_=Z(()=>{$e(),dh=(t,e)=>new(Go(e))(t)}),jo=Z(()=>{}),As,ci,pi,cl,pl,Ms,_o,hl,ch,J_=Z(()=>{rn(),jo(),As=new Map([[64,250],[128,200],[256,200],[512,200],[2048,230],[4096,200],[8192,50],[16384,50],[32768,50],[65536,50],[131072,50],[262144,50],[524288,50],[1048576,50],[2097152,30],[4194304,20],[8388608,10],[12582912,10],[16777216,10],[26214400,15],[33554432,22],[44236800,2],[58982400,6],[67108864,6],[134217728,6],[167772160,6]]),ci=[],pi=t=>Math.ceil(t/16)*16,cl=t=>{for(let e=0;e<ci.length;e++){let r=ci[e];if(t<=r)return r}return Math.ceil(t/16)*16},pl=1,Ms=()=>pl++,_o=async(t,e,r,n)=>{let a=pi(r),i=t.device.createBuffer({size:a,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let s=t.getCommandEncoder();t.endComputePass(),s.copyBufferToBuffer(e,0,i,0,a),t.flush(),await i.mapAsync(GPUMapMode.READ);let o=i.getMappedRange();if(n){let u=n();return u.set(new Uint8Array(o,0,r)),u}else return new Uint8Array(o.slice(0,r))}finally{i.destroy()}},hl=class{constructor(t){this.backend=t,this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.buffersForUploadingPending=[],this.buffersPending=[],this.externalBuffers=new Map,this.capturedPendingBuffers=new Map;for(let[e]of As)ci.push(e),this.freeBuffers.set(e,[]),this.freeUniformBuffers.set(e,[])}upload(t,e){let r=e.buffer,n=e.byteOffset,a=e.byteLength,i=pi(a),s=this.storageCache.get(t);if(!s)throw new Error("gpu data for uploading does not exist");if(s.originalSize!==a)throw new Error(`inconsistent data size. gpu data size=${s.originalSize}, data size=${a}`);let o=this.backend.device.createBuffer({mappedAtCreation:!0,size:i,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),u=o.getMappedRange();new Uint8Array(u).set(new Uint8Array(r,n,a)),o.unmap();let l=this.backend.getCommandEncoder();this.backend.endComputePass(),l.copyBufferToBuffer(o,0,s.gpuData.buffer,0,i),rt("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${t})`),this.buffersForUploadingPending.push(o)}memcpy(t,e){let r=this.storageCache.get(t);if(!r)throw new Error("source gpu data for memcpy does not exist");let n=this.storageCache.get(e);if(!n)throw new Error("destination gpu data for memcpy does not exist");if(r.originalSize!==n.originalSize)throw new Error("inconsistent source and destination gpu data size");let a=pi(r.originalSize),i=this.backend.getCommandEncoder();this.backend.endComputePass(),i.copyBufferToBuffer(r.gpuData.buffer,0,n.gpuData.buffer,0,a)}registerExternalBuffer(t,e,r){let n;if(r){if(n=this.externalBuffers.get(r),n===void 0)throw new Error("previous buffer is not registered");if(t===r)return rt("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${e}) => id=${n}, buffer is the same, skip.`),n;if(this.backend.capturedCommandList.has(this.backend.currentSessionId))throw new Error(`Registering a different external buffer under graph capture mode is not supported yet.
             Please use the previous external buffer!`);this.externalBuffers.delete(r)}else n=Ms();return this.storageCache.set(n,{gpuData:{id:n,type:0,buffer:t},originalSize:e}),this.externalBuffers.set(t,n),rt("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${e}) => id=${n}, registered.`),n}unregisterExternalBuffer(t){let e=this.externalBuffers.get(t);e!==void 0&&(this.storageCache.delete(e),this.externalBuffers.delete(t),rt("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${e}`))}create(t,e=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let r=cl(t),n,a=(e&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,i=(e&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM;if(a||i){let o=(a?this.freeBuffers:this.freeUniformBuffers).get(r);o?o.length>0?n=o.pop():n=this.backend.device.createBuffer({size:r,usage:e}):n=this.backend.device.createBuffer({size:r,usage:e})}else n=this.backend.device.createBuffer({size:r,usage:e});let s={id:Ms(),type:0,buffer:n};return this.storageCache.set(s.id,{gpuData:s,originalSize:t}),rt("verbose",()=>`[WebGPU] GpuDataManager.create(size=${t}) => id=${s.id}`),s}get(t){return this.storageCache.get(t)?.gpuData}release(t){let e=this.storageCache.get(t);if(!e)throw new Error("releasing data does not exist");return rt("verbose",()=>`[WebGPU] GpuDataManager.release(id=${t}), gpuDataId=${e.gpuData.id}`),this.storageCache.delete(t),this.buffersPending.push(e.gpuData.buffer),e.originalSize}async download(t,e){let r=this.storageCache.get(t);if(!r)throw new Error("data does not exist");await _o(this.backend,r.gpuData.buffer,r.originalSize,e)}refreshPendingBuffers(){for(let t of this.buffersForUploadingPending)t.destroy();if(this.buffersForUploadingPending=[],this.buffersPending.length!==0)if(this.backend.sessionStatus==="default"){for(let t of this.buffersPending){let e=As.get(t.size);if((t.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let r=this.freeBuffers.get(t.size)||[];e===void 0||r.length>=e?t.destroy():r.push(t)}else if((t.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let r=this.freeUniformBuffers.get(t.size)||[];e===void 0||r.length>=e?t.destroy():r.push(t)}else t.destroy()}this.buffersPending=[]}else{let t=this.capturedPendingBuffers.get(this.backend.currentSessionId);t||(t=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,t));for(let e of this.buffersPending)t.push(e);this.buffersPending=[]}}dispose(){this.freeBuffers.forEach(t=>{t.forEach(e=>{e.destroy()})}),this.freeUniformBuffers.forEach(t=>{t.forEach(e=>{e.destroy()})}),this.storageCache.forEach(t=>{t.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(t=>{t.forEach(e=>{e.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onReleaseSession(t){let e=this.capturedPendingBuffers.get(t);e&&(e.forEach(r=>{r.destroy()}),this.capturedPendingBuffers.delete(t))}},ch=(...t)=>new hl(...t)}),fl,qe,lt=Z(()=>{fl=class{constructor(t){Object.assign(this,t)}get cacheKey(){return this.key||(this.key=Object.getOwnPropertyNames(this).sort().map(t=>`${this[t]}`).join(";")),this.key}},qe=t=>new fl(t)}),ml,kn,Y,Fi,ph,qo,Ko,Ae=Z(()=>{ml=class{static calcMatMulShape(t,e){return t[1]!==e[0]?void 0:[t[0],e[1]]}},kn=class{static calcShape(t,e,r=!1){let n=t.length,a=e.length;if(n===0)return e;if(a===0)return t;let i=Math.max(t.length,e.length),s=new Array(i);if(r){if(n<2||a<2)return;let o=ml.calcMatMulShape([t[n-2],t[n-1]],[e[a-2],e[a-1]]);if(o===void 0)return;[s[i-2],s[i-1]]=o}for(let o=r?3:1;o<=i;o++){let u=n-o<0?1:t[n-o],l=a-o<0?1:e[a-o];if(u!==l&&u>1&&l>1)return;let p=Math.max(u,l);if(u&&l)s[i-o]=Math.max(u,l);else{if(p>1)return;s[i-o]=0}}return s}static isValidBroadcast(t,e){let r=t.length,n=e.length;if(r>n)return!1;for(let a=1;a<=r;a++)if(t[r-a]!==1&&t[r-a]!==e[n-a])return!1;return!0}},Y=class Ei{static size(e){return Ei.getSizeFromDimensionRange(e,0,e.length)}static convertShape(e,r=4){let n=e.length;if(n===0)return[];let a=new Array(n),i=n-1;for(;i>=0;){if(e[i]%r===0){a[i]=e[i]/r;break}if(r%e[i]!==0)throw new Error("cannot convert shape");a[i]=1,r/=e[i],i--}for(i--;i>=0;i--)a[i]=e[i];return a}static sizeFromDimension(e,r){if(r<0||r>e.length)throw new Error(`invalid dimension of ${r} for sizeFromDimension as Tensor has ${e.length} dimensions.`);return Ei.getSizeFromDimensionRange(e,r,e.length)}static sizeToDimension(e,r){if(r<0||r>e.length)throw new Error(`invalid dimension of ${r} for sizeToDimension as Tensor has ${e.length} dimensions.`);return Ei.getSizeFromDimensionRange(e,0,r)}static getSizeFromDimensionRange(e,r,n){let a=1;for(let i=r;i<n;i++){if(e[i]<0)throw new Error("cannot get valid size from specified dimension range. Most likely the range contains negative values in them.");a*=e[i]}return a}static computeStrides(e){let r=e.length;if(r===0)return[];if(r===1)return[1];let n=new Array(r);n[r-1]=1,n[r-2]=e[r-1];for(let a=r-3;a>=0;--a)n[a]=n[a+1]*e[a+1];return n}static normalizeAxis(e,r){if(e<-r&&e>=r)throw new Error("unsupported axis for this operation.");return e<0?e+r:e}static normalizeAxes(e,r){return e.map(n=>this.normalizeAxis(n,r??e.length))}static sortBasedOnPerm(e,r){return r?r.map(n=>e[n]):e.slice().reverse()}static padShape(e,r){let n=e.length;return e.map((a,i)=>a+r[i]+r[i+n])}static areEqual(e,r){return e.length!==r.length?!1:e.every((n,a)=>n===r[a])}},Fi=class oa{static adjustPoolAttributes(e,r,n,a,i,s){if(!e&&n.length!==r.length-2)throw new Error("length of specified kernel shapes should be 2 less than length of input dimensions");if(e)for(let o=0;o<r.length-2;o++)o>=n.length?n.push(r[o+2]):n[o]=r[o+2];for(let o=0;o<n.length;o++)if(o<a.length){if(a[o]<0)throw new Error("strides should be greater than or equal to 1")}else a.push(1);for(let o=0;o<n.length;o++)if(o<i.length){if(i[o]<0)throw new Error("dilations should be greater than or equal to 1")}else i.push(1);for(let o=0;o<n.length*2;o++)if(o<s.length){if(s[o]<0)throw new Error("pad should be greater than or equal to 1")}else s.push(0);for(let o=0;o<n.length;o++){if(n[o]<=0)throw new Error("kernel shapes need to be greater than 0");if(s[o]>=n[o]||s[o+n.length]>=n[o])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(e,r,n,a,i,s,o){if(o){if(i.length!==2*(e.length-2))throw new Error("length of pads should be twice the length of data dimensions");if(r.length!==e.length-2)throw new Error("length of strides should be the length of data dimensions");if(a.length!==e.length-2)throw new Error("length of kernel shapes should be the length of data dimensions");for(let u=0;u<e.length-2;u++)oa.adjustPadAndReturnShape(e[u+(s?1:2)],r[u],n[u],a[u],i,u,u+e.length-2,o)}}static computePoolOutputShape(e,r,n,a,i,s,o){if(r.length<=0)throw new Error("input shape must be of size greater than 0");let u=[r[0],r[1]];return oa.computeShapeHelper(e,r,u,n,a,i,s,o),u}static computeConvOutputShape(e,r,n,a,i,s,o){if(e.length<=0||r.length<=0)throw new Error("invalid input tensor dims or invalid filter tensor dims");let u=[e[0],r[0]];return oa.computeShapeHelper(!1,e,u,n,a,i,s,o),u}static computeShapeHelper(e,r,n,a,i,s,o,u){if(e)for(let l=0;l<r.length-2;l++)n.push(1);else for(let l=0;l<r.length-2;l++)n.push(oa.adjustPadAndReturnShape(r[l+2],a[l],i[l],s[l],o,l,l+r.length-2,u))}static adjustPadAndReturnShape(e,r,n,a,i,s,o,u){let l=n*(a-1)+1;if(u&&u!=="NOTSET")switch(u){case"VALID":return i[s]=0,i[o]=0,Math.floor((e-l)/r+1);case"SAME_LOWER":case"SAME_UPPER":if(n!==1)throw new Error("Dilation not supported for SAME_UPPER or SAME_LOWER");{let p=((e+r-1)/r-1)*r+a-e;return i[s]=Math.floor(u==="SAME_LOWER"?(p+1)/2:p/2),i[o]=p-i[s],Math.floor((e+p-a)/r+1)}default:throw new Error("Unsupported AutoPad type")}else return Math.floor((e+i[s]+i[o]-l)/r+1)}},ph=class{static getShapeOfGemmResult(t,e,r,n,a){if(t.length!==2||r.length!==2)throw new Error("shape need to be of size 2");let i,s,o;e?(i=t[1],s=t[0]):(i=t[0],s=t[1]);let u=-1;if(n?(o=r[0],u=1):(o=r[1],u=0),r[u]!==s)throw new Error("dimension mismatch");if(i<=0||o<=0||s<=0)throw new Error("invalid shape specified");if(a&&!kn.isValidBroadcast(a,[i,o]))throw new Error("gemm: invalid bias shape for broadcast");return[i,o,s]}},qo=-34028234663852886e22,Ko=34028234663852886e22}),En,hi,pt,Tt,ye,it,yr,xn,Mr,Se,fi,Q,ge,Yo,gl,hh,ma,Te=Z(()=>{$e(),Ae(),En=64,hi=(t,e)=>{if(e===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(t){case 10:return e>1?`vec${e}<f16>`:"f16";case 1:return e>1?`vec${e}<f32>`:"f32";case 6:return e>1?`vec${e}<i32>`:"i32";case 12:return e>1?`vec${e}<u32>`:"u32";case 7:if(e>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","i32"];case 13:if(e>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","u32"];case 9:if(e!==4)throw new Error("bool must be vec4");return["u32","vec4<bool>"];default:throw new Error(`Unknown data type: ${t}`)}},pt=(t,e=1)=>{let r=hi(t,e);return typeof r=="string"?r:r[0]},Tt=(t,e=1)=>{let r=hi(t,e);return typeof r=="string"?r:r[1]},ye=(...t)=>{let e=[];return t.forEach(r=>{r.length!==0&&e.push({type:12,data:r},{type:12,data:Y.computeStrides(r)})}),e},it=t=>t%4===0?4:t%2===0?2:1,yr=(t="f32",e,r="0")=>!e||e===1?`${t}(${r})`:`vec${e}<${t}>(${r})`,xn=(t,e,r)=>t==="f32"?r:e===1?`f32(${r})`:`vec${e}<f32>(${r})`,Mr=(t,e)=>e===4?`(${t}.x + ${t}.y + ${t}.z + ${t}.w)`:e===2?`(${t}.x + ${t}.y)`:e===3?`(${t}.x + ${t}.y + ${t}.z)`:t,Se=(t,e,r,n)=>t.startsWith("uniforms.")&&r>4?typeof e=="string"?n==="f16"?`${t}[(${e}) / 8][(${e}) % 8 / 4][(${e}) % 8 % 4]`:`${t}[(${e}) / 4][(${e}) % 4]`:n==="f16"?`${t}[${Math.floor(e/8)}][${Math.floor(e%8/4)}][${e%8%4}]`:`${t}[${Math.floor(e/4)}][${e%4}]`:r>1?`${t}[${e}]`:t,fi=(t,e,r,n,a)=>{let i=typeof r=="number",s=i?r:r.length,o=[...new Array(s).keys()],u=s<2?"u32":s<=4?`vec${s}<u32>`:`array<u32, ${s}>`,l=hi(e,a),p=typeof l=="string"?l:l[1],f=typeof l=="string"?l:l[0],m={indices:u,value:p,storage:f,tensor:e},c=N=>typeof N=="string"?N:`${N}u`,y={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},w=i?"uniforms.":"",v=`${w}${t}_shape`,k=`${w}${t}_strides`,$="";for(let N=0;N<s-1;N++)$+=`
    let dim${N} = current / ${Se(k,N,s)};
    let rest${N} = current % ${Se(k,N,s)};
    indices[${N}] = dim${N};
    current = rest${N};
    `;$+=`indices[${s-1}] = current;`;let C=s<2?"":`
  fn o2i_${t}(offset: u32) -> ${m.indices} {
    var indices: ${m.indices};
    var current = offset;
    ${$}
    return indices;
  }`,T=N=>(y.offsetToIndices=!0,s<2?N:`o2i_${t}(${N})`),A=[];if(s>=2)for(let N=s-1;N>=0;N--)A.push(`${Se(k,N,s)} * (indices[${N}])`);let B=s<2?"":`
  fn i2o_${t}(indices: ${m.indices}) -> u32 {
    return ${A.join("+")};
  }`,R=N=>(y.indicesToOffset=!0,s<2?N:`i2o_${t}(${N})`),D=(...N)=>s===0?"0u":`${m.indices}(${N.map(c).join(",")})`,K=(N,ae)=>s<2?`${N}`:`${Se(N,ae,s)}`,j=(N,ae,fe)=>s<2?`${N}=${fe};`:`${Se(N,ae,s)}=${fe};`,ie={},te=(N,ae)=>{y.broadcastedIndicesToOffset=!0;let fe=`${ae.name}broadcastedIndicesTo${t}Offset`;if(fe in ie)return`${fe}(${N})`;let Ce=[];for(let Be=s-1;Be>=0;Be--){let Ke=ae.indicesGet("outputIndices",Be+ae.rank-s);Ce.push(`${K(k,Be)} * (${Ke} % ${K(v,Be)})`)}return ie[fe]=`fn ${fe}(outputIndices: ${ae.type.indices}) -> u32 {
             return ${Ce.length>0?Ce.join("+"):"0u"};
           }`,`${fe}(${N})`},oe=(N,ae)=>(()=>{if(m.storage===m.value)return`${t}[${N}]=${ae};`;if(m.storage==="vec2<u32>"&&m.value==="i32")return`${t}[${N}]=vec2<u32>(u32(${ae}), select(0u, 0xFFFFFFFFu, ${ae} < 0));`;if(m.storage==="vec2<u32>"&&m.value==="u32")return`${t}[${N}]=vec2<u32>(u32(${ae}), 0u);`;if(m.storage==="u32"&&m.value==="vec4<bool>")return`${t}[${N}]=dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(${ae}));`;throw new Error(`not supported combination of storage type ${m.storage} and value type ${m.value} yet`)})(),re=N=>(()=>{if(m.storage===m.value)return`${t}[${N}]`;if(m.storage==="vec2<u32>"&&m.value==="i32")return`i32(${t}[${N}].x)`;if(m.storage==="vec2<u32>"&&m.value==="u32")return`u32(${t}[${N}].x)`;if(m.storage==="u32"&&m.value==="vec4<bool>")return`vec4<bool>(bool(${t}[${N}] & 0xFFu), bool(${t}[${N}] & 0xFF00u), bool(${t}[${N}] & 0xFF0000u), bool(${t}[${N}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${m.storage} and value type ${m.value} yet`)})(),M=s<2?"":`
  fn get_${t}ByIndices(indices: ${m.indices}) -> ${p} {
    return ${re(`i2o_${t}(indices)`)};
  }`,P=s<2?"":(()=>{let N=o.map(fe=>`d${fe}: u32`).join(", "),ae=o.map(fe=>`d${fe}`).join(", ");return`
  fn get_${t}(${N}) -> ${p} {
    return get_${t}ByIndices(${D(ae)});
  }`})(),H=(...N)=>{if(N.length!==s)throw new Error(`indices length must be ${s}`);let ae=N.map(c).join(",");return s===0?re("0u"):s===1?re(ae[0]):(y.get=!0,y.getByIndices=!0,y.indicesToOffset=!0,`get_${t}(${ae})`)},le=N=>s<2?re(N):(y.getByIndices=!0,y.indicesToOffset=!0,`get_${t}ByIndices(${N})`),G=s<2?"":`
  fn set_${t}ByIndices(indices: ${m.indices}, value: ${p}) {
    ${oe(`i2o_${t}(indices)`,"value")}
  }`,ne=s<2?"":(()=>{let N=o.map(fe=>`d${fe}: u32`).join(", "),ae=o.map(fe=>`d${fe}`).join(", ");return`
  fn set_${t}(${N}, value: ${p}) {
    set_${t}ByIndices(${D(ae)}, value);
  }`})();return{impl:()=>{let N=[],ae=!1;return y.offsetToIndices&&(N.push(C),ae=!0),y.indicesToOffset&&(N.push(B),ae=!0),y.broadcastedIndicesToOffset&&(Object.values(ie).forEach(fe=>N.push(fe)),ae=!0),y.set&&(N.push(ne),ae=!0),y.setByIndices&&(N.push(G),ae=!0),y.get&&(N.push(P),ae=!0),y.getByIndices&&(N.push(M),ae=!0),!i&&ae&&N.unshift(`const ${v} = ${m.indices}(${r.join(",")});`,`const ${k} = ${m.indices}(${Y.computeStrides(r).join(",")});`),N.join(`
`)},type:m,offsetToIndices:T,indicesToOffset:R,broadcastedIndicesToOffset:te,indices:D,indicesGet:K,indicesSet:j,set:(...N)=>{if(N.length!==s+1)throw new Error(`indices length must be ${s}`);let ae=N[s];if(typeof ae!="string")throw new Error("value must be string");let fe=N.slice(0,s).map(c).join(",");return s===0?oe("0u",ae):s===1?oe(fe[0],ae):(y.set=!0,y.setByIndices=!0,y.indicesToOffset=!0,`set_${t}(${fe}, ${ae})`)},setByOffset:oe,setByIndices:(N,ae)=>s<2?oe(N,ae):(y.setByIndices=!0,y.indicesToOffset=!0,`set_${t}ByIndices(${N}, ${ae});`),get:H,getByOffset:re,getByIndices:le,usage:n,name:t,strides:k,shape:v,rank:s}},Q=(t,e,r,n=1)=>fi(t,e,r,"input",n),ge=(t,e,r,n=1)=>fi(t,e,r,"output",n),Yo=(t,e,r,n=1)=>fi(t,e,r,"internal",n),gl=class{constructor(t,e){this.normalizedDispatchGroup=t,this.limits=e,this.internalVariables=[],this.variables=[],this.uniforms=[],this.variableIndex=0}guardAgainstOutOfBoundsWorkgroupSizes(t){return`if (global_idx >= ${typeof t=="number"?`${t}u`:t}) { return; }`}mainStart(t=En){let e=typeof t=="number"?t:t[0],r=typeof t=="number"?1:t[1],n=typeof t=="number"?1:t[2];if(e>this.limits.maxComputeWorkgroupSizeX||r>this.limits.maxComputeWorkgroupSizeY||n>this.limits.maxComputeWorkgroupSizeZ)throw new Error(`workgroup size [${e}, ${r}, ${n}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(e*r*n>this.limits.maxComputeInvocationsPerWorkgroup)throw new Error(`workgroup size [${e}, ${r}, ${n}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let a=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,i=a?`@builtin(global_invocation_id) global_id : vec3<u32>,
    @builtin(workgroup_id) workgroup_id : vec3<u32>,
    @builtin(local_invocation_id) local_id : vec3<u32>`:`@builtin(global_invocation_id) global_id : vec3<u32>,
                                             @builtin(local_invocation_id) local_id : vec3<u32>,
    @builtin(local_invocation_index) local_idx : u32,
    @builtin(workgroup_id) workgroup_id : vec3<u32>,
    @builtin(num_workgroups) num_workgroups : vec3<u32>`,s=a?"let global_idx = global_id.x; let local_idx = local_id.x;":`let global_idx = (workgroup_id.z * num_workgroups[0] * num_workgroups[1] +
          workgroup_id.y * num_workgroups[0] + workgroup_id.x) * ${e*r*n}u + local_idx;`;return`@compute @workgroup_size(${e}, ${r}, ${n})
  fn main(${i}) {
    ${s}
  `}appendVariableUniforms(t){t.rank!==0&&(t.shape.startsWith("uniforms.")&&this.uniforms.push({name:t.shape.replace("uniforms.",""),type:"u32",length:t.rank}),t.strides.startsWith("uniforms.")&&this.uniforms.push({name:t.strides.replace("uniforms.",""),type:"u32",length:t.rank}))}declareVariable(t,e){if(t.usage==="internal")throw new Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(t),this.appendVariableUniforms(t);let r=t.usage==="input"?"read":"read_write",n=t.type.storage;return`@group(0) @binding(${e}) var<storage, ${r}> ${t.name}: array<${n}>;`}declareVariables(...t){return t.map(e=>this.declareVariable(e,this.variableIndex++)).join(`
`)}registerInternalVariable(t){if(t.usage!=="internal")throw new Error("cannot use input or output variable with registerInternalVariable(). use declareVariables() instead.");this.internalVariables.push(t),this.appendVariableUniforms(t)}registerInternalVariables(...t){return t.forEach(e=>this.registerInternalVariable(e)),this}registerUniform(t,e,r=1){return this.uniforms.push({name:t,type:e,length:r}),this}registerUniforms(t){return this.uniforms=this.uniforms.concat(t),this}uniformDeclaration(){if(this.uniforms.length===0)return"";let t=[];for(let{name:e,type:r,length:n}of this.uniforms)if(n&&n>4)r==="f16"?t.push(`@align(16) ${e}:array<mat2x4<${r}>, ${Math.ceil(n/8)}>`):t.push(`${e}:array<vec4<${r}>, ${Math.ceil(n/4)}>`);else{let a=n==null||n===1?r:`vec${n}<${r}>`;t.push(`${e}:${a}`)}return`
      struct Uniforms { ${t.join(", ")} };
      @group(0) @binding(${this.variableIndex}) var<uniform> uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(t=>t.impl()).join(`
`)+this.internalVariables.map(t=>t.impl()).join(`
`)}get variablesInfo(){if(this.uniforms.length===0)return;let t=e=>[12,10,1,6][["u32","f16","f32","i32"].indexOf(e)];return this.uniforms.map(e=>[t(e.type),e.length??1])}},hh=(t,e)=>new gl(t,e),ma=(t,e)=>{let r=t.length,n=[];for(let a=0;a<r;a++){let i=r-1-a,s=t[i]||1;(e[e.length-1-a]||1)>1&&s===1&&n.unshift(i)}return n}}),_l,Os,yl,wl,br,fh,mh,$a=Z(()=>{$e(),Ae(),lt(),Te(),_l=t=>{if(!t||t.length!==1)throw new Error("Transpose requires 1 input.")},Os=(t,e)=>e&&e.length!==t?[...new Array(t).keys()].reverse():e,yl=(t,e)=>Y.sortBasedOnPerm(t,Os(t.length,e)),wl=(t,e,r,n)=>{let a=[];a.push(`fn perm(i: ${n.type.indices}) -> ${r.type.indices} {
    var a: ${r.type.indices};`);for(let i=0;i<e;++i)a.push(r.indicesSet("a",t[i],`i[${i}]`));return a.push("return a;}"),a.join(`
`)},br=(t,e)=>{let r=t.dataType,n=t.dims.length,a=Os(n,e),i=yl(t.dims,a),s=ge("output",r,i.length),o=Q("a",r,n),u=l=>`
  ${l.registerUniform("output_size","u32").declareVariables(o,s)}

  ${wl(a,n,o,s)}

  ${l.mainStart()}
    ${l.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}

    let indices = ${s.offsetToIndices("global_idx")};
    let aIndices = perm(indices);

    ${s.setByOffset("global_idx",o.getByIndices("aIndices"))}
  }`;return{name:"Transpose",shaderCache:{hint:`${e}`,inputDependencies:["rank"]},getRunData:l=>{let p=Y.size(i);return{outputs:[{dims:i,dataType:l[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:[{type:12,data:p},...ye(l[0].dims,i)]}},getShaderSource:u}},fh=(t,e)=>{_l(t.inputs),t.compute(br(t.inputs[0],e.perm))},mh=t=>qe({perm:t.perm})}),bl,vl,$l,xl,Sl,kl,El,Cl,Tl,Il,Kt,gh,_h,yh,wh,bh,vh,$h,xh,Sh,kh,Z_=Z(()=>{$e(),Ae(),Te(),Xo(),$a(),bl={max:"select(bestValue, candidate, candidate > bestValue)",min:"select(bestValue, candidate, candidate < bestValue)",mean:"bestValue + candidate",sum:"bestValue + candidate",prod:"bestValue * candidate",sumSquare:"bestValue + candidate * candidate",logSumExp:"bestValue + exp(candidate)",l1:"bestValue + abs(candidate)",l2:"bestValue + candidate * candidate",logSum:"bestValue + candidate"},vl={max:"select(bestValue, candidate, candidate > bestValue)",min:"select(bestValue, candidate, candidate < bestValue)",mean:"bestValue + candidate",sum:"bestValue + candidate",prod:"bestValue * candidate",sumSquare:"bestValue + candidate",logSumExp:"bestValue + candidate",l1:"bestValue + candidate",l2:"bestValue + candidate",logSum:"bestValue + candidate"},$l={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},xl={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},Sl=(t,e)=>{let r=[];for(let n=e-t;n<e;++n)r.push(n);return r},kl=(t,e)=>{let r=[],n=t.length;for(let i=0;i<n;i++)e.indexOf(i)===-1&&r.push(t[i]);let a=e.map(i=>t[i]);return[r,a]},El=(t,e)=>{let r=t.length+e.length,n=[],a=0;for(let i=0;i<r;i++)e.indexOf(i)===-1?n.push(t[a++]):n.push(1);return n},Cl=(t,e)=>{for(let r=0;r<t.length;++r)if(t[t.length-r-1]!==e-1-r)return!1;return!0},Tl=(t,e)=>{let r=[];if(!Cl(t,e)){for(let n=0;n<e;++n)t.indexOf(n)===-1&&r.push(n);t.forEach(n=>r.push(n))}return r},Il=(t,e,r,n,a,i,s)=>{let o=r[0].dims,u=Y.size(i),l=Y.size(s),p=Q("_A",r[0].dataType,o),f=ge("output",a,i),m=32,c=`
          var<workgroup> aBestValues : array<f32, ${m}>;
       `;return{name:t,shaderCache:e,getShaderSource:y=>`
        ${y.registerUniform("reduceSize","u32").declareVariables(p,f)}
        ${c}
        fn DIV_CEIL(a : u32, b : u32) -> u32 {
          return ((a - 1u) / b + 1u);
         }
         ${y.mainStart(m)}

          let outputIndex = global_idx / ${m};
          let offset = outputIndex * uniforms.reduceSize;

          var bestValue = f32(${$l[n]});
          let Length = uniforms.reduceSize;
          for (var k = local_idx; k < Length; k = k + ${m}) {
           let candidate = f32(${p.getByOffset("offset + k")});
           bestValue = ${bl[n]};
          }
          aBestValues[local_idx] = bestValue;
          workgroupBarrier();

         var reduceSize = min(Length, ${m}u);
         for (var currentSize = reduceSize / 2u; reduceSize > 1u;
             currentSize = reduceSize / 2u) {
           let interval = DIV_CEIL(reduceSize, 2u);
           if (local_idx < currentSize) {
            let candidate = aBestValues[local_idx + interval];
            bestValue = ${vl[n]};
            aBestValues[local_idx] = bestValue;
           }
           reduceSize = interval;
           workgroupBarrier();
         }

         if (local_idx == 0u) {
          ${f.setByOffset("outputIndex",`${n==="mean"?`${f.type.storage}(bestValue / f32(uniforms.reduceSize))`:`${f.type.storage}(${xl[n]})`}`)};
         }
        }`,getRunData:()=>({outputs:[{dims:i,dataType:a}],dispatchGroup:{x:u},programUniforms:[{type:12,data:l}]})}},Kt=(t,e,r,n)=>{let a=t.inputs.length===1?r:yo(t.inputs,r),i=a.axes;i.length===0&&!a.noopWithEmptyAxes&&(i=t.inputs[0].dims.map((c,y)=>y));let s=Y.normalizeAxes(i,t.inputs[0].dims.length),o=s,u=t.inputs[0],l=Tl(o,t.inputs[0].dims.length);l.length>0&&(u=t.compute(br(t.inputs[0],l),{inputs:[0],outputs:[-1]})[0],o=Sl(o.length,u.dims.length));let[p,f]=kl(u.dims,o),m=p;a.keepDims&&(m=El(p,s)),t.compute(Il(e,{hint:a.cacheKey,inputDependencies:["type"]},[u],n,t.inputs[0].dataType,m,f),{inputs:[u]})},gh=(t,e)=>{Kt(t,"ReduceMeanShared",e,"mean")},_h=(t,e)=>{Kt(t,"ReduceL1Shared",e,"l1")},yh=(t,e)=>{Kt(t,"ReduceL2Shared",e,"l2")},wh=(t,e)=>{Kt(t,"ReduceLogSumExpShared",e,"logSumExp")},bh=(t,e)=>{Kt(t,"ReduceMaxShared",e,"max")},vh=(t,e)=>{Kt(t,"ReduceMinShared",e,"min")},$h=(t,e)=>{Kt(t,"ReduceProdShared",e,"prod")},xh=(t,e)=>{Kt(t,"ReduceSumShared",e,"sum")},Sh=(t,e)=>{Kt(t,"ReduceSumSquareShared",e,"sumSquare")},kh=(t,e)=>{Kt(t,"ReduceLogSumShared",e,"logSum")}}),Yt,Al,Li,yo,Xt,Ml,Ol,zl,Pl,Rl,Bl,Dl,Nl,Fl,Ll,Qt,Eh,Ch,Th,Ih,Ah,Mh,Oh,zh,Ph,Rh,Xo=Z(()=>{$e(),Ae(),lt(),Te(),Z_(),Yt=t=>{if(!t||t.length===0||t.length>2)throw new Error("Reduce op requires 1 or 2 inputs.");if(t.length===2&&t[1].dims.length!==1)throw new Error("Invalid axes input dims.")},Al=t=>["","",`var value = ${t.getByIndices("input_indices")};`,""],Li=(t,e,r,n,a,i,s=!1,o=!1)=>{let u=[],l=r[0].dims,p=l.length,f=Y.normalizeAxes(a,p),m=!o&&f.length===0;l.forEach((w,v)=>{m||f.indexOf(v)>=0?s&&u.push(1):u.push(w)});let c=u.length,y=Y.size(u);return{name:t,shaderCache:e,getShaderSource:w=>{let v=[],k=Q("_A",r[0].dataType,p),$=ge("output",i,c),C=n(k,$,f),T=C[2];for(let A=0,B=0;A<p;A++)m||f.indexOf(A)>=0?(s&&B++,T=`for(var j${A}: u32 = 0; j${A} < ${l[A]}; j${A}++) {
                  ${C[2].includes("last_index")?`let last_index = j${A};`:""}
                  ${k.indicesSet("input_indices",A,`j${A}`)}
                  ${T}
                }`):(v.push(`${k.indicesSet("input_indices",A,$.indicesGet("output_indices",B))};`),B++);return`

        ${w.registerUniform("output_size","u32").declareVariables(k,$)}

        ${w.mainStart()}
          ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
          var input_indices: ${k.type.indices};
          let output_indices = ${$.offsetToIndices("global_idx")};

          ${v.join(`
`)}
          ${C[0]}       // init ops for reduce max/min
          ${C[1]}
          ${T}
          ${C[3]}
          ${C.length===4?$.setByOffset("global_idx","value"):C.slice(4).join(`
`)}
        }`},getRunData:()=>({outputs:[{dims:u,dataType:i}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:[{type:12,data:y},...ye(l,u)]})}},yo=(t,e)=>{let r=[];return t[1].dims[0]>0&&t[1].getBigInt64Array().forEach(n=>r.push(Number(n))),qe({axes:r,keepDims:e.keepDims,noopWithEmptyAxes:e.noopWithEmptyAxes})},Xt=(t,e,r,n)=>{let a=t.inputs,i=a.length===1?r:yo(a,r);t.compute(Li(e,{hint:i.cacheKey,inputDependencies:["rank"]},[a[0]],i.noopWithEmptyAxes&&i.axes.length===0?Al:n,i.axes,a[0].dataType,i.keepDims,i.noopWithEmptyAxes),{inputs:[0]})},Ml=(t,e)=>{Yt(t.inputs),Xt(t,"ReduceLogSum",e,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,"value = log(value);"])},Ol=(t,e)=>{Yt(t.inputs),Xt(t,"ReduceL1",e,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += abs(${r.getByIndices("input_indices")});`,""])},zl=(t,e)=>{Yt(t.inputs),Xt(t,"ReduceL2",e,(r,n)=>[`var t = ${n.type.value}(0); var value = ${n.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},Pl=(t,e)=>{Yt(t.inputs),Xt(t,"ReduceLogSumExp",e,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += exp(${r.getByIndices("input_indices")});`,"value = log(value);"])},Rl=(t,e)=>{Yt(t.inputs),Xt(t,"ReduceMax",e,(r,n,a)=>{let i=[];for(let s=0;s<r.rank;s++)(a.indexOf(s)>=0||a.length===0)&&i.push(r.indicesSet("input_indices",s,0));return[`${i.join(`
`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = max(value, ${r.getByIndices("input_indices")});`,""]})},Bl=(t,e)=>{Yt(t.inputs),Xt(t,"ReduceMean",e,(r,n,a)=>{let i=1;for(let s=0;s<r.rank;s++)(a.indexOf(s)>=0||a.length===0)&&(i*=t.inputs[0].dims[s]);return["var sum = f32(0);","",`sum += f32(${r.getByIndices("input_indices")});`,`let value = ${n.type.value}(sum / ${i});`]})},Dl=(t,e)=>{Yt(t.inputs),Xt(t,"ReduceMin",e,(r,n,a)=>{let i=[];for(let s=0;s<r.rank;s++)(a.indexOf(s)>=0||a.length===0)&&i.push(`input_indices[${s}] = 0;`);return[`${i.join(`
`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = min(value, ${r.getByIndices("input_indices")});`,""]})},Nl=(t,e)=>{Yt(t.inputs),Xt(t,"ReduceProd",e,(r,n)=>[`var value = ${n.type.storage}(1);`,"",`value *= ${r.getByIndices("input_indices")};`,""])},Fl=(t,e)=>{Yt(t.inputs),Xt(t,"ReduceSum",e,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,""])},Ll=(t,e)=>{Yt(t.inputs),Xt(t,"ReduceSumSquare",e,(r,n)=>[`var t = ${n.type.value}(0); var value = ${n.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += t * t;`,""])},Qt=(t,e,r)=>{if(e.length===0)return r;let n=1,a=1;for(let i=0;i<e.length;i++)e.indexOf(i)===-1?n*=t[i]:a*=t[i];return a<32&&n>1024},Eh=(t,e)=>{Qt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Bl(t,e):gh(t,e)},Ch=(t,e)=>{Qt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Ol(t,e):_h(t,e)},Th=(t,e)=>{Qt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?zl(t,e):yh(t,e)},Ih=(t,e)=>{Qt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Pl(t,e):wh(t,e)},Ah=(t,e)=>{Qt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Rl(t,e):bh(t,e)},Mh=(t,e)=>{Qt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Dl(t,e):vh(t,e)},Oh=(t,e)=>{Qt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Nl(t,e):$h(t,e)},zh=(t,e)=>{Qt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Fl(t,e):xh(t,e)},Ph=(t,e)=>{Qt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Ll(t,e):Sh(t,e)},Rh=(t,e)=>{Qt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Ml(t,e):kh(t,e)}}),zs,Bh,Dh,wo,ey=Z(()=>{$e(),lt(),Xo(),zs=t=>{if(!t||t.length===0||t.length>2)throw new Error("ArgMinMaxOp op requires 1 or 2 inputs.");if(t[0].dataType!==1)throw new Error("Invalid input type.")},Bh=(t,e)=>{zs(t.inputs);let r=(n,a,i)=>{let s=[];for(let o=0;o<n.rank;o++)(i.indexOf(o)>=0||i.length===0)&&s.push(`input_indices[${o}] = 0;`);return[`${s.join(`
`)}`,`var value = ${n.getByIndices("input_indices")};
var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${e.selectLastIndex>0?"<=":"<"} value) {
         value = ${n.getByIndices("input_indices")};
         best_index = i32(last_index);
       }`,"",a.setByOffset("global_idx","best_index")]};t.compute(Li("ArgMin",{hint:e.cacheKey,inputDependencies:["rank"]},[t.inputs[0]],r,[e.axis],7,e.keepDims),{inputs:[0]})},Dh=(t,e)=>{zs(t.inputs);let r=(n,a,i)=>{let s=[];for(let o=0;o<n.rank;o++)(i.indexOf(o)>=0||i.length===0)&&s.push(`input_indices[${o}] = 0;`);return[`${s.join(`
`)}`,`var value = ${n.getByIndices("input_indices")};
var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${e.selectLastIndex>0?">=":">"} value) {
         value = ${n.getByIndices("input_indices")};
         best_index = i32(last_index);
       }`,"",a.setByOffset("global_idx","best_index")]};t.compute(Li("argMax",{hint:e.cacheKey,inputDependencies:["rank"]},[t.inputs[0]],r,[e.axis],7,e.keepDims),{inputs:[0]})},wo=t=>qe(t)}),Ul,Wl,Vl,Ui,Nh,Fh,Lh=Z(()=>{$e(),Ae(),lt(),Te(),Ul=(t,e)=>{if(!t||t.length<1)throw new Error("too few inputs");let r=0,n=t[r],a=n.dataType,i=n.dims.length;t.forEach((s,o)=>{if(o!==r){if(s.dataType!==a)throw new Error("input tensors should be one type");if(s.dims.length!==i)throw new Error("input tensors should have the same shape");s.dims.forEach((u,l)=>{if(l!==e&&u!==n.dims[l])throw new Error("non concat dimensions must match")})}})},Wl=(t,e)=>`
  fn calculateInputIndex(index: u32) -> u32 {
    let sizeInConcatAxis = array<u32, ${t}u>(${e});
    for (var i: u32 = 0u; i < ${t}; i += 1u ) {
      if (index < sizeInConcatAxis[i]) {
        return i;
      }
    }
    return ${t}u;
  }`,Vl=(t,e)=>{let r=t.length,n=[];for(let a=0;a<r;++a){let i=e.setByOffset("global_idx",t[a].getByIndices("indices"));r===1?n.push(i):a===0?n.push(`if (inputIndex == ${a}u) { ${i} }`):a===r-1?n.push(`else { ${i} }`):n.push(`else if (inputIndex == ${a}) { ${i} }`)}return n.join(`
`)},Ui=(t,e,r,n)=>{let a=Y.size(r),i=new Array(t.length),s=new Array(t.length),o=0,u=[],l=[],p=[{type:12,data:a}];for(let w=0;w<t.length;++w)o+=t[w].dims[e],i[w]=o,l.push(t[w].dims.length),s[w]=Q(`input${w}`,n,l[w]),u.push("rank"),p.push({type:12,data:i[w]});for(let w=0;w<t.length;++w)p.push(...ye(t[w].dims));p.push(...ye(r));let f=ge("output",n,r.length),m=f.indicesGet("indices",e),c=Array.from(Array(i.length).keys()).map(w=>`uniforms.sizeInConcatAxis${w}`).join(","),y=w=>`

  ${(()=>{w.registerUniform("outputSize","u32");for(let v=0;v<t.length;v++)w.registerUniform(`sizeInConcatAxis${v}`,"u32");return w.declareVariables(...s,f)})()}

  ${Wl(i.length,c)}

  ${w.mainStart()}
    ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}

    var indices = ${f.offsetToIndices("global_idx")};

    let inputIndex = calculateInputIndex(${m});
    if (inputIndex != 0u) {
      let sizeInConcatAxis = array<u32, ${i.length}u>(${c});
      ${m} -= sizeInConcatAxis[inputIndex - 1u];
    }

    ${Vl(s,f)}
  }`;return{name:"Concat",shaderCache:{hint:`${e}`,inputDependencies:u},getRunData:()=>({outputs:[{dims:r,dataType:n}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:p}),getShaderSource:y}},Nh=(t,e)=>{let r=t.inputs,n=r[0].dims,a=Y.normalizeAxis(e.axis,n.length);Ul(r,a);let i=n.slice();i[a]=r.reduce((o,u)=>o+(u.dims.length>a?u.dims[a]:0),0);let s=r.filter(o=>Y.size(o.dims)>0);t.compute(Ui(s,a,i,r[0].dataType),{inputs:s})},Fh=t=>qe({axis:t.axis})}),Gl,Hl,jl,ql,Wi,Kl,Uh,Wh=Z(()=>{$e(),jo(),Te(),Lh(),Gl=(t,e)=>{let r=t[0],n=t[1],a=t[2],i=t[3],s=t[4],o=t[5];if(s&&o)throw new Error("Attention cannot have both past and relative_position_bias");if(r.dims.length!==3)throw new Error('Input "input" must have 3 dimensions');let u=r.dims[0],l=r.dims[1],p=r.dims[2];if(a.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimensions');if(n.dims.length!==2)throw new Error('Input "weights" is expected to have 2 dimensions');if(n.dims[0]!==p)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(a.dims[0]!==n.dims[1])throw new Error('Input "bias" dimension 0 should have same length as dimension 1 of input "weights"');let f=a.dims[0]/3,m=f,c=m;if(e.qkvHiddenSizes.length>0){if(e.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let C of e.qkvHiddenSizes)if(C%e.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");f=e.qkvHiddenSizes[0],m=e.qkvHiddenSizes[1],c=e.qkvHiddenSizes[2]}let y=l;if(f!==m)throw new Error("qkv_hidden_sizes first element should be same as the second");if(a.dims[0]!==f+m+c)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let w=0;if(s){if(m!==c)throw new Error('Input "past" expect k_hidden_size == v_hidden_size');if(s.dims.length!==5)throw new Error('Input "past" must have 5 dimensions');if(s.dims[0]!==2)throw new Error('Input "past" first dimension must be 2');if(s.dims[1]!==u)throw new Error('Input "past" second dimension must be batch_size');if(s.dims[2]!==e.numHeads)throw new Error('Input "past" third dimension must be num_heads');if(s.dims[4]!==m/e.numHeads)throw new Error('Input "past" fifth dimension must be k_hidden_size / num_heads');e.pastPresentShareBuffer||(w=s.dims[3])}let v=y+w,k=-1,$=0;if(i)throw new Error("Mask not supported");if(s)throw new Error("past is not supported");return{batchSize:u,sequenceLength:l,pastSequenceLength:w,kvSequenceLength:y,totalSequenceLength:v,maxSequenceLength:k,inputHiddenSize:p,hiddenSize:f,vHiddenSize:c,headSize:Math.floor(f/e.numHeads),vHeadSize:Math.floor(c/e.numHeads),numHeads:e.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:e.maskFilterValue,maskType:$,scale:e.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},Hl=(t,e,r,n)=>{let a=it(n),i=64,s=n/a;s<i?i=1:s/8<64&&(i=Math.ceil(s/8));let o=Math.ceil(n/a/i),u=[{type:e.dataType,data:1/n},{type:12,data:s},{type:12,data:o}],l=pt(e.dataType,a),p=Tt(1,a),f=m=>{let c=ge("x",e.dataType,e.dims,a),y=[{name:"d_inv",type:Tt(e.dataType)},{name:"d_comp",type:"u32"},{name:"elements_per_thread",type:"u32"}];return`
  var<workgroup> thread_max: array<f32, ${i}>;
  var<workgroup> thread_sum: array<f32, ${i}>;
  ${m.registerUniforms(y).declareVariables(c)}
  ${m.mainStart([i,1,1])}
    let local_offset = local_idx * uniforms.elements_per_thread;
    let offset = workgroup_id.x * uniforms.d_comp + local_offset;

    var thread_max_vector = ${p}(-3.402823e+38f);
    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {
      thread_max_vector = max(${p}(x[offset + i]), thread_max_vector);
    }
    thread_max[local_idx] = ${(()=>{switch(a){case 1:return"thread_max_vector";case 2:return"max(thread_max_vector.x, thread_max_vector.y)";case 4:return"max(max(thread_max_vector.x, thread_max_vector.y), max(thread_max_vector.z, thread_max_vector.w))";default:throw new Error(`Unsupported components: ${a}`)}})()};
    workgroupBarrier();

    var max_value =  f32(-3.402823e+38f);
    for (var i = 0u; i < ${i}; i++) {
      max_value = max(thread_max[i], max_value);
    }

    var sum_vector = ${p}(0);
    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {
      sum_vector += exp(${p}(x[offset + i]) - max_value);
    }
    thread_sum[local_idx] = ${(()=>{switch(a){case 1:return"sum_vector";case 2:return"sum_vector.x + sum_vector.y";case 4:return"sum_vector.x + sum_vector.y + sum_vector.z + sum_vector.w";default:throw new Error(`Unsupported components: ${a}`)}})()};
    workgroupBarrier();

    var sum: f32 = 0;
    for (var i = 0u; i < ${i}; i++) {
      sum += thread_sum[i];
    }

    if (sum == 0) {
      for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {
        x[offset + i] = ${c.type.value}(uniforms.d_inv);
      }
    } else {
      for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {
        var f32input = ${p}(x[offset + i]);
        x[offset + i] = ${c.type.value}(exp(f32input - max_value) / sum);
      }
    }
  }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${i};${l};${a}`},getShaderSource:f,getRunData:()=>({outputs:[],dispatchGroup:{x:r},programUniforms:u})}},jl=(t,e,r,n,a,i,s)=>{let o=s+a.kvSequenceLength,u=[a.batchSize,a.numHeads,a.sequenceLength,o],l=i.scale===0?1/Math.sqrt(a.headSize):i.scale,p=it(a.headSize),f=a.headSize/p,m=12,c={x:Math.ceil(o/m),y:Math.ceil(a.sequenceLength/m),z:a.batchSize*a.numHeads},y=[{type:12,data:a.sequenceLength},{type:12,data:f},{type:12,data:o},{type:12,data:a.numHeads},{type:1,data:l}],w=n?["type","type","type"]:["type","type"],v=k=>{let $=Q("q",e.dataType,e.dims,p),C=Q("key",r.dataType,r.dims,p),T=[$,C];n&&T.push(Q("relative_position_bias",n.dataType,n.dims));let A=ge("output",e.dataType,u),B=Tt(1,p),R=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"alpha",type:"f32"}];return`
  const TILE_SIZE = ${m}u;

  var<workgroup> tileQ: array<${$.type.storage}, ${m*m}>;
  var<workgroup> tileK: array<${$.type.storage}, ${m*m}>;
  ${k.registerUniforms(R).declareVariables(...T,A)}
  ${k.mainStart([m,m,1])}
    // x holds the N and y holds the M
    let headIdx = workgroup_id.z;
    let m = workgroup_id.y * TILE_SIZE;
    let n = workgroup_id.x * TILE_SIZE;
    let qOffset = uniforms.M * uniforms.K * headIdx + m * uniforms.K;
    let kOffset = uniforms.N * uniforms.K * headIdx + n * uniforms.K;

    var value = ${B}(0);
    for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {
      if (global_id.y < uniforms.M && w + local_id.x < uniforms.K) {
        tileQ[TILE_SIZE * local_id.y + local_id.x] = q[qOffset + local_id.y * uniforms.K + w + local_id.x];
      }
      if (n + local_id.y < uniforms.N && w + local_id.x < uniforms.K) {
        tileK[TILE_SIZE * local_id.y + local_id.x] = key[kOffset + local_id.y * uniforms.K + w + local_id.x];
      }
      workgroupBarrier();

      for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) {
        value += ${B}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]);
      }

      workgroupBarrier();
    }

    let headOffset = headIdx * uniforms.M * uniforms.N;
    if (global_id.y < uniforms.M && global_id.x < uniforms.N) {
      let outputIdx = headOffset + global_id.y * uniforms.N + global_id.x;
      var sum: f32 = ${(()=>{switch(p){case 1:return"value";case 2:return"value.x + value.y";case 4:return"value.x + value.y + value.z + value.w";default:throw new Error(`Unsupported components: ${p}`)}})()};
        output[outputIdx] = ${A.type.value} (sum * uniforms.alpha) + ${n?"relative_position_bias[outputIdx]":"0.0"};
    }
  }`};return{name:"AttentionProbs",shaderCache:{hint:`${p}`,inputDependencies:w},getRunData:()=>({outputs:[{dims:u,dataType:e.dataType,gpuDataType:0}],dispatchGroup:c,programUniforms:y}),getShaderSource:v}},ql=(t,e,r,n,a)=>{let i=a+n.kvSequenceLength,s=[n.batchSize,n.sequenceLength,n.vHiddenSize],o=12,u={x:Math.ceil(n.vHeadSize/o),y:Math.ceil(n.sequenceLength/o),z:n.batchSize*n.numHeads},l=[{type:12,data:n.sequenceLength},{type:12,data:i},{type:12,data:n.vHeadSize},{type:12,data:n.numHeads},{type:12,data:n.vHiddenSize}];return{name:"AttentionScore",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:s,dataType:e.dataType,gpuDataType:0}],dispatchGroup:u,programUniforms:l}),getShaderSource:p=>{let f=Q("probs",e.dataType,e.dims),m=Q("v",r.dataType,r.dims),c=ge("output",e.dataType,s),y=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"v_hidden_size",type:"u32"}];return`
  const TILE_SIZE = ${o}u;
  var<workgroup> tileQ: array<${f.type.value}, ${o*o}>;
  var<workgroup> tileK: array<${f.type.value}, ${o*o}>;
  ${p.registerUniforms(y).declareVariables(f,m,c)}
  ${p.mainStart([o,o,1])}
   let headIdx = workgroup_id.z;
   let m = global_id.y;
   let n = global_id.x;

   let offsetA = headIdx * (uniforms.M * uniforms.K) + m * uniforms.K;
   let offsetB = headIdx * (uniforms.N * uniforms.K) + n;

   var value = ${f.type.storage}(0);
   for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {
     if (m < uniforms.M && w + local_id.x < uniforms.K) {
       tileQ[TILE_SIZE * local_id.y + local_id.x] = probs[offsetA + w + local_id.x];
     }
     if (n < uniforms.N && w + local_id.y < uniforms.K) {
       tileK[TILE_SIZE * local_id.y + local_id.x] = v[offsetB + (w + local_id.y) * uniforms.N];
     }
     workgroupBarrier();
     for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) {
       value += tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * k + local_id.x];
     }
     workgroupBarrier();
   }

   // we need to transpose output from BNSH_v to BSND_v
   let batchIdx = workgroup_id.z / uniforms.num_heads;
   let currentBatchHeadNumber = workgroup_id.z % uniforms.num_heads;
   if (m < uniforms.M && n < uniforms.N) {
     let outputIdx = batchIdx * uniforms.M * uniforms.v_hidden_size + m * uniforms.v_hidden_size
       + currentBatchHeadNumber * uniforms.N + n;
     output[outputIdx] = value;
   }
  }`}}},Wi=(t,e,r,n,a,i,s,o,u,l,p)=>{let f=t.outputCount>1,m=t.outputCount>2,c=f&&m?l.pastSequenceLength:0,y=c+l.kvSequenceLength,w=[l.batchSize,l.numHeads,y,l.headSize],v=s?[s,r]:[r],k=f?t.compute(Ui(v,2,w,r.dataType),{inputs:v,outputs:[1]})[0]:r,$=[l.batchSize,l.numHeads,y,l.headSize],C=o?[o,n]:[n],T=m?t.compute(Ui(C,2,$,n.dataType),{inputs:C,outputs:[2]})[0]:n,A=[e,k];u&&A.push(u);let B=t.compute(jl(t,e,k,u,l,p,c),{inputs:A,outputs:[-1]})[0];t.compute(Hl(t,B,l.batchSize*l.numHeads*l.sequenceLength,y),{inputs:[B],outputs:[]});let R=[B,T];t.compute(ql(t,B,T,l,c),{inputs:R,outputs:[0]})},Kl=(t,e)=>{let r=[e.batchSize,e.numHeads,e.sequenceLength,e.headSize],n=e.sequenceLength,a=e.inputHiddenSize,i=e.headSize,s=12,o={x:Math.ceil(e.headSize/s),y:Math.ceil(e.sequenceLength/s),z:e.batchSize*e.numHeads},u=[t.inputs[0],t.inputs[1],t.inputs[2]],l=[{type:12,data:n},{type:12,data:a},{type:12,data:i},{type:12,data:e.numHeads},{type:12,data:e.headSize},{type:12,data:e.hiddenSize},{type:12,data:e.hiddenSize+e.hiddenSize+e.vHiddenSize}],p=f=>{let m=ge("output_q",u[0].dataType,r),c=ge("output_k",u[0].dataType,r),y=ge("output_v",u[0].dataType,r),w=Q("input",u[0].dataType,u[0].dims),v=Q("weight",u[1].dataType,u[1].dims),k=Q("bias",u[2].dataType,u[2].dims),$=w.type.storage,C=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"hidden_size",type:"u32"},{name:"ldb",type:"u32"}];return`
  const TILE_SIZE = ${s}u;
  var<workgroup> tileInput: array<${$}, ${s*s}>;
  var<workgroup> tileWeightQ: array<${$}, ${s*s}>;
  var<workgroup> tileWeightK: array<${$}, ${s*s}>;
  var<workgroup> tileWeightV: array<${$}, ${s*s}>;
  ${f.registerUniforms(C).declareVariables(w,v,k,m,c,y)}
  ${f.mainStart([s,s,1])}
    let batchIndex = workgroup_id.z / uniforms.num_heads;
    let headNumber = workgroup_id.z % uniforms.num_heads;
    let m = global_id.y;
    let n = global_id.x;

    let inputOffset = batchIndex * (uniforms.M * uniforms.K) + m * uniforms.K;
    let biasOffsetQ = headNumber * uniforms.head_size;
    let biasOffsetK = uniforms.hidden_size + biasOffsetQ;
    let biasOffsetV = uniforms.hidden_size + biasOffsetK;

    var valueQ = ${$}(0);
    var valueK = ${$}(0);
    var valueV = ${$}(0);
    for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {
      if (m < uniforms.M && w + local_id.x < uniforms.K) {
        tileInput[TILE_SIZE * local_id.y + local_id.x] = input[inputOffset + w + local_id.x];
      }
      if (n < uniforms.N && w + local_id.y < uniforms.K) {
        let offset = n + (w + local_id.y) * uniforms.ldb;
        tileWeightQ[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetQ + offset];
        tileWeightK[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetK + offset];
        tileWeightV[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetV + offset];
      }
      workgroupBarrier();
      for (var k: u32 = 0u; k<TILE_SIZE && w+k < uniforms.K; k++) {
        let inputTileOffset = TILE_SIZE * local_id.y + k;
        let weightTileOffset = TILE_SIZE * k + local_id.x;
        valueQ += tileInput[inputTileOffset] * tileWeightQ[weightTileOffset];
        valueK += tileInput[inputTileOffset] * tileWeightK[weightTileOffset];
        valueV += tileInput[inputTileOffset] * tileWeightV[weightTileOffset];
      }

      workgroupBarrier();
    }

    let headOffset = (m * uniforms.N + n) % uniforms.head_size;
    valueQ += bias[headOffset + biasOffsetQ];
    valueK += bias[headOffset + biasOffsetK];
    valueV += bias[headOffset + biasOffsetV];

    let offset = workgroup_id.z * uniforms.M * uniforms.N;
    if (m < uniforms.M && n < uniforms.N) {
      let outputIdx = offset + m * uniforms.N + n;
      output_q[outputIdx] = valueQ;
      output_k[outputIdx] = valueK;
      output_v[outputIdx] = valueV;
    }
  }`};return t.compute({name:"AttentionPrepare",shaderCache:{inputDependencies:["type","type","type"]},getRunData:()=>({outputs:[{dims:r,dataType:t.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:t.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:t.inputs[0].dataType,gpuDataType:0}],dispatchGroup:o,programUniforms:l}),getShaderSource:p},{inputs:u,outputs:[-1,-1,-1]})},Uh=(t,e)=>{let r=Gl(t.inputs,e),[n,a,i]=Kl(t,r);return Wi(t,n,a,i,t.inputs[4],void 0,void 0,void 0,t.inputs[5],r,e)}}),Yl,Xl,Ql,Vh,ty=Z(()=>{ar(),$e(),Ae(),lt(),Te(),Yl=(t,e)=>{if(!t||t.length!==5)throw new Error("BatchNormalization requires 5 inputs");let r=(n,a,i)=>{let s=a.length;if(s!==n.length)throw new Error(`${i}: num dimensions != ${s}`);a.forEach((o,u)=>{if(o!==n[u])throw new Error(`${i}: dim[${u}] do not match`)})};if(t[0].dims.length>1){let n=e.format==="NHWC"?e.spatial?t[0].dims.slice(-1):t[0].dims.slice(-1).concat(t[0].dims.slice(1,t[0].dims.length-1)):t[0].dims.slice(1,e.spatial?2:void 0);r(t[1].dims,n,"Invalid input scale"),r(t[2].dims,n,"Invalid input B"),r(t[3].dims,n,"Invalid input mean"),r(t[4].dims,n,"Invalid input var")}else r(t[1].dims,[1],"Invalid input scale"),r(t[2].dims,[1],"Invalid input B"),r(t[3].dims,[1],"Invalid input mean"),r(t[4].dims,[1],"Invalid input var")},Xl=(t,e)=>{let{epsilon:r,spatial:n,format:a}=e,i=t[0].dims,s=n?it(i[i.length-1]):1,o=a==="NHWC"&&i.length>1?s:1,u=Y.size(i)/s,l=n,p=l?i.length:i,f=Q("x",t[0].dataType,t[0].dims,s),m=Q("scale",t[1].dataType,t[1].dims,o),c=Q("bias",t[2].dataType,t[2].dims,o),y=Q("inputMean",t[3].dataType,t[3].dims,o),w=Q("inputVar",t[4].dataType,t[4].dims,o),v=ge("y",t[0].dataType,p,s),k=()=>{let C="";if(n)C=`let cOffset = ${i.length===1?"0u":a==="NHWC"?`outputIndices[${i.length-1}] / ${s}`:"outputIndices[1]"};`;else if(a==="NCHW")C=`
            ${v.indicesSet("outputIndices","0","0")}
            let cOffset = ${v.indicesToOffset("outputIndices")};`;else{C=`var cIndices = ${m.type.indices}(0);
                       cIndices[0] = outputIndices[${i.length-1}];`;for(let T=1;T<m.rank;T++)C+=`cIndices[${T}] = outputIndices[${T}];`;C+=`let cOffset = ${m.indicesToOffset("cIndices")};`}return C},$=C=>`
  const epsilon = ${r};
  ${C.registerUniform("outputSize","u32").declareVariables(f,m,c,y,w,v)}
  ${C.mainStart()}
  ${C.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
    var outputIndices = ${v.offsetToIndices(`global_idx * ${s}`)};
    ${k()}
    let scale = ${m.getByOffset("cOffset")};
    let bias = ${c.getByOffset("cOffset")};
    let inputMean = ${y.getByOffset("cOffset")};
    let inputVar = ${w.getByOffset("cOffset")};
    let x = ${f.getByOffset("global_idx")};
    let value = (x - inputMean) * inverseSqrt(inputVar + epsilon) * scale + bias;
    ${v.setByOffset("global_idx","value")}
  }`;return{name:"BatchNormalization",shaderCache:{hint:`${e.epsilon}_${e.format}_${n}_${s}`,inputDependencies:l?["rank","type","type","type","type"]:void 0},getShaderSource:$,getRunData:()=>({outputs:[{dims:t[0].dims,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:l?[{type:12,data:u},...ye(i)]:[{type:12,data:u}]})}},Ql=t=>qe(t),Vh=(t,e)=>{let{inputs:r,outputCount:n}=t,a=Ql({...e,outputCount:n});if(Ue.webgpu.validateInputContent&&Yl(r,a),e.trainingMode)throw new Error("BatchNormalization trainingMode is not supported yet.");t.compute(Xl(r,a))}}),Jl,Zl,Gh,ry=Z(()=>{Ae(),Te(),Jl=t=>{if(t[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![320,640,1280].includes(t[0].dims[2]))throw new Error("number of channels should be 320, 640 or 1280");if(t[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(t[0].dims[2]!==t[1].dims[0])throw new Error("last dimension of input and bias are not the same")},Zl=t=>{let e=t[0].dims,r=t[0].dims[2],n=Y.size(e)/4,a=t[0].dataType,i=Q("input",a,e,4),s=Q("bias",a,[r],4),o=Q("residual",a,e,4),u=ge("output",a,e,4);return{name:"BiasAdd",getRunData:()=>({outputs:[{dims:e,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(n/64)}}),getShaderSource:l=>`
  const channels = ${r}u / 4;
  ${l.declareVariables(i,s,o,u)}

  ${l.mainStart()}
    ${l.guardAgainstOutOfBoundsWorkgroupSizes(n)}
    let value = ${i.getByOffset("global_idx")}
      + ${s.getByOffset("global_idx % channels")} + ${o.getByOffset("global_idx")};
    ${u.setByOffset("global_idx","value")}
  }`}},Gh=t=>{Jl(t.inputs),t.compute(Zl(t.inputs))}}),ed,Re,Hh,jh,qh,Kh,Yh,Xh,Qh,Jh,Zh,td,ef,tf,rf,nf,Ci,af,Ti,sf,of,uf,lf,df,cf,pf,hf,ff,mf,gf,_f,yf,wf,bf,vf,Ps,$f,bo,vo,xf,Sf,kf,Qo=Z(()=>{$e(),Ae(),lt(),Te(),ed=(t,e,r,n,a,i)=>{let s=Math.ceil(e/4),o="";typeof a=="string"?o=`${a}(a)`:o=a("a");let u=Q("inputData",r,[s],4),l=ge("outputData",n,[s],4);return`
      ${t.registerUniform("vec_size","u32").declareVariables(u,l)}

  ${i??""}

  ${t.mainStart()}
    ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}

    let a = ${u.getByOffset("global_idx")};
    ${l.setByOffset("global_idx",o)}
  }`},Re=(t,e,r,n,a,i=t.dataType)=>({name:e,shaderCache:{hint:a,inputDependencies:["type"]},getShaderSource:s=>ed(s,Y.size(t.dims),t.dataType,i,r,n),getRunData:s=>({outputs:[{dims:t.dims,dataType:i}],dispatchGroup:{x:Math.ceil(Y.size(s[0].dims)/64/4)},programUniforms:[{type:12,data:Math.ceil(Y.size(t.dims)/4)}]})}),Hh=t=>{t.compute(Re(t.inputs[0],"Abs","abs"))},jh=t=>{t.compute(Re(t.inputs[0],"Acos","acos"))},qh=t=>{t.compute(Re(t.inputs[0],"Acosh","acosh"))},Kh=t=>{t.compute(Re(t.inputs[0],"Asin","asin"))},Yh=t=>{t.compute(Re(t.inputs[0],"Asinh","asinh"))},Xh=t=>{t.compute(Re(t.inputs[0],"Atan","atan"))},Qh=t=>{t.compute(Re(t.inputs[0],"Atanh","atanh"))},Jh=t=>qe(t),Zh=(t,e)=>{let r;switch(e.to){case 10:r="vec4<f16>";break;case 1:r="vec4<f32>";break;case 12:r="vec4<u32>";break;case 6:r="vec4<i32>";break;case 9:r="vec4<bool>";break;default:throw new RangeError(`not supported type (specified in attribute 'to' from 'Cast' operator): ${e.to}`)}t.compute(Re(t.inputs[0],"Cast",r,void 0,e.cacheKey,e.to))},td=t=>{let e=t.length>=2&&t[1].data!==0?t[1].getFloat32Array()[0]:qo,r=t.length>=3&&t[2].data!==0?t[2].getFloat32Array()[0]:Ko;return qe({min:e,max:r})},ef=(t,e)=>{let r=t.inputs.length===1?e:td(t.inputs),n=Tt(t.inputs[0].dataType);t.compute(Re(t.inputs[0],"Clip",a=>`clamp(${a}, clip_min_, clip_max_)`,`
    const clip_min_: vec4<${n}> = vec4(${n}(${r.min}));
    const clip_max_: vec4<${n}> = vec4(${n}(${r.max}));
`,r.cacheKey),{inputs:[0]})},tf=t=>{t.compute(Re(t.inputs[0],"Ceil","ceil"))},rf=t=>{t.compute(Re(t.inputs[0],"Cos","cos"))},nf=t=>{t.compute(Re(t.inputs[0],"Cosh","cosh"))},Ci=t=>qe(t),af=(t,e)=>{let r=Tt(t.inputs[0].dataType);t.compute(Re(t.inputs[0],"Elu",n=>`elu_vf32(${n})`,`
  const elu_alpha_ = ${r}(${e.alpha});

  fn elu_f32(a: ${r}) -> ${r} {
  return select((exp(a) - 1.0) * elu_alpha_, a, a >= 0.0);
  }

  fn elu_vf32(v: vec4<${r}>) -> vec4<${r}> {
  return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w));
  }`,e.cacheKey))},Ti=(t="f32")=>`
const r0: ${t} = 0.3275911;
const r1: ${t} = 0.254829592;
const r2: ${t} = -0.284496736;
const r3: ${t} = 1.421413741;
const r4: ${t} = -1.453152027;
const r5: ${t} = 1.061405429;

fn erf_vf32(v: vec4<${t}>) -> vec4<${t}> {
  let absv = abs(v);
  let x = 1.0 / (1.0 + r0 * absv);
  return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv));
}`,sf=t=>{let e=Tt(t.inputs[0].dataType);t.compute(Re(t.inputs[0],"Erf",r=>`erf_vf32(${r})`,Ti(e)))},of=t=>{t.compute(Re(t.inputs[0],"Exp","exp"))},uf=t=>{t.compute(Re(t.inputs[0],"Floor","floor"))},lf=t=>{let e=Tt(t.inputs[0].dataType);t.compute(Re(t.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,Ti(e)))},df=(t,e)=>{let r=Tt(t.inputs[0].dataType);t.compute(Re(t.inputs[0],"LeakyRelu",n=>`select(leaky_relu_alpha_ * ${n}, ${n}, ${n} >= vec4<${r}>(0.0))`,`const leaky_relu_alpha_ = ${r}(${e.alpha});`,e.cacheKey))},cf=t=>{t.compute(Re(t.inputs[0],"Not",e=>`!${e}`))},pf=t=>{t.compute(Re(t.inputs[0],"Neg",e=>`-${e}`))},hf=t=>{t.compute(Re(t.inputs[0],"Reciprocal",e=>`1.0/${e}`))},ff=t=>{let e=Tt(t.inputs[0].dataType);t.compute(Re(t.inputs[0],"Relu",r=>`select(vec4<${e}>(0.0), ${r}, ${r} > vec4<${e}>(0.0))`))},mf=t=>{t.compute(Re(t.inputs[0],"Sigmoid",e=>`(1.0 / (1.0 + exp(-${e})))`))},gf=t=>qe(t),_f=(t,e)=>{let r=Tt(t.inputs[0].dataType);t.compute(Re(t.inputs[0],"HardSigmoid",n=>`max(vec4<${r}>(0.0), min(vec4<${r}>(1.0), ${e.alpha} * ${n} + vec4<${r}>(${e.beta})))`,void 0,e.cacheKey))},yf=t=>{t.compute(Re(t.inputs[0],"Sin","sin"))},wf=t=>{t.compute(Re(t.inputs[0],"Sinh","sinh"))},bf=t=>{t.compute(Re(t.inputs[0],"Sqrt","sqrt"))},vf=t=>{t.compute(Re(t.inputs[0],"Tan","tan"))},Ps=t=>`sign(${t}) * (1 - exp(-2 * abs(${t}))) / (1 + exp(-2 * abs(${t})))`,$f=t=>{t.compute(Re(t.inputs[0],"Tanh",Ps))},bo=(t="f32")=>`
const fast_gelu_a: ${t} = 0.5;
const fast_gelu_b: ${t} = 0.7978845608028654;
const fast_gelu_c: ${t} = 0.035677408136300125;

fn tanh_v(v: vec4<${t}>) -> vec4<${t}> {
  return ${Ps("v")};
}
`,vo=t=>`(fast_gelu_a + fast_gelu_a * tanh_v(${t} * (fast_gelu_c * ${t} * ${t} + fast_gelu_b))) * ${t}`,xf=t=>{let e=Tt(t.inputs[0].dataType);t.compute(Re(t.inputs[0],"FastGelu",vo,bo(e),void 0,t.inputs[0].dataType))},Sf=(t,e)=>{let r=Tt(t.inputs[0].dataType);return t.compute(Re(t.inputs[0],"ThresholdedRelu",n=>`select(vec4<${r}>(0.0), ${n}, ${n} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${e.alpha});`,e.cacheKey)),0},kf=t=>{t.compute(Re(t.inputs[0],"Log","log"))}}),rd,nd,Ef,ny=Z(()=>{Ae(),Te(),Qo(),rd=t=>{if(t[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![2560,5120,10240].includes(t[0].dims[2]))throw new Error("hidden state should be 2560, 5120 or 10240");if(t[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(t[0].dims[2]!==t[1].dims[0])throw new Error("last dimension of input and bias are not the same")},nd=t=>{let e=t[0].dims.slice();e[2]=e[2]/2;let r=Q("input",t[0].dataType,t[0].dims,4),n=Q("bias",t[0].dataType,[t[0].dims[2]],4),a=ge("output",t[0].dataType,e,4),i=Y.size(e)/4,s=pt(t[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:e,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)}}),getShaderSource:o=>`
  const M_SQRT2 = sqrt(2.0);
  const halfChannels = ${t[0].dims[2]/4/2}u;

  ${o.declareVariables(r,n,a)}

  ${Ti(s)}

  ${o.mainStart()}
    ${o.guardAgainstOutOfBoundsWorkgroupSizes(i)}
    let biasIdx = global_idx % halfChannels;
    let batchIndex = global_idx / halfChannels;
    let inputOffset = biasIdx + batchIndex * halfChannels * 2;
    let valueLeft = input[inputOffset] + bias[biasIdx];
    let valueRight = input[inputOffset + halfChannels] + bias[biasIdx + halfChannels];
    let geluRight = valueRight * 0.5 * (erf_vf32(valueRight / M_SQRT2) + 1);

    ${a.setByOffset("global_idx","valueLeft * geluRight")}
  }`}},Ef=t=>{rd(t.inputs),t.compute(nd(t.inputs))}}),ad,id,Jt,Cf,Tf,If,Af,Mf,Of,zf,Pf,Rf,Bf,ay=Z(()=>{$e(),Ae(),Te(),ad=(t,e,r,n,a,i,s,o,u,l,p,f)=>{let m,c;typeof o=="string"?m=c=($,C)=>`${o}((${$}),(${C}))`:typeof o=="function"?m=c=o:(m=o.scalar,c=o.vector);let y=ge("outputData",p,n.length,4),w=Q("aData",u,e.length,4),v=Q("bData",l,r.length,4),k;if(a)if(i){let $=Y.size(e)===1,C=Y.size(r)===1,T=e.length>0&&e[e.length-1]%4===0,A=r.length>0&&r[r.length-1]%4===0;$||C?k=y.setByOffset("global_idx",c($?`${w.type.value}(${w.getByOffset("0")}.x)`:w.getByOffset("global_idx"),C?`${v.type.value}(${v.getByOffset("0")}.x)`:v.getByOffset("global_idx"))):k=`
            let outputIndices = ${y.offsetToIndices("global_idx * 4u")};
            let offsetA = ${w.broadcastedIndicesToOffset("outputIndices",y)};
            let offsetB = ${v.broadcastedIndicesToOffset("outputIndices",y)};
            ${y.setByOffset("global_idx",c(s||T?w.getByOffset("offsetA / 4u"):`${w.type.value}(${w.getByOffset("offsetA / 4u")}[offsetA % 4u])`,s||A?v.getByOffset("offsetB / 4u"):`${v.type.value}(${v.getByOffset("offsetB / 4u")}[offsetB % 4u])`))}
          `}else k=y.setByOffset("global_idx",c(w.getByOffset("global_idx"),v.getByOffset("global_idx")));else{if(!i)throw new Error("no necessary to use scalar implementation for element-wise binary op implementation.");let $=(C,T,A="")=>{let B=`aData[indexA${T}][componentA${T}]`,R=`bData[indexB${T}][componentB${T}]`;return`
            let outputIndices${T} = ${y.offsetToIndices(`global_idx * 4u + ${T}u`)};
            let offsetA${T} = ${w.broadcastedIndicesToOffset(`outputIndices${T}`,y)};
            let offsetB${T} = ${v.broadcastedIndicesToOffset(`outputIndices${T}`,y)};
            let indexA${T} = offsetA${T} / 4u;
            let indexB${T} = offsetB${T} / 4u;
            let componentA${T} = offsetA${T} % 4u;
            let componentB${T} = offsetB${T} % 4u;
            ${C}[${T}] = ${A}(${m(B,R)});
          `};p===9?k=`
            var data = vec4<u32>(0);
            ${$("data",0,"u32")}
            ${$("data",1,"u32")}
            ${$("data",2,"u32")}
            ${$("data",3,"u32")}
            outputData[global_idx] = dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(data));`:k=`
            ${$("outputData[global_idx]",0)}
            ${$("outputData[global_idx]",1)}
            ${$("outputData[global_idx]",2)}
            ${$("outputData[global_idx]",3)}
          `}return`
        ${t.registerUniform("vec_size","u32").declareVariables(w,v,y)}

        ${f??""}

        ${t.mainStart()}
        ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
        ${k}
      }`},id=(t,e,r,n,a,i,s=r.dataType)=>{let o=!Y.areEqual(r.dims,n.dims),u=r.dims,l=Y.size(r.dims),p=!1,f=!1,m=[o];if(o){let c=kn.calcShape(r.dims,n.dims,!1);if(!c)throw new Error("Can't perform binary op on the given tensors");u=c,l=Y.size(u);let y=Y.size(r.dims)===1,w=Y.size(n.dims)===1,v=r.dims.length>0&&r.dims[r.dims.length-1]%4===0,k=n.dims.length>0&&n.dims[n.dims.length-1]%4===0;m.push(y),m.push(w),m.push(v),m.push(k);let $=1;for(let C=1;C<u.length;C++){let T=r.dims[r.dims.length-C]??1,A=n.dims[n.dims.length-C]??1;if(T===A)$*=T;else break}$%4===0?(f=!0,p=!0):(y||w||v||k)&&(p=!0)}else p=!0;return m.push(p),{name:t,shaderCache:{hint:e+m.map(c=>c.toString()).join("_"),inputDependencies:["rank","rank"]},getShaderSource:c=>ad(c,r.dims,n.dims,u,p,o,f,a,r.dataType,n.dataType,s,i),getRunData:()=>({outputs:[{dims:u,dataType:s}],dispatchGroup:{x:Math.ceil(l/64/4)},programUniforms:[{type:12,data:Math.ceil(Y.size(u)/4)},...ye(r.dims,n.dims,u)]})}},Jt=(t,e,r,n,a,i)=>{t.compute(id(e,a??"",t.inputs[0],t.inputs[1],r,n,i))},Cf=t=>{Jt(t,"Add",(e,r)=>`${e}+${r}`)},Tf=t=>{Jt(t,"Div",(e,r)=>`${e}/${r}`)},If=t=>{Jt(t,"Equal",{scalar:(e,r)=>`u32(${e}==${r})`,vector:(e,r)=>`vec4<u32>(${e}==${r})`},void 0,void 0,9)},Af=t=>{Jt(t,"Mul",(e,r)=>`${e}*${r}`)},Mf=t=>{let e=Q("input",t.inputs[0].dataType,t.inputs[0].dims).type.value;Jt(t,"Pow",{scalar:(r,n)=>`pow_custom(${r},${n})`,vector:(r,n)=>`pow_vector_custom(${r},${n})`},`
    fn pow_custom(a : ${e}, b : ${e}) -> ${e} {
      if (b == ${e}(0.0)) {
        return ${e}(1.0);
      } else if (a < ${e}(0.0) && f32(b) != floor(f32(b))) {
        return ${e}(pow(f32(a), f32(b))); // NaN
      }
      return select(sign(a), ${e}(1.0), round(f32(abs(b) % ${e}(2.0))) != 1.0) * ${e}(${e==="i32"?"round":""}(pow(f32(abs(a)), f32(b))));
    }
    fn pow_vector_custom(a : vec4<${e}>, b : vec4<${e}>) -> vec4<${e}> {
      // TODO: implement vectorized pow
      return vec4<${e}>(pow_custom(a.x, b.x), pow_custom(a.y, b.y), pow_custom(a.z, b.z), pow_custom(a.w, b.w));
    }
      `)},Of=t=>{Jt(t,"Sub",(e,r)=>`${e}-${r}`)},zf=t=>{Jt(t,"Greater",{scalar:(e,r)=>`u32(${e}>${r})`,vector:(e,r)=>`vec4<u32>(${e}>${r})`},void 0,void 0,9)},Pf=t=>{Jt(t,"Less",{scalar:(e,r)=>`u32(${e}<${r})`,vector:(e,r)=>`vec4<u32>(${e}<${r})`},void 0,void 0,9)},Rf=t=>{Jt(t,"GreaterOrEqual",{scalar:(e,r)=>`u32(${e}>=${r})`,vector:(e,r)=>`vec4<u32>(${e}>=${r})`},void 0,void 0,9)},Bf=t=>{Jt(t,"LessOrEqual",{scalar:(e,r)=>`u32(${e}<=${r})`,vector:(e,r)=>`vec4<u32>(${e}<=${r})`},void 0,void 0,9)}}),Zr,en,tn,Jo,nn=Z(()=>{$e(),Ae(),Zr=(t,e,r="f32")=>{switch(t.activation){case"Relu":return`value = max(value, ${e}(0.0));`;case"Sigmoid":return`value = (${e}(1.0) / (${e}(1.0) + exp(-value)));`;case"Clip":return`value = clamp(value, ${e}(${r}(uniforms.clip_min)), ${e}(${r}(uniforms.clip_max)));`;case"HardSigmoid":return`value = max(${e}(0.0), min(${e}(1.0), ${r}(uniforms.alpha) * value + ${r}(uniforms.beta)));`;case"LeakyRelu":return`value = select(${r}(uniforms.alpha) * value, value, value >= ${e}(0.0));`;case"":return"";default:throw new Error(`Unsupported activation ${t.activation}`)}},en=(t,e)=>{t.activation==="Clip"?e.push({type:1,data:t.clipMax},{type:1,data:t.clipMin}):t.activation==="HardSigmoid"?e.push({type:1,data:t.alpha},{type:1,data:t.beta}):t.activation==="LeakyRelu"&&e.push({type:1,data:t.alpha})},tn=(t,e)=>{t.activation==="Clip"?e.push({name:"clip_max",type:"f32"},{name:"clip_min",type:"f32"}):t.activation==="HardSigmoid"?e.push({name:"alpha",type:"f32"},{name:"beta",type:"f32"}):t.activation==="LeakyRelu"&&e.push({name:"alpha",type:"f32"})},Jo=t=>{let e=t?.activation||"";if(e==="HardSigmoid"){let[r,n]=t?.activation_params||[.2,.5];return{activation:e,alpha:r,beta:n}}else if(e==="Clip"){let[r,n]=t?.activation_params||[qo,Ko];return{activation:e,clipMax:n,clipMin:r}}else if(e==="LeakyRelu"){let[r]=t?.activation_params||[.01];return{activation:e,alpha:r}}return{activation:e}}}),bt,Zo,eu=Z(()=>{bt=(t,e)=>{switch(t){case 1:return e;case 2:return`vec2<${e}>`;case 3:return`vec3<${e}>`;case 4:return`vec4<${e}>`;default:throw new Error(`${t}-component is not supported.`)}},Zo=t=>`
      ${t?"value = value + getBiasByOutputCoords(coords);":""}
      `}),tu,Df=Z(()=>{tu=t=>`
fn getIndexFromCoords4D(coords : vec4<i32>, shape : vec4<i32>) -> i32 {
  return dot(coords, vec4<i32>(
      shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1));
}
fn getOutputIndexFromCoords(coords : vec4<i32>) -> i32 {
  return dot(coords, vec4<i32>(
    i32(${t}.x), i32(${t}.y), i32(${t}.z), 1));
}
`}),sd,od,Vi,Rs,ud,Gi,ld,ru,Xi=Z(()=>{$e(),Ae(),Te(),nn(),eu(),sd=(t,e)=>t?`
        mm_Asub[inputRow][inputCol] = mm_readA(batch,
          kStart + inputRow,
          globalRowStart / innerElementSize + inputCol${e?", batchIndices":""});
        `:`
        mm_Asub[inputRow][inputCol] = mm_readA(batch,
          globalRow + innerRow,
          kStart / innerElementSize + inputCol${e?", batchIndices":""});
        `,od=(t,e)=>t?`
        let ACached0 = mm_Asub[k * innerElementSize][localRow];
        let ACached1 = mm_Asub[k * innerElementSize + 1][localRow];
        let ACached2 = mm_Asub[k * innerElementSize + 2][localRow];
        ${e===3?"":"let ACached3 = mm_Asub[k * innerElementSize + 3][localRow];"}
        for (var i = 0; i < rowPerThread; i = i + 1) {
          acc[i] = BCached0 * ACached0[i] + acc[i];
          acc[i] = BCached1 * ACached1[i] + acc[i];
          acc[i] = BCached2 * ACached2[i] + acc[i];
          ${e===3?"":"acc[i] = BCached3 * ACached3[i] + acc[i];"}
        }`:`
        for (var i = 0; i < rowPerThread; i = i + 1) {
          let ACached = mm_Asub[tileRow + i][k];
          acc[i] = BCached0 * ACached.x + acc[i];
          acc[i] = BCached1 * ACached.y + acc[i];
          acc[i] = BCached2 * ACached.z + acc[i];
          ${e===3?"":"acc[i] = BCached3 * ACached.w + acc[i];"}
        }`,Vi=(t,e,r="f32",n,a=!1,i=32,s=!1,o=32)=>{let u=e[1]*t[1],l=e[0]*t[0],p=a?u:i,f=a?i:u,m=p/e[0],c=i/e[1];if(!((a&&m===4&&t[1]===4||!a&&(m===3||m===4))&&p%e[0]===0&&i%e[1]===0&&t[0]===4))throw new Error(`If transposeA ${a} is true, innerElementSize ${m} and workPerThread[1] ${t[1]} must be 4.
      Otherwise, innerElementSize ${m} must be 3 or 4.
  tileAWidth ${p} must be divisible by workgroupSize[0]${e[0]}. tileInner ${i} must be divisible by workgroupSize[1] ${e[1]}. colPerThread ${t[0]} must be 4.`);return`
var<workgroup> mm_Asub: array<array<vec${m}<${r}>, ${p/m}>, ${f}>;
var<workgroup> mm_Bsub: array<array<vec4<${r}>, ${l/t[0]}>, ${i}>;

const rowPerThread = ${t[1]};
const colPerThread = ${t[0]};
const innerElementSize = ${m};
const tileInner = ${i};

@compute @workgroup_size(${e[0]}, ${e[1]}, ${e[2]})
fn main(@builtin(local_invocation_id) localId : vec3<u32>,
        @builtin(global_invocation_id) globalId : vec3<u32>,
        @builtin(workgroup_id) workgroupId : vec3<u32>) {
  let localRow = i32(localId.y);
  let tileRow = localRow * rowPerThread;
  let tileCol = i32(localId.x);

  let globalRow =i32(globalId.y) * rowPerThread;
  let globalCol = i32(globalId.x);
  let batch = ${s?"0":"i32(globalId.z)"};
  ${n?`let batchIndices = ${n.offsetToIndices("u32(batch)")};`:""}
  let globalRowStart = i32(workgroupId.y) * ${u};

  let num_tiles = ${s?`${Math.ceil(o/i)}`:"(uniforms.dim_inner - 1) / tileInner + 1"};
  var kStart = ${s?`i32(globalId.z) * ${o}`:"0"};

  var acc: array<vec4<${r}>, rowPerThread>;

  // Loop over shared dimension.
  let tileRowB = localRow * ${c};
  for (var t = 0; t < num_tiles; t = t + 1) {
      // Load one tile of A into local memory.
      for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
          let inputRow = tileRow + innerRow;
          let inputCol = tileCol;
          ${sd(a,n)}
      }

      // Load one tile of B into local memory.
      for (var innerRow = 0; innerRow < ${c}; innerRow = innerRow + 1) {
          let inputRow = tileRowB + innerRow;
          let inputCol = tileCol;
          mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${n?", batchIndices":""});
      }
      kStart = kStart + tileInner;
      workgroupBarrier();

      // Compute acc values for a single thread.
      for (var k = 0; k < tileInner / innerElementSize; k = k + 1) {
          let BCached0 = mm_Bsub[k * innerElementSize][tileCol];
          let BCached1 = mm_Bsub[k * innerElementSize + 1][tileCol];
          let BCached2 = mm_Bsub[k * innerElementSize + 2][tileCol];
          ${m===3?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"}

          ${od(a,m)}
      }

      workgroupBarrier();
  }

  for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
      mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]);
  }
}`},Rs=(t,e)=>t?`
            mm_Asub[inputRow][inputCol] = mm_readA(batch,
              kStart + inputRow,
              globalRowStart + inputCol${e?", batchIndices":""});
            `:`
            mm_Asub[inputRow][inputCol] = mm_readA(batch,
              globalRowStart + inputRow,
              kStart + inputCol${e?", batchIndices":""});
            `,ud=t=>t?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",Gi=(t,e,r="f32",n,a=!1,i=32,s=!1,o=32,u=!1)=>{let l=t[1]*e[1],p=t[0]*e[0],f=a?l:i,m=a?i:l;if(!(m%e[1]===0&&f%e[0]===0&&i%e[1]===0))throw new Error(`tileAHight ${m} must be divisible by workgroupSize[1]${e[1]}, tileAWidth ${f} must be divisible by workgroupSize[0]${e[0]}, tileInner ${i} must be divisible by workgroupSize[1]${e[1]}`);let c=m/e[1],y=f/e[0],w=i/e[1],v=u?`
    let localRow = i32(localId.y);
    let localCol = i32(localId.x);
    let globalRowStart = i32(workgroupId.y) * ${l};
    let globalColStart = i32(workgroupId.x) * ${p};

    // Loop over shared dimension.
    for (var t = 0; t < num_tiles; t = t + 1) {
      // Load one tile of A into local memory.
      for (var inputRow = localRow; inputRow < ${m}; inputRow = inputRow + ${e[1]}) {
        for (var inputCol = localCol; inputCol < ${f}; inputCol = inputCol + ${e[0]}) {
          ${Rs(a,n)}
        }
      }
      // Load one tile of B into local memory.
      for (var inputRow = localRow; inputRow < ${i}; inputRow = inputRow + ${e[1]}) {
            for (var inputCol = localCol; inputCol < ${p}; inputCol = inputCol + ${e[0]}) {
          mm_Bsub[inputRow][inputCol] = mm_readB(batch,
            kStart + inputRow,
            globalColStart + inputCol${n?", batchIndices":""});
        }
      }
      kStart = kStart + tileInner;
      workgroupBarrier();

      // Compute acc values for a single thread.
      var BCached : array<${r}, colPerThread>;
      for (var k = 0; k < tileInner; k = k + 1) {
        for (var inner = 0; inner < colPerThread; inner = inner + 1) {
          BCached[inner] = mm_Bsub[k][localCol + inner * ${e[0]}];
        }
        for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
          let ACached = ${a?`mm_Asub[k][localRow + innerRow * ${e[1]}];`:`mm_Asub[localRow + innerRow * ${e[1]}][k];`}
          for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
            acc[innerRow][innerCol] = acc[innerRow][innerCol] +
                ACached * BCached[innerCol];
          }
        }
      }
      workgroupBarrier();
    }
    for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
      let gRow = globalRowStart + localRow + innerRow * ${e[1]};
      for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
        let gCol = globalColStart + localCol + innerCol * ${e[0]};
        mm_write(batch, gRow, gCol, acc[innerRow][innerCol]);
      }
    }
    `:`
let tileRow = i32(localId.y) * rowPerThread;
let tileCol = i32(localId.x) * colPerThread;

let globalRow = i32(globalId.y) * rowPerThread;
let globalCol = i32(globalId.x) * colPerThread;
let globalRowStart = i32(workgroupId.y) * ${l};

let tileRowA = i32(localId.y) * ${c};
let tileColA = i32(localId.x) * ${y};
let tileRowB = i32(localId.y) * ${w};
// Loop over shared dimension.
for (var t = 0; t < num_tiles; t = t + 1) {
  // Load one tile of A into local memory.
  for (var innerRow = 0; innerRow < ${c}; innerRow = innerRow + 1) {
    for (var innerCol = 0; innerCol < ${y}; innerCol = innerCol + 1) {
      let inputRow = tileRowA + innerRow;
      let inputCol = tileColA + innerCol;
      ${Rs(a,n)}
    }
  }

  // Load one tile of B into local memory.
  for (var innerRow = 0; innerRow < ${w}; innerRow = innerRow + 1) {
    for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
      let inputRow = tileRowB + innerRow;
      let inputCol = tileCol + innerCol;
      mm_Bsub[inputRow][inputCol] = mm_readB(batch,
        kStart + inputRow,
        globalCol + innerCol${n?", batchIndices":""});
    }
  }
  kStart = kStart + tileInner;
  workgroupBarrier();

  // Compute acc values for a single thread.
  var BCached : array<${r}, colPerThread>;
  for (var k = 0; k < tileInner; k = k + 1) {
    for (var inner = 0; inner < colPerThread; inner = inner + 1) {
      BCached[inner] = mm_Bsub[k][tileCol + inner];
    }

    for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
      ${ud(a)}
      for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
        acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol];
      }
    }
  }

  workgroupBarrier();
}

for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
  for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
    mm_write(batch, globalRow + innerRow, globalCol + innerCol,
        acc[innerRow][innerCol]);
  }
}
`;return`
  var<workgroup> mm_Asub : array<array<${r}, ${f}>, ${m}>;
  var<workgroup> mm_Bsub : array<array<${r}, ${p}>, ${i}>;
  const rowPerThread = ${t[1]};
  const colPerThread = ${t[0]};
  const tileInner = ${i};

@compute @workgroup_size(${e[0]}, ${e[1]}, ${e[2]})
fn main(@builtin(local_invocation_id) localId : vec3<u32>,
        @builtin(global_invocation_id) globalId : vec3<u32>,
        @builtin(workgroup_id) workgroupId : vec3<u32>) {
    let batch = ${s?"0":"i32(globalId.z)"};
    ${n?`let batchIndices = ${n.offsetToIndices("u32(batch)")};`:""}
    let num_tiles = ${s?`${Math.ceil(o/i)}`:"(uniforms.dim_inner - 1) / tileInner + 1"};
    var kStart = ${s?`i32(globalId.z) * ${o}`:"0"};

    var acc : array<array<${r}, colPerThread>, rowPerThread>;

    // Without this initialization strange values show up in acc.
    for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
      for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
        acc[innerRow][innerCol] = 0.0;
      }
    }
    ${v}
  }
`},ld=(t,e,r,n,a,i=!1)=>{let[s,o,u]=a,[l,p,f,m]=n,c=ma(s,u),y=ma(o,u),w=pt(n[0].type.tensor),v=()=>{let $=p.rank,C=l.rank,T=`var aIndices: ${p.type.indices};`;for(let A=$-2-1,B=C-1;A>=0;A--,B--)T+=`
aIndices[${A}] = ${C>1?`batchIndices[${B}]`:"batchIndices"};`;return c.forEach(A=>{T+=`
aIndices[${A}] = 0;`}),T+=`
aIndices[${$-2}] = u32(row);
                   aIndices[${$-1}] = u32(colIn);`,T},k=()=>{let $=f.rank,C=l.rank,T=`var bIndices: ${f.type.indices};`;for(let A=$-2-1,B=C-1;A>=0;A--,B--)T+=`
bIndices[${A}] = ${C>1?`batchIndices[${B}]`:"batchIndices"};`;return y.forEach(A=>{T+=`
bIndices[${A}] = 0;`}),T+=`
bIndices[${$-2}] = u32(row);
                   bIndices[${$-1}] = u32(colIn);`,T};return`
    fn mm_readA(batch: i32, row: i32, colIn: i32, batchIndices: ${l.type.indices}) -> ${bt(t,w)} {
      var value = ${bt(t,w)}(0.0);
      let col = colIn * ${t};
      if(row < uniforms.dim_a_outer && col < uniforms.dim_inner)
      {
        ${v()}
        value = ${p.getByIndices("aIndices")};
      }
      return value;
    }

    fn mm_readB(batch: i32, row: i32, colIn: i32, batchIndices: ${l.type.indices}) -> ${bt(t,w)} {
      var value = ${bt(t,w)}(0.0);
      let col = colIn * ${t};
      if(row < uniforms.dim_inner && col < uniforms.dim_b_outer)
      {
        ${k()}
        value = ${f.getByIndices("bIndices")};
      }
      return value;
    }

    fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${bt(t,w)}) {
      let col = colIn * ${t};
      if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) {
        var value = valueIn;
        let coords = vec3<i32>(batch, row, colIn);
        ${e?`value = value + ${i?"bias[colIn]":`${bt(t,w)}(bias[row])`};`:""}
        ${r}
        ${m.setByIndices("vec3<u32>(coords)","value")}
      }
    }
    `},ru=(t,e,r,n,a=!1)=>{let i=t[0].dims,s=t[1].dims,o=i.slice(0,-2),u=s.slice(0,-2),l=n?n.slice(0,-2):r.slice(0,-2),p=Y.size(l),f=i[i.length-2],m=i[i.length-1],c=s[s.length-1],y=m%4===0&&c%4===0,w=f<=8?[4,1,1]:[4,4,1],v=[8,8,1],k=[Math.ceil(c/v[0]/w[0]),Math.ceil(f/v[1]/w[1]),Math.ceil(p/v[2]/w[2])],$=y?4:1,C=[...o,f,m/$],T=C.length,A=[...u,m,c/$],B=A.length,R=[p,f,c/$],D=[{type:6,data:f},{type:6,data:c},{type:6,data:m}];en(e,D),D.push(...ye(l,C,A));let K=["rank","rank"],j=t.length>2;j&&(D.push(...ye(t[2].dims)),K.push("rank")),D.push(...ye(R));let ie=te=>{let oe=l.length,re=Yo("batchDims",t[0].dataType,oe,1),M=pt(t[0].dataType),P=Q("a",t[0].dataType,T,$),H=Q("b",t[1].dataType,B,$),le=ge("result",t[0].dataType,R.length,$),G=[P,H];if(j){let Ce=a?$:1;G.push(Q("bias",t[2].dataType,t[2].dims.length,Ce))}let ne=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"}];tn(e,ne);let N=pt(le.type.tensor),ae=Zr(e,le.type.value,N),fe=ld($,j,ae,[re,P,H,le],[o,u,l],a);return`
  ${te.registerUniforms(ne).registerInternalVariables(re).declareVariables(...G,le)}
  ${fe}
  ${y?Vi(w,v,M,re):Gi(w,v,M,re)}
                   `};return{name:"MatMul",shaderCache:{hint:`${w};${e.activation};${y};${a}`,inputDependencies:K},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:k[0],y:k[1],z:k[2]},programUniforms:D}),getShaderSource:ie}}}),dd,Nf,iy=Z(()=>{$e(),rn(),Te(),nn(),eu(),Df(),Xi(),dd=(t,e,r,n,a=!1,i,s=4,o=4,u=4,l="f32")=>{let p=K=>{switch(K){case 1:return"resData = x[xIndex];";case 3:return`resData = vec3<${l}>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);`;case 4:return"resData = x[xIndex / 4];";default:throw new Error(`innerElementSize ${K} is not supported.`)}},f=K=>{switch(K){case 1:return"return w[row * i32(uniforms.w_shape[3]) + colIn];";case 4:return"return w[row * i32(uniforms.w_shape[3]) / 4 + colIn];";default:throw new Error(`innerElementSize ${K} is not supported.`)}},m=t?`
    let coord = vec4<i32>(batch, xRow, xCol, xCh);
    `:`
    let coord = vec4<i32>(batch, xCh, xRow, xCol);
    `,c=t?`
    let coords = vec4<i32>(
      batch,
      row / outWidth,
      row % outWidth,
      col);
    `:`
    let coords = vec4<i32>(
      batch,
      row,
      col / outWidth,
      col % outWidth);
    `,y=t?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",w=t?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",v=t?"row":"col",k=t?"col":"row",$=`
    let inChannels = i32(uniforms.w_shape[2]);
    let outWidth = ${t?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};
    let outRow = ${v} / outWidth;
    let outCol = ${v} % outWidth;

    let WRow = ${k} / (i32(uniforms.w_shape[1]) * inChannels);
    let WCol = ${k} / inChannels % i32(uniforms.w_shape[1]);
    let xRow = outRow * uniforms.stride[0] + uniforms.dilation[0] * WRow - uniforms.pad[0];
    let xCol = outCol * uniforms.stride[1] + uniforms.dilation[1] * WCol - uniforms.pad[1];
    let xCh = ${k} % inChannels;
    var resData = ${bt(s,l)}(0.0);
    // The bounds checking is always needed since we use it to pad zero for
    // the 'same' padding type.
    if (xRow >= 0 && xRow < ${y} && xCol >= 0 && xCol < ${w}) {
      ${m}
      let xIndex = getIndexFromCoords4D(coord, vec4<i32>(uniforms.x_shape));
      ${p(s)}
    }
    return resData;`,C=t?e&&n?`
    let col = colIn * ${s};
    ${$}`:`
    let col = colIn * ${s};
    if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) {
      ${$}
    }
    return ${bt(s,l)}(0.0);`:n&&r?`
    let col = colIn * ${s};
    ${$}`:`
    let col = colIn * ${s};
    if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) {
      ${$}
    }
    return ${bt(s,l)}(0.0);`,T=`${f(o)}`,A=bt(u,l),B=bt(t?s:o,l),R=bt(t?o:s,l),D=Zr(i,A,l);return`
    fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${B} {
      ${t?C:T}
    }

    fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${R} {
      ${t?T:C}
    }

    fn mm_write(batch: i32, row : i32, colIn : i32, valueIn : ${A}) {
      let col = colIn * ${u};
      if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer)
      {
      var value = valueIn;
      let outWidth = ${t?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};
      ${c}
      ${Zo(a)}
      ${D}
      setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
      }
    }`},Nf=(t,e,r,n,a,i,s,o)=>{let u=e.format==="NHWC",l=u?t[0].dims[3]:t[0].dims[1],p=r[0],f=u?r[2]:r[3],m=u?r[1]:r[2],c=u?r[3]:r[1],y=u&&(l%4===0||l%3===0)&&c%4===0,w=u?c:f*m,v=u?f*m:c,k=[8,8,1],$=n<=8?[4,1,1]:[4,4,1],C=[Math.ceil(w/k[0]/$[0]),Math.ceil(v/k[1]/$[1]),Math.ceil(p/k[2]/$[2])];rt("verbose",()=>`[conv2d_mm_webgpu] dispatch = ${C}`);let T=y?u&&l%4!==0?3:4:1,A=k[1]*$[1],B=k[0]*$[0],R=Math.max(k[0]*T,k[1]),D=n%A===0,K=a%B===0,j=i%R===0,ie=y?[T,4,4]:[1,1,1],te=[{type:6,data:n},{type:6,data:a},{type:6,data:i},{type:6,data:[e.pads[0],e.pads[1]]},{type:6,data:e.strides},{type:6,data:e.dilations}];en(e,te),te.push(...ye(t[0].dims,t[1].dims));let oe=["rank","rank"];s&&(te.push(...ye(t[2].dims)),oe.push("rank")),te.push(...ye(r));let re=M=>{let P=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"},{name:"pad",type:"i32",length:2},{name:"stride",type:"i32",length:2},{name:"dilation",type:"i32",length:2}];tn(e,P);let H=y?4:1,le=pt(t[0].dataType),G=`
      fn setOutputAtIndex(flatIndex : i32, value : ${y?`vec4<${le}>`:le}) {
        result[flatIndex] = ${y?`vec4<${le}>`:le}(value);
      }
      fn setOutputAtCoords(d0 : i32, d1 : i32, d2 : i32, d3 : i32, value : ${y?`vec4<${le}>`:le}) {
        let flatIndex = getOutputIndexFromCoords(vec4<i32>(d0, d1, d2, d3));
        setOutputAtIndex(flatIndex ${y?"/ 4":""}, value);
      }`,ne=Q("x",t[0].dataType,t[0].dims.length,T===3?1:T),N=Q("w",t[1].dataType,t[1].dims.length,H),ae=[ne,N],fe=ge("result",t[0].dataType,r.length,H);if(s){let Ce=Q("bias",t[2].dataType,t[2].dims.length,H);ae.push(Ce),G+=`
        fn getBiasByOutputCoords(coords : vec4<i32>) -> ${y?`vec4<${le}>`:le} {
          return bias[coords.${u?"w":"y"}${y?"/ 4":""}];
        }`}return`
        ${tu("uniforms.result_strides")}
        //struct Uniforms { xShape : vec4<i32>, wShape : vec4<i32>, outShape : vec4<i32>,
        //  outShapeStrides: vec3<i32>, filterDims : vec2<i32>, pad : vec2<i32>, stride : vec2<i32>,
        //  dilation : vec2<i32>, dimAOuter : i32, dimBOuter : i32, dimInner : i32 };
        ${M.registerUniforms(P).declareVariables(...ae,fe)}
        ${G}
        ${dd(u,D,K,j,s,e,ie[0],ie[1],ie[2],le)}
        ${y?Vi($,k,le,void 0,!u,R):Gi($,k,le,void 0,!u,R,!1,void 0,o)}`};return{name:"Conv2DMatMul",shaderCache:{hint:`${e.cacheKey};${T};${y};${D};${K};${j};${A};${B};${R}`,inputDependencies:oe},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:C[0],y:C[1],z:C[2]},programUniforms:te}),getShaderSource:re}}}),$o,Ff,sy=Z(()=>{$e(),Ae(),Te(),Wf(),nn(),$o=(t,e,r)=>{let n=t.length>2,a=n?"value += b[output_channel];":"",i=t[0].dims,s=t[1].dims,o=s[0]/e.group,u=e.format==="NHWC",l=Ii(i,s,e.dilations,e.pads,e.strides,u),p=Y.size(l),f=[{type:12,data:p},{type:12,data:e.dilations},{type:12,data:[e.strides[0],e.strides[1]]},{type:12,data:[e.pads[0],e.pads[1]]},{type:12,data:o}];en(e,f),f.push(...ye(i,s));let m=["rank","rank"];n&&(f.push(...ye(t[2].dims)),m.push("rank")),f.push(...ye(l));let c=y=>{let w=ge("output",t[0].dataType,l.length),v=pt(w.type.tensor),k=Zr(e,w.type.value,v),$=Q("x",t[0].dataType,i.length),C=Q("w",t[1].dataType,s.length),T=[$,C];n&&T.push(Q("b",t[2].dataType,t[2].dims.length));let A=[{name:"output_size",type:"u32"},{name:"dilations",type:"u32",length:e.dilations.length},{name:"strides",type:"u32",length:2},{name:"pads",type:"u32",length:2},{name:"output_channels_per_group",type:"u32"}];return tn(e,A),`
  ${y.registerUniforms(A).declareVariables(...T,w)}

  ${y.mainStart()}
    ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}

    let outputIndices = ${w.offsetToIndices("global_idx")};
    let batch: u32 = outputIndices[0];
    let output_channel: u32 = outputIndices[${u?3:1}];
    let xRCCorner: vec2<u32> = vec2<u32>(outputIndices[${u?1:2}], outputIndices[${u?2:3}]) * uniforms.strides - uniforms.pads;
    let group_id: u32 = output_channel / uniforms.output_channels_per_group;

    var value: ${w.type.value} = ${w.type.value}(0);
    for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[1]; wInChannel++) {
      let input_channel = group_id * uniforms.w_shape[1] + wInChannel;
      for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[2]; wHeight++) {
        let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0];

        if (xHeight < 0u || xHeight >= uniforms.x_shape[${u?1:2}]) {
          continue;
        }

        for (var wWidth: u32 = 0u; wWidth < uniforms.w_shape[3]; wWidth++) {
          let xWidth = xRCCorner.y + wWidth * uniforms.dilations[1];
          if (xWidth < 0u || xWidth >= uniforms.x_shape[${u?2:3}]) {
            continue;
          }

          let xVal = ${u?$.get("batch","xHeight","xWidth","input_channel"):$.get("batch","input_channel","xHeight","xWidth")};
          let wVal = ${C.get("output_channel","wInChannel","wHeight","wWidth")};
          value += xVal*wVal;
        }
      }
    }
    ${a}
    ${k}
    ${w.setByOffset("global_idx","value")}
  }`};return{name:"GroupedConv",shaderCache:{hint:e.cacheKey,inputDependencies:m},getRunData:()=>({outputs:[{dims:r?r(l):l,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:f}),getShaderSource:c}},Ff=(t,e,r)=>{let n=t.length>2,a=it(r[3]),i=it(r[2]),s=Y.size(r)/a/i,o=[t[0].dims[0],t[0].dims[1],t[0].dims[2],t[0].dims[3]/a],u=[t[1].dims[0],t[1].dims[1],t[1].dims[2],t[1].dims[3]/a],l=[r[0],r[1],r[2],r[3]/a],p=[{type:12,data:s},{type:6,data:[e.strides[0],e.strides[1]]},{type:6,data:[e.pads[0],e.pads[1]]}];en(e,p),p.push(...ye(o,u,l));let f=(i-1)*e.strides[1]+u[1],m=c=>{let y=ge("output",t[0].dataType,l.length,a),w=pt(y.type.tensor),v=Zr(e,y.type.value,w),k=Q("x",t[0].dataType,o.length,a),$=Q("w",t[1].dataType,u.length,a),C=[k,$];n&&C.push(Q("b",t[2].dataType,t[2].dims,a));let T=n?"value += b[output_channel];":"",A=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return tn(e,A),`
  ${c.registerUniforms(A).declareVariables(...C,y)}
  ${c.mainStart()}
    ${c.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
    let width0 = uniforms.output_shape[3];
    let output_channel = global_idx % width0;
    var index1 = global_idx / width0;
    let width1 = uniforms.output_shape[2] / ${i}u;
    let col = (index1 % width1) * ${i}u;
    index1 = index1 / width1;
    let row = index1 % uniforms.output_shape[1];
    let batch = index1 / uniforms.output_shape[1];

    let x_corner = vec2<i32>(i32(row), i32(col)) * uniforms.strides - uniforms.pads;

    var x_vals: array<${k.type.value}, ${f}>;
    var values: array<${y.type.value}, ${i}>;
    let input_channel = output_channel;
    // Use constant instead of uniform can give better performance for w's height/width.
    for (var w_height: u32 = 0u; w_height < ${u[0]}; w_height++) {
      let x_height = x_corner.x + i32(w_height);
      if (x_height >= 0 && u32(x_height) < uniforms.x_shape[1]) {
        for (var i = 0; i < ${f}; i++) {
          let x_width = x_corner.y + i;
          if (x_width >= 0 && u32(x_width) < uniforms.x_shape[2]) {
            x_vals[i] = ${k.get("batch","u32(x_height)","u32(x_width)","input_channel")};
          } else {
            x_vals[i] = ${k.type.value}(0);
          }
        }
        for (var w_width: u32 = 0u; w_width < ${u[1]}; w_width++) {
          let w_val = ${$.get("w_height","w_width","0","output_channel")};
          for (var i = 0u; i < ${i}u; i++) {
            values[i] = fma(x_vals[i * u32(uniforms.strides[1]) + w_width], w_val, values[i]);
          }
        }
      }
    }

    for (var i = 0u; i < ${i}u; i++) {
      var value = values[i];
      ${T}
      ${v}
      ${y.set("batch","row","col + i","output_channel","value")};
    }
  }`};return{name:"GroupedConv-Vectorize",shaderCache:{hint:`${e.cacheKey};${a};${i};${f};${u[0]};${u[1]}`,inputDependencies:n?["rank","rank","type"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:p}),getShaderSource:m}}}),xo,cd,Lf,Uf=Z(()=>{$e(),Ae(),Xi(),Te(),nn(),xo=(t,e,r,n,a=!1)=>{let i=t[0].dims,s=t[1].dims,o=i[i.length-2],u=s[s.length-1],l=i[i.length-1],p=it(u),f=it(l),m=it(o),c=Y.size(r)/p/m,y=t.length>2,w=n?n.slice(0,-2):r.slice(0,-2),v=[Y.size(w),o,u],k=[{type:12,data:c},{type:12,data:o},{type:12,data:u},{type:12,data:l}];en(e,k),k.push(...ye(w,i,s)),y&&k.push(...ye(t[2].dims)),k.push(...ye(v));let $=C=>{let T=Yo("batch_dims",t[0].dataType,w.length),A=Q("a",t[0].dataType,i.length,f),B=Q("b",t[1].dataType,s.length,p),R=ge("output",t[0].dataType,v.length,p),D=pt(R.type.tensor),K=Zr(e,R.type.value,D),j=[A,B],ie="";if(y){let G=a?p:1;j.push(Q("bias",t[2].dataType,t[2].dims.length,G)),ie=`${a?`value += bias[col / ${G}];`:`value += ${R.type.value}(bias[row + i]);`}`}let te=i.slice(0,-2),oe=s.slice(0,-2),re=ma(te,w),M=ma(oe,w),P=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];tn(e,P);let H=(G,ne)=>{let N=G.rank,ae=G.name;if(N===2)return`var ${ae}_indices = ${G.type.indices}(0u, 0u);`;let fe=T.rank,Ce=`var ${ae}_indices: ${G.type.indices};`;for(let Be=N-2-1,Ke=fe-1;Be>=0;Be--,Ke--)Ce+=`
${ae}_indices[${Be}] = ${fe>1?`batch_indices[${Ke}]`:"batch_indices"};`;return ne.forEach(Be=>{Ce+=`
${ae}_indices[${Be}] = 0;`}),Ce+=`${ae}_indices[${N-2}] = 0u;
                     ${ae}_indices[${N-1}] = 0u;`,Ce},le=()=>{let G=`var a_data: ${A.type.value};`;for(let ne=0;ne<f;ne++)G+=`
              let b_data${ne} = b[(b_offset + (k + ${ne}) * uniforms.N + col) / ${p}];`;for(let ne=0;ne<m;ne++){G+=`a_data = a[(a_offset + (row + ${ne}) * uniforms.K + k) / ${f}];`;for(let N=0;N<f;N++)G+=`
            values[${ne}] = fma(${B.type.value}(a_data${f===1?"":`[${N}]`}), b_data${N}, values[${ne}]);
`}return G};return`
  ${C.registerUniforms(P).registerInternalVariables(T).declareVariables(...j,R)}
  ${C.mainStart()}
    ${C.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
    let col = (global_idx % (uniforms.N / ${p})) * ${p};
    var index1 = global_idx / (uniforms.N / ${p});
    let stride1 = uniforms.M / ${m};
    let row = (index1 % stride1) * ${m};
    let batch = index1 / stride1;

    ${r.length===2?"":`let batch_indices = ${T.offsetToIndices("batch")};`}
    ${H(A,re)}
    let a_offset = ${A.indicesToOffset("a_indices")};
    ${H(B,M)}
    let b_offset = ${B.indicesToOffset("b_indices")};
    var values: array<${R.type.value}, ${m}>;
    for (var k: u32 = 0u; k < uniforms.K; k = k + ${f}) {
      ${le()}
    }
    for (var i = 0u; i < ${m}u; i++) {
      var value = values[i];
      ${ie}
      ${K}
      let cur_indices = ${R.type.indices}(batch, row + i, col);
      let offset = ${R.indicesToOffset("cur_indices")};
      ${R.setByOffset(`offset / ${p}`,"value")};
    }
  }
  `};return{name:"MatMulNaive",shaderCache:{hint:`${e.activation};${p};${f};${m};${a}`,inputDependencies:y?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(c/64)},programUniforms:k}),getShaderSource:$}},cd=t=>{if(!t||t.length!==2)throw new Error("MatMul requires 2 inputs.");if(t[0].dims[t[0].dims.length-1]!==t[1].dims[t[1].dims.length-2])throw new Error("shared dimension does not match.")},Lf=t=>{cd(t.inputs);let e=kn.calcShape(t.inputs[0].dims,t.inputs[1].dims,!0);if(!e)throw new Error("Can't use matmul on the given tensors");let r=e[e.length-1],n=t.inputs[0].dims[t.inputs[0].dims.length-1];r<8&&n<8?t.compute(xo(t.inputs,{activation:""},e)):t.compute(ru(t.inputs,{activation:""},e))}}),Ii,mi,pd,Bs,So,hd,fd,ko,Wf=Z(()=>{Ae(),iy(),Xi(),sy(),nn(),Uf(),$a(),Ii=(t,e,r,n,a,i)=>{let s=t[0],o=t.slice(i?1:2,i?3:4),u=o.length,l=e[0],p=e.slice(2).map((m,c)=>m+(m-1)*(r[c]-1)),f=o.map((m,c)=>m+n[c]+n[c+u]).map((m,c)=>Math.floor((m-p[c]+a[c])/a[c]));return f.splice(0,0,s),f.splice(i?3:1,0,l),f},mi=[2,3,1,0],pd=(t,e)=>{if(!t||t.length!==2&&t.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(t[0].dims.length!==4&&t[0].dims.length!==3)throw new Error("currently only support conv 1D and 2D");if(t[0].dims.length!==t[1].dims.length)throw new Error("filter does not have same dimension as input");let r=t[0].dims[e.format==="NHWC"?t[0].dims.length-1:1],n=t[1].dims[1]*e.group;if(r!==n)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");if(t.length===3&&(t[2].dims.length!==1||t[1].dims[0]!==t[2].dims[0]))throw new Error("invalid bias");let a=t[0].dims.length-2;if(e.dilations.length!==a)throw new Error(`dilations should be ${a}D`);if(e.strides.length!==a)throw new Error(`strides should be ${a}D`);if(e.pads.length!==a*2)throw new Error(`pads should be ${a*2}D`);if(e.kernelShape.length!==0&&e.kernelShape.length!==t[1].dims.length-2)throw new Error("invalid kernel shape")},Bs=(t,e)=>{let r=t.kernelShape.slice();for(let i=2;i<e[1].dims.length;++i)r[i-2]===0&&(r[i-2]=e[1].dims[i]);let n=t.pads.slice();Fi.adjustPadsBasedOnAutoPad(e[0].dims,t.strides,t.dilations,r,n,t.format==="NHWC",t.autoPad);let a=Object.assign({},t);return Object.assign(a,{kernelShape:r,pads:n}),a},So=t=>{let e=Jo(t),r=t.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][t.auto_pad],a=t.dilations,i=t.group,s=t.kernel_shape,o=t.pads,u=t.strides,l=t.w_is_const();return{autoPad:n,format:r,dilations:a,group:i,kernelShape:s,pads:o,strides:u,wIsConst:l,...e,cacheKey:`${t.format};${e.activation};`}},hd=(t,e,r)=>{let n=Bs(r,e),a=r.format==="NHWC";if(r.group!==1){if(!t.adapterInfo.isArchitecture("ampere")&&a&&e[1].dims[0]===r.group&&e[1].dims[1]===1&&r.dilations[0]===1&&r.dilations[1]===1){let B=Ii(e[0].dims,e[1].dims,r.dilations,n.pads,r.strides,a),R=t.kernelCustomData.wT??t.compute(br(e[1],mi),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!t.kernelCustomData.wT&&(t.kernelCustomData.wT=R);let D=[e[0],R];e.length===3&&D.push(e[2]),t.compute(Ff(D,n,B),{inputs:D})}else t.compute($o(e,n));return}let i=e.length===3,s=e[0].dims[a?1:2],o=e[0].dims[a?2:3],u=e[0].dims[a?3:1],l=e[1].dims[2],p=e[1].dims[3],f=Ii(e[0].dims,e[1].dims,r.dilations,n.pads,r.strides,a),m=f[a?1:2],c=f[a?2:3],y=f[a?3:1],w=a&&l===s&&p===o&&r.pads[0]===0&&r.pads[1]===0;if(w||l===1&&p===1&&r.dilations[0]===1&&r.dilations[1]===1&&r.strides[0]===1&&r.strides[1]===1&&r.pads[0]===0&&r.pads[1]===0){let B=f[0],R,D,K,j=[];if(a){let oe=t.kernelCustomData.wT??t.compute(br(e[1],mi),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!t.kernelCustomData.wT&&(t.kernelCustomData.wT=oe),w){let re=s*o*u;R=e[0].reshape([1,B,re]),D=oe.reshape([1,re,y]),K=[1,B,y]}else R=e[0].reshape([B,s*o,u]),D=oe.reshape([1,u,y]),K=[B,m*c,y];j.push(R),j.push(D)}else R=e[0].reshape([B,u,s*o]),D=e[1].reshape([1,y,u]),K=[B,y,m*c],j.push(D),j.push(R);i&&j.push(e[2]);let ie=K[2],te=j[0].dims[j[0].dims.length-1];ie<8&&te<8?t.compute(xo(j,n,f,K,a),{inputs:j}):t.compute(ru(j,n,f,K,a),{inputs:j});return}let v=!0,k=t.kernelCustomData.wT??t.compute(br(e[1],mi),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!t.kernelCustomData.wT&&(t.kernelCustomData.wT=k);let $=[e[0],k];i&&$.push(e[2]);let C=a?m*c:y,T=a?y:m*c,A=l*p*u;t.compute(Nf($,n,f,C,T,A,i,v),{inputs:$})},fd=(t,e)=>{let r=e.format==="NHWC",n=[t.inputs[0].reshape(r?[t.inputs[0].dims[0],1,t.inputs[0].dims[1],t.inputs[0].dims[2]]:[t.inputs[0].dims[0],t.inputs[0].dims[1],1,t.inputs[0].dims[2]]),t.inputs[1].reshape([t.inputs[1].dims[0],t.inputs[1].dims[1],1,t.inputs[1].dims[2]])];t.inputs.length===3&&n.push(t.inputs[2]);let a=[0,e.pads[0],0,e.pads[1]],i=[1].concat(e.strides),s=[1].concat(e.dilations),o=[1].concat(e.kernelShape),u=Bs({...e,pads:a,strides:i,dilations:s,kernelShape:o},n);t.compute($o(n,u,l=>r?[l[0],l[2],l[3]]:[]))},ko=(t,e)=>{pd(t.inputs,e),t.inputs[0].dims.length===3?fd(t,e):hd(t,t.inputs,e)}}),md,Vf,oy=Z(()=>{$e(),rn(),Te(),nn(),eu(),Df(),Xi(),md=(t,e=!1,r,n,a=4)=>{let i=v=>{switch(v){case 1:return"return w[getIndexFromCoords4D(coord, vec4<i32>(uniforms.w_shape))];";case 4:return`
            let coord1 = vec4<i32>(coordX, coordY, col + 1, rowInner);
            let coord2 = vec4<i32>(coordX, coordY, col + 2, rowInner);
            let coord3 = vec4<i32>(coordX, coordY, col + 3, rowInner);
            let v0 = w[getIndexFromCoords4D(coord, vec4<i32>(uniforms.w_shape))];
            let v1 = w[getIndexFromCoords4D(coord1, vec4<i32>(uniforms.w_shape))];
            let v2 = w[getIndexFromCoords4D(coord2, vec4<i32>(uniforms.w_shape))];
            let v3 = w[getIndexFromCoords4D(coord3, vec4<i32>(uniforms.w_shape))];
            return ${n}(v0, v1, v2, v3);
            `;default:throw new Error(`innerElementSize ${v} is not supported.`)}},s=t?`
      let coord = vec4<i32>(batch, iXR, iXC, xCh);
      `:`
      let coord = vec4<i32>(batch, xCh, iXR, iXC);
      `,o=t?`
    let coords = vec4<i32>(
      batch,
      row / outWidth,
      row % outWidth,
      col);
    `:`
    let coords = vec4<i32>(
      batch,
      row,
      col / outWidth,
      col % outWidth);
    `,u=t?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",l=t?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",p=t?"row":"col",f=t?"col":"row",m=`
      let inChannels = ${t?"i32(uniforms.x_shape[3])":"i32(uniforms.x_shape[1])"};
      let outWidth = ${t?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};
      let outRow = ${p} / outWidth;
      let outCol = ${p} % outWidth;

      let WRow = ${f} / (uniforms.filter_dims[1] * inChannels);
      let WCol = ${f} / inChannels % uniforms.filter_dims[1];
      let xR = f32(outRow - uniforms.pads[0] + uniforms.dilations[0] * WRow) / f32(uniforms.strides[0]);
      let xC = f32(outCol - uniforms.pads[1] + uniforms.dilations[1] * WCol) / f32(uniforms.strides[1]);
      if (xR < 0.0 || xR >= f32(${u}) || fract(xR) > 0.0) {
        return ${n}(0.0);
      }
      if (xC < 0.0 || xC >= f32(${l}) || fract(xC) > 0.0) {
        return ${n}(0.0);
      }
      let iXR = i32(xR);
      let iXC = i32(xC);
      let xCh = ${f} % inChannels;
      ${s}
      return x[getIndexFromCoords4D(coord, vec4<i32>(uniforms.x_shape))/${a}];`,c=t?`
      let col = colIn * ${a};
      if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) {
        ${m}
      }
      return ${n}(0.0);`:`
      let col = colIn * ${a};
      if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) {
        ${m}
      }
      return ${n}(0.0);`,y=`
      let col = colIn * ${a};
      let inChannels = ${t?"i32(uniforms.x_shape[3])":"i32(uniforms.x_shape[1])"};
      let coordX = uniforms.filter_dims[0] - 1 - row / (uniforms.filter_dims[1] * inChannels);
      let coordY = uniforms.filter_dims[1] - 1 - (row / inChannels) % uniforms.filter_dims[1];
      if (${t?"row < uniforms.dim_inner && col < uniforms.dim_b_outer":"row < uniforms.dim_inner && col < uniforms.dim_a_outer"}  && coordX >= 0 && coordY >= 0) {
        let rowInner = row % inChannels;
        let coord = vec4<i32>(coordX, coordY, col, rowInner);
        ${i(a)}
      }
      return ${n}(0.0);
      `,w=Zr(r,n);return`
  fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${n} {
    ${t?c:y}
  }

  fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${n} {
    ${t?y:c}
  }

  fn mm_write(batch: i32, row : i32, colIn : i32, valueInput : ${n}) {
    let col = colIn * ${a};
    if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) {
      var value = valueInput;
      let outWidth = ${t?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};
      ${o}
      ${Zo(e)}
      ${w}
      result[getIndexFromCoords4D(coords, vec4<i32>(uniforms.result_shape))/${a}] = value;
    }
  }`},Vf=(t,e,r,n,a,i,s,o)=>{let u=e.format==="NHWC",l=u?t[0].dims[3]:t[0].dims[1],p=r[0],f=u?r[2]:r[3],m=u?r[1]:r[2],c=u?r[3]:r[1],y=u&&l%4===0&&l%3&&c%4===0,w=u?c:f*m,v=u?f*m:c,k=[8,8,1],$=n<=8?[4,1,1]:[4,4,1],C=[Math.ceil(w/k[0]/$[0]),Math.ceil(v/k[1]/$[1]),Math.ceil(p/k[2]/$[2])];rt("verbose",()=>`[conv_backprop_mm_webgpu] dispatch = ${C}`);let T=y?4:1,A=Math.max(k[0]*T,k[1]),B=y?4:1,R=[e.kernelShape[u?1:2],e.kernelShape[u?2:3]],D=[R[0]+(e.dilations[0]<=1?0:(R[0]-1)*(e.dilations[0]-1)),R[1]+(e.dilations[1]<=1?0:(R[1]-1)*(e.dilations[1]-1))],K=[D[0]-1-Math.floor((e.pads[0]+e.pads[2])/2),D[1]-1-Math.floor((e.pads[1]+e.pads[3])/2)],j=[{type:6,data:n},{type:6,data:a},{type:6,data:i},{type:6,data:e.strides},{type:6,data:e.dilations},{type:6,data:R},{type:6,data:K}];en(e,j),j.push(...ye(t[0].dims,t[1].dims));let ie=["rank","rank"];s&&(j.push(...ye(t[2].dims)),ie.push("rank")),j.push(...ye(r));let te=oe=>{let re=Q("x",t[0].dataType,t[0].dims.length,B),M=Q("w",t[1].dataType,t[1].dims.length,1),P=ge("result",t[0].dataType,r.length,B),H=[re,M],le="";if(s){let N=Q("bias",t[2].dataType,t[2].dims.length,B);H.push(N),le+=`
          fn getBiasByOutputCoords(coords : vec4<i32>) -> ${N.type.value} {
            return bias[coords.${u?"w":"y"}${y?"/ 4":""}];
          }`}let G=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"},{name:"strides",type:"i32",length:2},{name:"dilations",type:"i32",length:2},{name:"filter_dims",type:"i32",length:R.length},{name:"pads",type:"i32",length:K.length}];tn(e,G);let ne=pt(t[0].dataType,1);if(ne!=="f16"&&ne!=="f32")throw new Error(`elemType ${ne} is not supported.`);return`
        ${tu("uniforms.result_strides")}
        ${oe.registerUniforms(G).declareVariables(...H,P)};
        ${le}
        ${md(u,s,e,re.type.value,T)}
        ${y?Vi($,k,ne,void 0,!u,A):Gi($,k,ne,void 0,!u,A,!1,void 0,o)}`};return{name:"Conv2DTransposeMatMul",shaderCache:{hint:`${e.cacheKey};${$};${k};${y}`,inputDependencies:ie},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:C[0],y:C[1],z:C[2]},programUniforms:j}),getShaderSource:te}}}),gd,Eo,uy=Z(()=>{$e(),rn(),Ae(),Te(),gd=(t,e,r,n,a,i=!1,s,o,u=!1)=>{let l=u?1:2,p=u?2:3,f=u?3:1,m=i?2:1,c=`
  fn setOutputAtIndex(flatIndex : u32, value : ${i?`vec4<${s}>`:s}) {
    result[flatIndex] = ${i?`vec4<${s}>`:s}(value);
  }`;n&&(c+=`
    fn getBiasByOutputCoords(coords : vec4<u32>) -> ${i?`vec4<${s}>`:s} {
      return bias[coords.${u?"w":"y"}${i?"/ 4":""}];
    }`);let y=i?4:1,w=Q("W",e[1].dataType,e[1].dims.length,y),v=Q("Dy",e[0].dataType,e[0].dims.length,y),k=[v,w];n&&k.push(Q("bias",e[2].dataType,[r[f]].length,y));let $=ge("result",e[0].dataType,r.length,y),C=`{
        let batch: u32 = ${a?"global_id.z":"workgroup_id.z"} / uniforms.result_shape[1];
        let r = ${a?"global_id.z":"workgroup_id.z"} % uniforms.result_shape[1];
        let c = ${a?"global_id.y":"workgroup_id.y"} * ${m};
        let d1: u32 = ${a?"global_id.x":"workgroup_id.x"} * 4;

        let dyCorner = vec2<i32>(i32(r), i32(c)) - vec2<i32>(uniforms.pads);

        // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
        // ? = to be determined. : = across all values in that axis.
        var dotProd: array<vec4<${s}>, ${m}>;
        for (var i = 0; i < ${m}; i++) {
          dotProd[i] = vec4<${s}>(0.0);
        }
        for (var wR: u32 = 0; wR < uniforms.filter_dims[0]; wR = wR + 1) {
          var dyR = (${s}(dyCorner.x) + ${s}(wR)) / ${s}(uniforms.strides.x);
          let wRPerm = uniforms.filter_dims[0] - 1 - wR;
          if (dyR < 0.0 || dyR >= ${s}(uniforms.Dy_shape[1]) ||
              fract(dyR) > 0.0 || wRPerm < 0) {
            continue;
          }
          let idyR: u32 = u32(dyR);

          for (var wC: u32 = 0; wC < uniforms.filter_dims[1]; wC = wC + 1) {
            let dyC = (${s}(dyCorner.y) + ${s}(wC)) / ${s}(uniforms.strides.y);
            let dyC2 = (${s}(dyCorner.y) + 1.0 + ${s}(wC)) / ${s}(uniforms.strides.y);
            let wCPerm = uniforms.filter_dims[1] - 1 - wC;
            if (wCPerm < 0) {
              continue;
            }
            var bDyCVal = true;
            var bDyCVal2 = true;
            if (dyC < 0.0 || dyC >= ${s}(uniforms.Dy_shape[2]) ||
                fract(dyC) > 0.0) {
              bDyCVal = false;
            }
            if (dyC2 < 0.0 || dyC2 >= ${s}(uniforms.Dy_shape[2]) ||
                fract(dyC2) > 0.0) {
              bDyCVal2 = false;
            }

            let idyC: u32 = u32(dyC);
            let idyC2: u32 = u32(dyC2);
            if (bDyCVal && bDyCVal2) {
              let d2Length = uniforms.Dy_shape[3];
              for (var d2 :u32 = 0; d2 < d2Length; d2 = d2 + 4) {
                let wValue0 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};
                let wValue1 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};
                let wValue2 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};
                let wValue3 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};

                var xValue = ${v.get("batch","idyR","idyC","d2")};
                let tmpval = vec4<${s}>(dot(xValue, wValue0),
                                      dot(xValue, wValue1),
                                      dot(xValue, wValue2),
                                      dot(xValue, wValue3));
                dotProd[0] = dotProd[0] + tmpval;

                xValue =  ${v.get("batch","idyR","idyC2","d2")};

                dotProd[1] = dotProd[1] + vec4<${s}>(dot(xValue, wValue0),
                                                    dot(xValue, wValue1),
                                                    dot(xValue, wValue2),
                                                    dot(xValue, wValue3));
              }
            } else if (bDyCVal) {
              let d2Length = uniforms.Dy_shape[${f}];
              for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) {
                let wValue0 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};
                let wValue1 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};
                let wValue2 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};
                let wValue3 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};

                var xValue = ${v.get("batch","idyR","idyC","d2")};
                let tmpval = vec4<${s}>(dot(xValue, wValue0),
                                      dot(xValue, wValue1),
                                      dot(xValue, wValue2),
                                      dot(xValue, wValue3));
                dotProd[0] = dotProd[0] + tmpval;
              }
            } else if (bDyCVal2) {
              let d2Length = uniforms.Dy_shape[3];
              for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) {
                let wValue0 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};
                let wValue1 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};
                let wValue2 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};
                let wValue3 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};

                var xValue = ${v.get("batch","idyR","idyC2","d2")};
                let tmpval = vec4<${s}>(dot(xValue, wValue0),
                                      dot(xValue, wValue1),
                                      dot(xValue, wValue2),
                                      dot(xValue, wValue3));
                dotProd[1] = dotProd[1] + tmpval;
              }
            }
          }
        }

        for (var i: u32 = 0; i < ${m}; i = i + 1) {
          let value = dotProd[i] + ${n?"bias[c+i]":`vec4<${s}>(0.0)`};
          ${$.set("batch","r","c + i","d1","value")};
        }
      }`,T=`
          let outputIndices = ${$.offsetToIndices("global_idx")};
          let batch = ${$.indicesGet("outputIndices",0)};
          let d1 = ${$.indicesGet("outputIndices",f)};
          let r = ${$.indicesGet("outputIndices",l)};
          let c = ${$.indicesGet("outputIndices",p)};
          let dyCorner = vec2<i32>(i32(r), i32(c)) - uniforms.pads;
          let dyRCorner = dyCorner.x;
          let dyCCorner = dyCorner.y;
          let groupId = d1 / uniforms.output_channels_per_group;
          let wOutChannel = d1 - groupId * uniforms.output_channels_per_group;
          // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
          // ? = to be determined. : = across all values in that axis.
          var dotProd = ${s}(0.0);
          for (var wR: u32 = 0; wR < uniforms.effective_filter_dims.x; wR = wR + 1) {
            if (wR % uniforms.dilations.x != 0) {
              continue;
            }
            let dyR = (${s}(dyRCorner) + ${s}(wR)) / ${s}(uniforms.strides[0]);
            let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x;
            if (dyR < 0.0 || dyR >= ${s}(uniforms.Dy_shape[${l}]) || fract(dyR) > 0.0 ||
                wRPerm < 0) {
              continue;
            }
            let idyR: u32 = u32(dyR);

            for (var wC: u32 = 0; wC < uniforms.effective_filter_dims.y; wC = wC + 1) {
              if (wC % uniforms.dilations.y != 0) {
                continue;
              }
              let dyC = (${s}(dyCCorner) + ${s}(wC)) / ${s}(uniforms.strides.y);
              let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y;
              if (dyC < 0.0 || dyC >= ${s}(uniforms.Dy_shape[${p}]) ||
                  fract(dyC) > 0.0 || wCPerm < 0) {
                continue;
              }
              let idyC: u32 = u32(dyC);
              var inputChannel = groupId * uniforms.input_channels_per_group;
              for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group; d2 = d2 + 1) {
                let xValue = ${u?v.get("batch","idyR","idyC","inputChannel"):v.get("batch","inputChannel","idyR","idyC")};
                let wValue = ${w.get("inputChannel","wOutChannel","u32(wRPerm)","u32(wCPerm)")};
                dotProd = dotProd + xValue * wValue;
                inputChannel = inputChannel + 1;
              }
            }
          }
          let value = dotProd + ${n?"bias[d1]":`${s}(0.0)`};
          ${$.setByOffset("global_idx","value")};
        `;return`
  ${t.registerUniforms(o).declareVariables(...k,$)}
  ${c}

    ${t.mainStart()}
    ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")};
  ${i?C:T}}`},Eo=(t,e,r)=>{let n=t.length>2,a=e.outputShape,i=Y.size(a),s=[Math.ceil(i/64),1,1];rt("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${s}`);let o=e.format==="NHWC",u=["rank","rank"],l=[e.strides[0],e.strides[1]],p=[e.kernelShape[o?1:2],e.kernelShape[o?2:3]],f=[e.dilations[0],e.dilations[1]],m=[p[0]+(e.dilations[0]<=1?0:(e.kernelShape[o?1:2]-1)*(e.dilations[0]-1)),p[1]+(e.dilations[1]<=1?0:(e.kernelShape[o?2:3]-1)*(e.dilations[1]-1))],c=[m[0]-1-Math.floor((e.pads[0]+e.pads[2])/2),m[1]-1-Math.floor(e.pads[1]+e.pads[3])/2],y=!1,w=e.group,v=t[1].dims,k=v[0]/w,$=v[1],C=[{type:12,data:i},{type:12,data:l},{type:12,data:p},{type:12,data:f},{type:12,data:m},{type:6,data:c},{type:12,data:k},{type:12,data:$},...ye(t[0].dims,t[1].dims)];n&&(C.push(...ye(t[2].dims)),u.push("rank")),C.push(...ye(a));let T=s[1]===1&&s[2]===1,A=B=>{let R=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:l.length},{name:"filter_dims",type:"u32",length:p.length},{name:"dilations",type:"u32",length:p.length},{name:"effective_filter_dims",type:"u32",length:m.length},{name:"pads",type:"i32",length:c.length},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],D=pt(t[0].dataType);return`${gd(B,t,a,n,T,y,D,R,o)}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${e.cacheKey};`,inputDependencies:u},getRunData:()=>({dispatchGroup:{x:s[0],y:s[1],z:s[2]},outputs:[{dims:r?r(a):a,dataType:t[0].dataType}],programUniforms:C}),getShaderSource:A}}}),_d,yd,wd,Ds,Gf,bd,vd,$d,xd,Hf,ly=Z(()=>{oy(),uy(),nn(),$a(),_d=(t,e,r,n,a,i)=>(t-1)*e+r+(n-1)*a+1-i,yd=(t,e,r,n,a)=>{let i=Math.floor(t/2);e==="SAME_UPPER"?(r[n]=i,r[a]=t-i):e==="SAME_LOWER"&&(r[n]=t-i,r[a]=i)},wd=(t,e,r,n,a,i,s,o,u,l)=>{let p=t.length-2,f=l.length===0;if(u.length===0)for(let y=0;y<p;++y)u.push(0);let m=t[0],c=e[o?3:1]*a;for(let y=0,w=t.length-p-(o?1:0);y<p;++y,++w){let v=t[w],k=f?v*s[y]:l[y],$=_d(v,s[y],i[y],e[w],r[y],k);yd($,n,i,y,y+p),f&&l.push(s[y]*(v-1)+u[y]+(e[w]-1)*r[y]+1-i[y]-i[y+p])}l.splice(0,0,m),l.splice(o?3:1,0,c)},Ds=(t,e)=>{let r=t.kernelShape.slice();if(t.kernelShape.length===0||t.kernelShape.reduce((f,m)=>f*m,1)===0){r.length=0;for(let f=2;f<e[1].dims.length;++f)r.push(e[1].dims[f])}let n=t.format==="NHWC";r.splice(0,0,e[1].dims[0]),r.splice(n?3:1,0,e[1].dims[1]);let a=t.pads.slice(),i=t.outputShape.slice(),s=t.outputPadding.slice(),o=e[0].dims,u=t.dilations.slice();if(u.reduce((f,m)=>f+m,0)===0){let f=e[0].dims.length-2;u=new Array(f).fill(1)}let l=t.strides.slice();if(l.reduce((f,m)=>f+m,0)===0){let f=e[0].dims.length-2;l=new Array(f).fill(1)}wd(o,r,u,t.autoPad,t.group,a,l,n,s,i);let p=Object.assign({},t);return Object.assign(p,{kernelShape:r,pads:a,outputPadding:s,outputShape:i,dilations:u,strides:l}),p},Gf=t=>{let e=Jo(t),r=t.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof t.autoPad>"u"?0:t.autoPad],a=t.dilations,i=t.group,s=t.kernelShape,o=t.pads,u=t.strides,l=t.wIsConst(),p=t.outputPadding,f=t.outputShape;return{autoPad:n,format:r,dilations:a,group:i,kernelShape:s,outputPadding:p,outputShape:f,pads:o,strides:u,wIsConst:l,...e,cacheKey:`${t.format};${e.activation};`}},bd=(t,e)=>{if(!t||t.length!==2&&t.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(t[0].dims.length!==4&&t[0].dims.length!==3)throw new Error("currently only support 2-dimensional conv");if(t[0].dims.length!==t[1].dims.length)throw new Error("filter does not have same dimension as input");let r=t[0].dims[e.format==="NHWC"?t[0].dims.length-1:1],n=t[1].dims[0];if(r!==n)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let a=t[1].dims[1]*e.group;if(t.length===3&&(t[2].dims.length!==1||t[2].dims[0]!==a))throw new Error("invalid bias");let i=t[0].dims.length-2;if(e.dilations.reduce((s,o)=>s+o,0)>0&&e.dilations.length!==i)throw new Error(`dilations should be ${i}D`);if(e.strides.reduce((s,o)=>s+o,0)>0&&e.strides.length!==i)throw new Error(`strides should be ${i}D`);if(e.pads.reduce((s,o)=>s+o,0)>0&&e.pads.length!==i*2)throw new Error(`pads should be ${i*2}D`);if(e.outputPadding.length!==i&&e.outputPadding.length!==0)throw new Error(`output_padding should be ${i}D`);if(e.kernelShape.reduce((s,o)=>s+o,0)>0&&e.kernelShape.length!==0&&e.kernelShape.length!==t[1].dims.length-2)throw new Error("invalid kernel shape");if(e.outputShape.length!==0&&e.outputShape.length!==t[0].dims.length-2)throw new Error("invalid output shape")},vd=[2,3,1,0],$d=(t,e,r)=>{let n=Ds(r,e),a=r.format==="NHWC",i=n.outputShape,s=i[a?3:1],o=e[0].dims[a?3:1];if(n.group!==1||s===1&&o===1){t.compute(Eo(e,n));return}let u=i[a?1:2],l=i[a?2:3],p=e[1].dims[2],f=e[1].dims[3],m=a?u*l:s,c=a?s:u*l,y=p*f*o,w=!0,v=t.kernelCustomData.wT??t.compute(br(e[1],vd),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!t.kernelCustomData.wT&&(t.kernelCustomData.wT=v);let k=[e[0],v],$=e.length===3;$&&(!a&&e[2].dims.length===1?k.push(e[2].reshape([e[2].dims[0],1,1])):k.push(e[2])),t.compute(Vf(k,n,i,m,c,y,$,w),{inputs:k})},xd=(t,e)=>{let r=e.format==="NHWC",n=[t.inputs[0].reshape(r?[t.inputs[0].dims[0],1,t.inputs[0].dims[1],t.inputs[0].dims[2]]:[t.inputs[0].dims[0],t.inputs[0].dims[1],1,t.inputs[0].dims[2]]),t.inputs[1].reshape([t.inputs[1].dims[0],t.inputs[1].dims[1],1,t.inputs[1].dims[2]])];t.inputs.length===3&&n.push(t.inputs[2]);let a=e.kernelShape;(a.length===0||a[0]===0)&&(a=[t.inputs[1].dims[2]]);let i=e.dilations;(i.length===0||i[0]===0)&&(i=[1]);let s=e.strides;(s.length===0||s[0]===0)&&(s=[1]);let o=e.pads;o.length===0&&(o=[0,0]),o=[0,o[0],0,o[1]],s=[1].concat(s),i=[1].concat(i),a=[1].concat(a);let u=Ds({...e,pads:o,strides:s,dilations:i,kernelShape:a},n);t.compute(Eo(n,u,l=>r?[l[0],l[2],l[3]]:[l[0],l[1],l[3]]))},Hf=(t,e)=>{bd(t.inputs,e),t.inputs[0].dims.length===3?xd(t,e):$d(t,t.inputs,e)}}),Sd,jf,qf,dy=Z(()=>{$e(),Ae(),lt(),Te(),Sd=(t,e,r,n)=>{let a=Y.size(e),i=e.length,s=Q("input",t,i),o=ge("output",t,i),u=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),l=Y.normalizeAxis(u,i),p=f=>{let m=` i32(${s.indicesGet("inputIndices","uniforms.axis")}) `,c=Se("uniforms.input_shape","uniforms.axis",i),y=n.reverse?m+(n.exclusive?" + 1":""):"0",w=n.reverse?c:m+(n.exclusive?"":" + 1");return`
                ${f.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(s,o)}
                ${f.mainStart()}
                  ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
                  var inputIndices = ${o.offsetToIndices("global_idx")};
                  var sum = ${o.type.value}(0);
                  let first : i32 = ${y};
                  let last : i32 = ${w};
                  for (var i : i32 = first; i < last; i++) {
                    ${s.indicesSet("inputIndices","uniforms.axis","u32(i)")};
                    sum = sum + ${s.getByIndices("inputIndices")};
                  }
                  ${o.setByOffset("global_idx","sum")};
                }`};return{name:"CumSum",shaderCache:{hint:n.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:e,dataType:t}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:[{type:12,data:a},{type:12,data:l},...ye(e,e)]}),getShaderSource:p}},jf=(t,e)=>{let r=t.inputs[0].dims,n=t.inputs[0].dataType,a=t.inputs[1];t.compute(Sd(n,r,a,e),{inputs:[0]})},qf=t=>{let e=t.exclusive===1,r=t.reverse===1;return qe({exclusive:e,reverse:r})}}),kd,Ed,Cd,Kf,Yf,cy=Z(()=>{$e(),Ae(),lt(),Te(),kd=t=>{if(!t||t.length!==1)throw new Error("DepthToSpace requires 1 input.");if(t[0].dims.length!==4)throw new Error("DepthToSpace requires 4D input.")},Ed=(t,e,r,n)=>{let a=[];a.push(`fn perm(i: ${n.type.indices}) -> ${r.type.indices} {
    var a: ${r.type.indices};`);for(let i=0;i<e;++i)a.push(r.indicesSet("a",t[i],`i[${i}]`));return a.push("return a;}"),a.join(`
`)},Cd=(t,e)=>{let r,n,a,i,s,o,u=e.format==="NHWC",l=e.blocksize,p=e.mode==="DCR";u?([r,n,a,i]=t.dims,s=p?[r,n,a,l,l,i/l**2]:[r,n,a,i/l**2,l,l],o=p?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([r,n,a,i]=[t.dims[0],t.dims[2],t.dims[3],t.dims[1]],s=p?[r,l,l,i/l**2,n,a]:[r,i/l**2,l,l,n,a],o=p?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let f=t.reshape(s),m=f.dims.length,c=t.dataType,y=Q("a",c,m),w=ge("output",c,m),v=k=>`
  ${k.registerUniform("output_size","u32").declareVariables(y,w)}

  ${Ed(o,m,y,w)}

  ${k.mainStart()}
    ${k.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}

    let indices = ${w.offsetToIndices("global_idx")};
    let aIndices = perm(indices);

    ${w.setByOffset("global_idx",y.getByIndices("aIndices"))}
  }`;return{name:"DepthToSpace",shaderCache:{hint:`${t.dims};${e.blocksize};${e.mode}`,inputDependencies:["rank"]},getRunData:k=>{let $=u?[r,n*l,a*l,i/l**2]:[r,i/l**2,n*l,a*l],C=Y.size($),T=f.dims,A=Y.sortBasedOnPerm(T,o);return{outputs:[{dims:$,dataType:k[0].dataType}],dispatchGroup:{x:Math.ceil(C/64)},programUniforms:[{type:12,data:C},...ye(T,A)]}},getShaderSource:v}},Kf=(t,e)=>{kd(t.inputs),t.compute(Cd(t.inputs[0],e))},Yf=t=>qe({blocksize:t.blocksize,mode:t.mode,format:t.format})}),gi,Zn,Ns,Td,Id,Ad,Md,Fs,Od,Xf,Qf,py=Z(()=>{$e(),Ae(),lt(),Te(),gi="[a-zA-Z]|\\.\\.\\.",Zn="("+gi+")+",Ns="^"+Zn+"$",Td="("+Zn+",)*"+Zn,Id="^"+Td+"$",Ad=class{constructor(t=-1){this.symbolToIndices=new Map,this.inputIndex=t}addSymbol(t,e){let r=this.symbolToIndices.get(t);r===void 0?r=[e]:r.push(e),this.symbolToIndices.set(t,r)}},Md=class{constructor(t,e){this.equation=e,this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[r,n]=e.includes("->")?e.split("->",2):[e,""];if(!r.match(RegExp(Id)))throw new Error("Invalid LHS term");if(r.split(",").forEach((a,i)=>{let s=t[i].dims.slice();if(!a.match(RegExp(Ns)))throw new Error("Invalid LHS term");let o=this.processTerm(a,!0,s,i);this.lhs.push(o)}),n==="")n+=[...this.symbolToInfo.entries()].filter(([a,i])=>i.count===1||a==="...").map(([a])=>a).join("");else if(!n.match(RegExp(Zn)))throw new Error("Invalid RHS");n.match(RegExp(gi,"g"))?.forEach(a=>{if(a==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let i=this.symbolToInfo.get(a);if(i===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(i.dimValue)}}),this.rhs=this.processTerm(n,!1,this.outputDims)}addSymbol(t,e,r){let n=this.symbolToInfo.get(t);if(n!==void 0){if(n.dimValue!==e&&n.count!==1)throw new Error("Dimension mismatch");n.count++,n.inputIndices.push(r)}else n={count:1,dimValue:e,inputIndices:[r]};this.symbolToInfo.set(t,n)}processTerm(t,e,r,n=-1){let a=r.length,i=!1,s=[],o=0;if(!t.match(RegExp(Ns))&&!e&&t!=="")throw new Error("Invalid LHS term");let u=t.match(RegExp(gi,"g")),l=new Ad(n);return u?.forEach((p,f)=>{if(p==="..."){if(i)throw new Error("Only one ellipsis is allowed per input term");i=!0;let m=a-u.length+1;if(m<0)throw new Error("Ellipsis out of bounds");if(s=r.slice(o,o+m),this.hasEllipsis){if(this.ellipsisDims.length!==s.length||this.ellipsisDims.toString()!==s.toString())throw new Error("Ellipsis dimensions mismatch")}else if(e)this.hasEllipsis=!0,this.ellipsisDims=s;else throw new Error("Ellipsis must be specified in the LHS");for(let c=0;c<s.length;c++){let y=String.fromCharCode(48+c);l.addSymbol(y,f+c),this.addSymbol(y,r[o++],n)}}else l.addSymbol(p,f+(this.hasEllipsis?this.ellipsisDims.length-1:0)),this.addSymbol(p,r[o++],n)}),l}},Fs=t=>t+"_max",Od=(t,e,r,n)=>{let a=t.map(l=>l.length).map((l,p)=>Q(`input${p}`,e,l)),i=Y.size(n),s=ge("output",e,n.length),o=[...r.symbolToInfo.keys()].filter(l=>!r.rhs.symbolToIndices.has(l)),u=l=>{let p=[],f="var prod = 1.0;",m="var sum = 0.0;",c="sum += prod;",y=[],w=[],v=[],k=[],$=r.symbolToInfo.size===r.rhs.symbolToIndices.size;r.symbolToInfo.forEach((T,A)=>{if(r.rhs.symbolToIndices.has(A)){let B=r.rhs.symbolToIndices.get(A)?.[0];B!==void 0&&r.lhs.forEach((R,D)=>{if(T.inputIndices.includes(D)){let K=R.symbolToIndices.get(A);if(K===void 0)throw new Error("Invalid symbol error");K.forEach(j=>{p.push(`${a[D].indicesSet(`input${D}Indices`,j,s.indicesGet("outputIndices",B))}`)})}})}else r.lhs.forEach((B,R)=>{if(T.inputIndices.includes(R)){let D=B.symbolToIndices.get(A);if(D===void 0)throw new Error("Invalid symbol error");D.forEach(K=>{y.push(`${a[R].indicesSet(`input${R}Indices`,K,`${A}`)}`)}),k.push(`prod *= ${a[R].getByIndices(`input${R}Indices`)};`)}}),w.push(`for(var ${A}: u32 = 0; ${A} < uniforms.${Fs(A)}; ${A}++) {`),v.push("}")});let C=$?[...p,`let sum = ${a.map((T,A)=>T.getByIndices(`input${A}Indices`)).join(" * ")};`]:[...p,m,...w,...y,f,...k,c,...v];return`
            ${l.registerUniforms(o.map(T=>({name:`${Fs(T)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...a,s)}

            ${l.mainStart()}
            ${l.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
            var outputIndices = ${s.offsetToIndices("global_idx")};
            ${a.map((T,A)=>`var input${A}Indices: ${a[A].type.indices};`).join(`
`)}
            ${C.join(`
`)};
            ${s.setByOffset("global_idx","sum")};
          }`};return{name:"Einsum",shaderCache:{hint:r.equation,inputDependencies:t.map(()=>"rank")},getRunData:()=>{let l=o.filter(f=>r.symbolToInfo.has(f)).map(f=>({type:12,data:r.symbolToInfo.get(f)?.dimValue||0}));l.push({type:12,data:i});let p=t.map((f,m)=>[...ye(f)]).reduce((f,m)=>f.concat(m),l);return p.push(...ye(n)),{outputs:[{dims:n,dataType:e}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:p}},getShaderSource:u}},Xf=(t,e)=>{let r=new Md(t.inputs,e.equation),n=r.outputDims,a=t.inputs.map((i,s)=>i.dims);t.compute(Od(a,t.inputs[0].dataType,r,n))},Qf=t=>{let e=t.equation.replace(/\s+/g,"");return qe({equation:e})}}),zd,Ls,Pd,Rd,Jf,hy=Z(()=>{$e(),Ae(),Te(),zd=t=>{if(!t||t.length!==2)throw new Error("Expand requires 2 input.");let e=t[0].dims,r=Array.from(t[1].getBigInt64Array(),Number),n=r.length<e.length?0:r.length-e.length,a=e.length<r.length?0:e.length-r.length;for(;n<r.length&&a<e.length;++n,++a)if(r[n]!==e[a]&&r[n]!==1&&e[a]!==1)throw new Error("Expand requires shape to be broadcastable to input")},Ls=(t,e)=>{let r=t.length-e.length,n=[];for(let a=0;a<r;++a)n.push(t[a]);for(let a=0;a<e.length;++a)n.push(e[a]===1?t[a+r]:e[a]);return n},Pd=(t,e)=>t.length>e.length?Ls(t,e):Ls(e,t),Rd=t=>{let e=t[0].dims,r=Array.from(t[1].getBigInt64Array(),Number),n=Pd(e,r),a=t[0].dataType,i=a===9?4:1,s=Math.ceil(Y.size(n)/i),o=l=>{let p=Q("input",a,e.length,i),f=ge("output",a,n.length,i),m;if(a===9){let c=(y,w,v="")=>`
          let outputIndices${w} = ${f.offsetToIndices(`outputOffset + ${w}u`)};
          let offset${w} = ${p.broadcastedIndicesToOffset(`outputIndices${w}`,f)};
          let index${w} = offset${w} / 4u;
          let component${w} = offset${w} % 4u;
          ${y}[${w}] = ${v}(${p.getByOffset(`index${w}`)}[component${w}]);
        `;m=`
        let outputOffset = global_idx * ${i};
        var data = vec4<u32>(0);
        ${c("data",0,"u32")}
        ${c("data",1,"u32")}
        ${c("data",2,"u32")}
        ${c("data",3,"u32")}
        ${f.setByOffset("global_idx","data")}
      }`}else m=`
        let outputIndices = ${f.offsetToIndices("global_idx")};
        let inputOffset = ${p.broadcastedIndicesToOffset("outputIndices",f)};
        ${f.setByOffset("global_idx",p.getByOffset("inputOffset"))}
      }`;return`
    ${l.registerUniform("vec_size","u32").declareVariables(p,f)}
    ${l.mainStart()}
    ${l.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
    ${m}`},u=[{type:12,data:s},...ye(e,n)];return{name:"Expand",shaderCache:{hint:`${n.length}`,inputDependencies:["rank"]},getShaderSource:o,getRunData:()=>({outputs:[{dims:n,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:u})}},Jf=t=>{zd(t.inputs),t.compute(Rd(t.inputs),{inputs:[0]})}}),Bd,Zf,fy=Z(()=>{$e(),Ae(),Te(),Qo(),Bd=t=>{let e=t[0].dataType,r=Y.size(t[0].dims),n=Y.size(t[1].dims),a=n%4===0,i=s=>{let o=Q("x",e,[1],4),u=Q("bias",e,[1],4),l=ge("y",e,[1],4),p=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],f=c=>`
      let bias${c}_offset: u32 = (global_idx * 4 + ${c}) % uniforms.bias_size;
      let bias${c} = ${u.getByOffset(`bias${c}_offset / 4`)}[bias${c}_offset % 4];`,m=a?`
      let bias = ${u.getByOffset("global_idx % (uniforms.bias_size / 4)")};`:`${f(0)}${f(1)}${f(2)}${f(3)}
      let bias = ${o.type.value}(bias0, bias1, bias2, bias3);`;return`${s.registerUniforms(p).declareVariables(o,u,l)}

    ${bo(Tt(e))}

    ${s.mainStart(En)}
      ${s.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_vec_size")}

      let x = ${o.getByOffset("global_idx")};
      ${m}
      let x_in = x + bias;
      ${l.setByOffset("global_idx",vo("x_in"))}
    }`};return{name:"FastGeluWithBias",shaderCache:{hint:`${a}`,inputDependencies:["type","type"]},getShaderSource:i,getRunData:s=>({outputs:[{dims:s[0].dims,dataType:s[0].dataType}],programUniforms:[{type:12,data:Math.ceil(r/4)},{type:12,data:n}],dispatchGroup:{x:Math.ceil(r/En/4)}})}},Zf=t=>{t.inputs.length<2||Y.size(t.inputs[1].dims)===0?xf(t):t.compute(Bd(t.inputs))}}),Dd,Nd,em,tm,my=Z(()=>{$e(),Ae(),lt(),Te(),Dd=t=>{if(!t||t.length!==2)throw new Error("Gather requires 2 inputs.")},Nd=(t,e)=>{let r=t[0].dims,n=t[1].dims,a=r.length,i=Y.normalizeAxis(e.axis,a),s=r.slice(0);s.splice(i,1,...n);let o=r[i],u=t[0].dataType===9?4:1,l=Math.ceil(Y.size(s)/u),p=[{type:12,data:l},{type:6,data:o},{type:12,data:i},...ye(t[0].dims,t[1].dims,s)],f=m=>{let c=Q("data",t[0].dataType,t[0].dims.length,u),y=Q("inputIndices",t[1].dataType,t[1].dims.length),w=ge("output",t[0].dataType,s.length,u),v=$=>{let C=n.length,T=`var indicesIndices${$}  = ${y.type.indices}(0);`;for(let A=0;A<C;A++)T+=`${C>1?`indicesIndices${$}[${A}]`:`indicesIndices${$}`} = ${s.length>1?`outputIndices${$}[uniforms.axis + ${A}]`:`outputIndices${$}`};`;T+=`
          var idx${$} = ${y.getByIndices(`indicesIndices${$}`)};
          if (idx${$} < 0) {
            idx${$} = idx${$} + uniforms.axisDimLimit;
          }
          var dataIndices${$} : ${c.type.indices};
        `;for(let A=0,B=0;A<a;A++)A===i?(T+=`${a>1?`dataIndices${$}[${A}]`:`dataIndices${$}`} = u32(idx${$});`,B+=C):(T+=`${a>1?`dataIndices${$}[${A}]`:`dataIndices${$}`} = ${s.length>1?`outputIndices${$}[${B}]`:`outputIndices${$}`};`,B++);return T},k;if(t[0].dataType===9){let $=(C,T,A="")=>`
          let outputIndices${T} = ${w.offsetToIndices(`outputOffset + ${T}u`)};
          ${v(T)};
          let offset${T} = ${c.indicesToOffset(`dataIndices${T}`)};
          let index${T} = offset${T} / 4u;
          let component${T} = offset${T} % 4u;
          ${C}[${T}] = ${A}(${c.getByOffset(`index${T}`)}[component${T}]);
        `;k=`
        let outputOffset = global_idx * ${u};
        var value = vec4<u32>(0);
        ${$("value",0,"u32")}
        ${$("value",1,"u32")}
        ${$("value",2,"u32")}
        ${$("value",3,"u32")}
        ${w.setByOffset("global_idx","value")}
      `}else k=`
      let outputIndices = ${w.offsetToIndices("global_idx")};
      ${v("")};
      let value = ${c.getByIndices("dataIndices")};
      ${w.setByOffset("global_idx","value")};
      `;return`
      ${m.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(c,y,w)}
      ${m.mainStart()}
        ${m.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
        ${k}
      }`};return{name:"Gather",shaderCache:{hint:e.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:s,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:p}),getShaderSource:f}},em=t=>qe({axis:t.axis}),tm=(t,e)=>{let r=t.inputs;Dd(r),t.compute(Nd(t.inputs,e))}}),Fd,Ld,rm,nm,gy=Z(()=>{$e(),Ae(),lt(),Te(),Fd=t=>{if(!t||t.length!==2)throw new Error("GatherElements requires 2 inputs.");if(t[0].dims.length<1)throw new Error("GatherElements requires that the data input be rank >= 1.");if(t[0].dims.length!==t[1].dims.length)throw new Error(`GatherElements requires that the data input and
                     indices input tensors be of same rank.`)},Ld=(t,e)=>{let r=t[0].dims,n=t[0].dataType,a=r.length,i=t[1].dims,s=t[1].dataType,o=Y.normalizeAxis(e.axis,a),u=r[o],l=i.slice(0),p=Y.size(l),f=Q("input",n,a),m=Q("indicesInput",s,i.length),c=ge("output",n,l.length),y=[{type:12,data:p},{type:6,data:u},{type:12,data:o}];return y.push(...ye(r,i,l)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:l,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:y}),getShaderSource:w=>`
      ${w.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(f,m,c)}
      ${w.mainStart()}
      ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}

      let outputIndices = ${c.offsetToIndices("global_idx")};

      var idx = ${m.getByOffset("global_idx")};
      if (idx < 0) {
        idx = idx + uniforms.axisDimLimit;
      }
      var inputIndices = ${f.type.indices}(outputIndices);
      ${f.indicesSet("inputIndices","uniforms.axis","u32(idx)")};
      let value = ${f.getByIndices("inputIndices")};

      ${c.setByOffset("global_idx","value")};
  }`}},rm=t=>qe({axis:t.axis}),nm=(t,e)=>{let r=t.inputs;Fd(r),t.compute(Ld(t.inputs,e))}}),Ud,Wd,am,im,_y=Z(()=>{$e(),Ae(),Te(),Ud=t=>{if(!t)throw new Error("Input is missing");if(t.length<2||t.length>3)throw new Error("Invaid input number.");if(t.length===3&&t[2].dims.length>2)throw new Error("Invalid input shape of C");if(t[0].dataType!==t[1].dataType||t.length===3&&t[0].dataType!==t[2].dataType)throw new Error("Input types are mismatched")},Wd=(t,e)=>{let r=t[0].dims.slice(),n=t[1].dims.slice(),[a,i,s]=ph.getShapeOfGemmResult(r,e.transA,n,e.transB,t.length===3?t[2].dims:void 0),o=[a,i];if(!o)throw new Error("Can't use gemm on the given tensors");let u=Y.size(o),l=[{type:12,data:u},{type:12,data:a},{type:12,data:i},{type:12,data:s},{type:1,data:e.alpha},{type:1,data:e.beta}],p=["type","type"];t.length===3&&(l.push(...ye(t[2].dims)),p.push("rank")),l.push(...ye(o));let f=m=>{let c="";e.transA&&e.transB?c="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":e.transA&&!e.transB?c="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!e.transA&&e.transB?c="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!e.transA&&!e.transB&&(c="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let y=e.alpha===1?"":"value *= uniforms.alpha;",w=Q("a",t[0].dataType,t[0].dims),v=Q("b",t[1].dataType,t[1].dims),k=w.type.value,$=null,C=[w,v];t.length===3&&($=Q("c",t[2].dataType,t[2].dims.length),C.push($));let T=ge("output",t[0].dataType,o.length);C.push(T);let A=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}];return`
  ${m.registerUniforms(A).declareVariables(...C)}

  ${m.mainStart()}
    ${m.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}

    let m = global_idx / uniforms.N;
    let n = global_idx % uniforms.N;

    var value = ${k}(0);
    for (var k: u32 = 0u; k < uniforms.K; k++) {
      ${c}
    }

    ${y}
    ${$!=null?`let cOffset = ${$.broadcastedIndicesToOffset("vec2(m, n)",T)}; value += ${k}(uniforms.beta) * ${$.getByOffset("cOffset")};`:""}
    output[global_idx] = value;
  }`};return{name:"Gemm",shaderCache:{hint:`${e.cacheKey}`,inputDependencies:p},getRunData:()=>({outputs:[{dims:o,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:l}),getShaderSource:f}},am=t=>{let e=t.transA,r=t.transB,n=t.alpha,a=t.beta;return{transA:e,transB:r,alpha:n,beta:a,cacheKey:`${t.transA};${t.transB};${t.alpha===1}`}},im=(t,e)=>{Ud(t.inputs),t.compute(Wd(t.inputs,e))}}),Vd,Gd,Hd,sm,yy=Z(()=>{$e(),Ae(),Te(),Vd=(t,e)=>{let r=t[0].dims,n=r,a=2,i=Y.sizeToDimension(r,a),s=Y.sizeFromDimension(r,a),o=it(s),u=s/o,l=[r[0],r[1],u],p=["rank","type","type"],f=[{type:12,data:s},{type:12,data:u}];f.push(...ye(l,l));let m=c=>{let y=Q("x",t[0].dataType,l.length,o),w=Q("scale",t[1].dataType,t[1].dims),v=Q("bias",t[2].dataType,t[2].dims),k=ge("output",t[0].dataType,l.length,o),$=[y,w,v,k],C=y.type.value,T=o===1?"f32":`vec${o}<f32>`,A=64,B=[{name:"normSize",type:"u32"},{name:"normPackedSize",type:"u32"}];return`
  var<workgroup> meanShared : f32;
  var<workgroup> squaredNormShared : f32;
  var<workgroup> workgroupShared : array<${T}, ${A}>;
  const workgroupSize = ${A}u;
  ${c.registerUniforms(B).declareVariables(...$)}
  ${c.mainStart(A)}
    let norm = global_idx / workgroupSize;
    let batch = norm / uniforms.x_shape[1];
    let channel = norm % uniforms.x_shape[1];
    let localIndex = local_id.x;

    // initialize workgroup memory
    var initial = ${T}(0);
    for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) {
      initial = initial + ${T}(${y.get("batch","channel","h")});
    }
    workgroupShared[localIndex] = initial;
    workgroupBarrier();

    // Calculate the mean of current channel data.
    for (var currSize = workgroupSize >> 1;  currSize > 0; currSize = currSize >> 1) {
      if (localIndex < currSize) {
        workgroupShared[localIndex] = workgroupShared[localIndex] + workgroupShared[localIndex + currSize];
      }
      workgroupBarrier();
    }
    if (localIndex == 0) {
      meanShared = ${Mr("workgroupShared[0]",o)} / f32(uniforms.normSize);
    }
    workgroupBarrier();

    // reinitialize workgroup memory.
    initial = ${T}(0);
    for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) {
      let deviation =  ${T}(${y.get("batch","channel","h")}) - ${T}(meanShared);
      initial = initial + deviation * deviation;
    }
    workgroupShared[localIndex] = initial;
    workgroupBarrier();

    // Calculate the sum of square of deviation of current channel data.
    for (var currSize = workgroupSize >> 1;  currSize > 0; currSize = currSize >> 1) {
      if (localIndex < currSize) {
        workgroupShared[localIndex] = workgroupShared[localIndex] + workgroupShared[localIndex + currSize];
      }
      workgroupBarrier();
    }
    if (localIndex == 0) {
      squaredNormShared = ${Mr("workgroupShared[0]",o)};
    }
    workgroupBarrier();

    let invStdDev = inverseSqrt(squaredNormShared / f32(uniforms.normSize) + f32(${e.epsilon}));
    let channelScale = invStdDev * f32(${w.getByOffset("channel")});
    let channelShift = f32(${v.getByOffset("channel")}) - meanShared * channelScale;
    for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) {
      let value = ${y.get("batch","channel","h")} * ${C}(${T}(channelScale)) + ${C}(${T}(channelShift));
      ${k.set("batch","channel","h","value")};
    }
  }`};return{name:"InstanceNormalization",shaderCache:{hint:`${e.epsilon};${o}`,inputDependencies:p},getRunData:()=>({outputs:[{dims:n,dataType:t[0].dataType}],dispatchGroup:{x:i},programUniforms:f}),getShaderSource:m}},Gd=(t,e,r,n,a,i,s,o)=>{let u=it(s),l=64,p=u===1?"vec2f":`mat2x${u}f`,f=u===1?"f32":`vec${u}f`,m=(B,R)=>`${p}(${B}, ${R})`,c=a*s/u,y=Math.ceil(i/l),w=["type"],v=[{type:12,data:y},{type:12,data:i},{type:12,data:Math.floor(s/u)},{type:12,data:Math.floor(i*s/u)}],k=B=>{let R=Q("input",e.dataType,e.dims,u);return`
  ${B.declareVariables(R)}
  @group(0) @binding(1) var<storage, read_write> output : array<${p}>;
  struct Uniforms {wg_size:u32, H:u32, C:u32, image_size:u32};
  @group(0) @binding(2) var<uniform> uniforms: Uniforms;

  ${B.mainStart(l)}
    let currentImageNumber = global_idx / ${l} / uniforms.C;
    let currentChannelNumber = (global_idx / ${l}) % uniforms.C;
    let wgOffset = local_id.x * uniforms.wg_size;
    if (wgOffset >= uniforms.H) {
        return;
    }
    let wgMax = min(wgOffset + uniforms.wg_size, uniforms.H);

    let offset = currentImageNumber * uniforms.image_size + currentChannelNumber;
    var sum = ${yr("f32",u)};
    var squaredSum = ${yr("f32",u)};
    for (var i: u32 = wgOffset; i < wgMax; i++) {
        let value = ${f}(input[offset + i * uniforms.C]);
        sum += value;
        squaredSum += value * value;
    }
    output[global_idx] = ${m("sum","squaredSum")};
  }`},$=t.compute({name:"InstanceNormComputeMean",shaderCache:{hint:`${u}`,inputDependencies:w},getRunData:()=>({outputs:[{dims:[a,s,l,2],dataType:1}],dispatchGroup:{x:a*s/u},programUniforms:v}),getShaderSource:k},{inputs:[e],outputs:[-1]})[0],C=[{type:12,data:c},{type:12,data:i},{type:12,data:Math.floor(s/u)},{type:12,data:Math.floor(l*s/u)}],T=["type","type","type"],A=B=>{let R=Q("scale",r.dataType,r.dims,u),D=Q("bias",n.dataType,n.dims,u);return`
  @group(0) @binding(0) var<storage, read> input : array<${p}>;
  @group(0) @binding(1) var<storage, read> scale : array<${R.type.storage}>;
  @group(0) @binding(2) var<storage, read> bias : array<${D.type.storage}>;
  @group(0) @binding(3) var<storage, read_write> output : array<${p}>;
  struct Uniforms {units_of_work : u32, H: u32, C : u32, image_size : u32};
  @group(0) @binding(4) var<uniform> uniforms: Uniforms;

  ${B.mainStart()}
    ${B.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.units_of_work")}
    let currentImageNumber = global_idx / uniforms.C;
    let currentChannelNumber = global_idx % uniforms.C;

    let offset = currentImageNumber * uniforms.image_size;
    var sum = ${yr("f32",u)};
    var squaredSum = ${yr("f32",u)};
    for (var i: u32 = 0; i < min(${l}, uniforms.H); i++) {
        let value = input[offset + i + currentChannelNumber * ${l}];
        sum += value[0];
        squaredSum += value[1];
    }
    sum = sum / f32(uniforms.H);
    squaredSum = squaredSum / f32(uniforms.H);
    let invStdDev = inverseSqrt(squaredSum - sum * sum + f32(${o}));
    let channelScale = invStdDev * ${f}(scale[currentChannelNumber]);
    let channelShift = ${f}(bias[currentChannelNumber]) - sum * channelScale;

    output[global_idx] = ${m("channelScale","channelShift")};
  }`};return t.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${u};${o}`,inputDependencies:T},getRunData:()=>({outputs:[{dims:[a,s,2],dataType:1}],dispatchGroup:{x:Math.ceil(c/64)},programUniforms:C}),getShaderSource:A},{inputs:[$,r,n],outputs:[-1]})[0]},Hd=(t,e,r)=>{let n=e[0].dims,a=n,i=n[0],s=n[n.length-1],o=Y.sizeFromDimension(n,1)/s,u=it(s),l=Y.size(a)/u,p=[{type:12,data:o},{type:12,data:Math.floor(s/u)}],f=["type","type"],m=Gd(t,e[0],e[1],e[2],i,o,s,r.epsilon),c=y=>{let w=pt(e[0].dataType),v=u===1?"vec2f":`mat2x${u}f`,k=u===1?w:`vec${u}<${w}>`,$=Q("input",e[0].dataType,e[0].dims,u),C=ge("output",e[0].dataType,a,u);return`
  @group(0) @binding(0) var<storage, read> input : array<${$.type.storage}>;
  @group(0) @binding(1) var<storage, read> scaleInput : array<${v}>;
  @group(0) @binding(2) var<storage, read_write> output : array<${C.type.storage}>;
  struct Uniforms {H: u32, C : u32};
  @group(0) @binding(3) var<uniform> uniforms: Uniforms;

  ${y.mainStart()}
    let currentImageNumber = global_idx / (uniforms.C * uniforms.H);
    let currentChannelNumber = global_idx % uniforms.C;

    let scaleOffset = currentImageNumber * uniforms.C + currentChannelNumber;
    let scale = scaleInput[scaleOffset];
    output[global_idx] = fma(input[global_idx], ${k}(scale[0]), ${k}(scale[1]));
  }`};t.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${u}`,inputDependencies:f},getRunData:()=>({outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:p}),getShaderSource:c},{inputs:[e[0],m]})},sm=(t,e)=>{e.format==="NHWC"?Hd(t,t.inputs,e):t.compute(Vd(t.inputs,e))}}),jd,qd,om,wy=Z(()=>{$e(),Ae(),Te(),jd=t=>{if(!t||t.length<2)throw new Error("layerNorm requires at least 2 inputs.")},qd=(t,e,r)=>{let n=e.simplified,a=t[0].dims,i=t[1],s=!n&&t[2],o=a,u=Y.normalizeAxis(e.axis,a.length),l=Y.sizeToDimension(a,u),p=Y.sizeFromDimension(a,u),f=Y.size(i.dims),m=s?Y.size(s.dims):0;if(f!==p||s&&m!==p)throw new Error(`Size of X.shape()[axis:] == ${p}.
       Size of scale and bias (if provided) must match this.
       Got scale size of ${f} and bias size of ${m}`);let c=[];for(let A=0;A<a.length;++A)A<u?c.push(a[A]):c.push(1);let y=it(p),w=["type","type"],v=[{type:12,data:l},{type:1,data:p},{type:12,data:Math.floor(p/y)},{type:1,data:e.epsilon}];s&&w.push("type");let k=r>1,$=r>2,C=A=>{let B=pt(t[0].dataType),R=[Q("x",t[0].dataType,t[0].dims,y),Q("scale",i.dataType,i.dims,y)];s&&R.push(Q("bias",s.dataType,s.dims,y)),R.push(ge("output",t[0].dataType,o,y)),k&&R.push(ge("mean_data_output",1,c)),$&&R.push(ge("inv_std_output",1,c));let D=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return`
  ${A.registerUniforms(D).declareVariables(...R)}
  ${A.mainStart()}
    ${A.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")}
    let offset = global_idx * uniforms.norm_size_vectorized;
    var mean_vector = ${yr("f32",y)};
    var mean_square_vector = ${yr("f32",y)};

    for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) {
      let value = ${xn(B,y,"x[h + offset]")};
      mean_vector += value;
      mean_square_vector += value * value;
    }
    let mean = ${Mr("mean_vector",y)} / uniforms.norm_size;
    let inv_std_dev = inverseSqrt(${Mr("mean_square_vector",y)} / uniforms.norm_size ${n?"":"- mean * mean"} + uniforms.epsilon);

    for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) {
      let f32input = ${xn(B,y,"x[j + offset]")};
      let f32scale = ${xn(B,y,"scale[j]")};
      output[j + offset] = ${R[0].type.value}((f32input ${n?"":"- mean"}) * inv_std_dev * f32scale
        ${s?`+ ${xn(B,y,"bias[j]")}`:""}
      );
    }

    ${k?"mean_data_output[global_idx] = mean":""};
    ${$?"inv_std_output[global_idx] = inv_std_dev":""};
  }`},T=[{dims:o,dataType:t[0].dataType}];return k&&T.push({dims:c,dataType:1}),$&&T.push({dims:c,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${y};${r};${n}`,inputDependencies:w},getRunData:()=>({outputs:T,dispatchGroup:{x:Math.ceil(l/64)},programUniforms:v}),getShaderSource:C}},om=(t,e)=>{jd(t.inputs),t.compute(qd(t.inputs,e,t.outputCount))}}),Kd,Yd,um,lm,by=Z(()=>{$e(),Ae(),lt(),Te(),Kd=(t,e)=>{if(t.length<3||t.length>4)throw new Error("MatMulNBits requires 3 or 4 inputs");let r=t[0],n=r.dims.length;if(r.dims[n-1]!==e.k)throw new Error("The last dim of input shape does not match the k value");let a=Math.floor((e.k+e.blockSize-1)/e.blockSize),i=e.blockSize/8*e.bits,s=t[1];if(!Y.areEqual(s.dims,[e.n,a,i]))throw new Error("The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize");let o=t[2].dims;if(Y.size(o)!==e.n*a)throw new Error("scales input size error.");if(t.length===4){let u=t[3].dims,l=e.bits>4?e.n*a:e.n*Math.floor((a+1)/2);if(Y.size(u)!==l)throw new Error("zeroPoints input size error.")}},Yd=(t,e,r,n)=>{let a=t[0].dims,i=a.length,s=Math.floor((e.k+e.blockSize-1)/e.blockSize),o=a[i-2],u=e.k,l=e.n,p=a.slice(0,i-2),f=Y.size(p),m=e.blockSize/8*e.bits/4,c=t[0].dataType,y=it(o),w=it(e.k),v=it(m),k=fa(c),$=o*s*k,C=Math.floor(n/$),T=s<=r[0]&&C>0,A=!T||C>=4?it(l):C>=2&&it(l)>=2?2:1,B=p.concat([o,l]),R=Y.size(B)/A/y,D=T?[]:[{type:12,data:R},{type:12,data:e.blockSize}],K=[f,o,u/w],j=Y.convertShape(t[1].dims).slice();j.splice(-1,1,m/v),D.push(...ye(K)),D.push(...ye(j)),D.push(...ye(t[2].dims)),t.length===4&&D.push(...ye(Y.convertShape(t[3].dims)));let ie=[f,o,l/A];D.push(...ye(ie));let te=oe=>{let re=K.length,M=Q("a",t[0].dataType,re,w),P=Q("b",12,j.length,v),H=Q("scales",t[2].dataType,t[2].dims.length),le=[M,P,H],G=t.length===4?Q("zero_points",12,t[3].dims.length):void 0;G&&le.push(G);let ne=ie.length,N=ge("output",t[0].dataType,ne,A),ae=[{name:"output_size",type:"u32"},{name:"block_size",type:"u32"}],fe=pt(t[0].dataType),Ce=(()=>{switch(w){case 1:return`array<${fe}, 8>`;case 2:return`mat4x2<${fe}>`;case 4:return`mat2x4<${fe}>`;default:throw new Error(`${w}-component is not supported.`)}})(),Be=`
        for (var word: u32 = 0; word < ${m}; word += ${v}) {
          ${P.indicesSet("b_indices","2","word")};
          let b_data = ${P.getByIndices("b_indices")};
          for (var i: u32 = 0; i < ${v}; i++) {
            let b_value: u32 = ${v===1?"b_data":"b_data[word + i]"};
            let b_mask: u32 = 0x0F0F0F0Fu;
            let b_value_lower: vec4<u32> = unpack4xU8(b_value & b_mask);
            let b_value_upper: vec4<u32> = unpack4xU8((b_value >> 4) & b_mask);
            let b_quantized_values = ${Ce}(${Array.from({length:4},(Qe,We)=>`${fe}(b_value_lower[${We}]), ${fe}(b_value_upper[${We}])`).join(", ")});
            let b_dequantized_values = ${w===1?`${Ce}(${Array.from({length:8},(Qe,We)=>`(b_quantized_values[${We}] - zero_point) * scale`).join(", ")});`:`(b_quantized_values - ${Ce}(${Array(8).fill("zero_point").join(",")})) * scale;`};
            // Number of B elements per 32-bit word is 32/bits = 32/4 = 8
            for (var m: u32 = 0; m < ${T?o:y}u; m++) {
              ${M.indicesSet("a_indices",re-2,T?"m":`row * ${y} + m`)};
              ${M.indicesSet("a_indices",re-1,"word_offset")};
              var input_offset = ${M.indicesToOffset("a_indices")};
              var a_data: ${Ce};
              for (var j: u32 = 0; j < ${8/w}; j++) {
                a_data[j] = ${M.getByOffset("input_offset")};
                input_offset++;
              }
              ${T?"workgroup_shared[workgroup_shared_offset + m]":"output_values[m]"}${A>1?"[c]":""} += ${Array.from({length:8/w},(Qe,We)=>`${w===1?`a_data[${We}] * b_dequantized_values[${We}]`:`dot(a_data[${We}], b_dequantized_values[${We}])`}`).join(" + ")};
            }
            word_offset += ${8/w};
          }
        }`,Ke=G?`
          zero_point_offset += 4;
          if (zero_point_offset == 32) {
            zero_point_offset = 0;
            zero_point_index++;
            zero_point_word = ${G.getByOffset("zero_point_index")};
          }`:"";return T?`
        var<workgroup> workgroup_shared: array<${N.type.value}, ${o*s}>;
        ${oe.declareVariables(...le,N)}
        ${oe.mainStart([s,1,1])}
          var a_indices: ${M.type.indices};
          var block = local_id.x;
          var col = workgroup_id.y;
          var batch = workgroup_id.z;
          ${M.indicesSet("a_indices","0","batch")};
          // Two zero points are packed into one byte when uniforms.bits is 4.
          for (var c: u32 = 0; c < ${A}; c++) {
            let col_times_components_plus_c = col * ${A} + c;
              ${G?`
            var zero_point_bytes_per_col: u32 = (${s} + 1) / 2;
            var zero_point_byte_count: u32 = col_times_components_plus_c * zero_point_bytes_per_col + (block >> 0x1u);
            var zero_point_word_index: u32 = zero_point_byte_count >> 0x2u;
            var zero_point_byte_offset: u32 = zero_point_byte_count & 0x3u;
            var zero_point_nibble_offset: u32 = block & 0x1u;
            var zero_point_bits_offset: u32 = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2);
            var zero_point_word: u32 = ${G.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;`:""}
            var b_indices: ${P.type.indices};
            ${P.indicesSet("b_indices","0","col_times_components_plus_c")};
            // The scale and zero points are computed per block.
            var scales_index = col_times_components_plus_c * ${s} + block;
            let scale = ${H.getByOffset("scales_index")};
            // The default zero point is 8 for unsigned 4-bit quantization.
            let zero_point = ${fe}(${G?"(zero_point_word) & 0xFu":8});
            ${P.indicesSet("b_indices","1","block")};
            var word_offset: u32 = block * ${e.blockSize/w};
            var workgroup_shared_offset: u32 = block * ${o};
            ${Be}
          }
          workgroupBarrier();
          if (local_id.x == 0u) {
            var output_indices: ${N.type.indices};
            ${N.indicesSet("output_indices","0","batch")};
            ${N.indicesSet("output_indices",ne-1,"col")};
            ${N.indicesSet("output_indices",ne-2,"0")};
            var output_offset = ${N.indicesToOffset("output_indices")};
            for (var m: u32 = 0u; m < ${o}u; m++) {
              var output_value: ${N.type.value} = ${N.type.value}(0);
              var workgroup_shared_offset: u32 = m;
              for (var b: u32 = 0u; b < ${s}u; b++) {
                output_value += workgroup_shared[workgroup_shared_offset];
                workgroup_shared_offset += ${o};
              }
              ${N.setByOffset("output_offset","output_value")};
              output_offset += ${l/A};
            }
          }
        }`:`
        ${oe.registerUniforms(ae).declareVariables(...le,N)}
        ${oe.mainStart()}
          ${oe.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
          var output_values: array<${N.type.value}, ${y}>;
          var output_indices = ${N.offsetToIndices("global_idx")};
          var col = ${N.indicesGet("output_indices",ne-1)};
          var row = ${N.indicesGet("output_indices",ne-2)};
          var a_indices: ${M.type.indices} = output_indices;
          // Two zero points are packed into one byte because uniforms.bits <= 4.
          // zero_point_offset is either 0 or 4. It is bit offset within one byte.
          // TODO support zero_point_offset for bits > 4
          ${G?`
          var zero_point_abs_offset = col * ${A} * ((${s} + 1) / 2);
          var zero_point_index: u32 = zero_point_abs_offset / 4;
          var zero_point_word: u32 = ${G.getByOffset("zero_point_index")};
          var zero_point_offset: u32 = (zero_point_abs_offset % 4) * 8;`:""}
          var scale_index = col * ${s*A};
          var b_indices: ${P.type.indices};
          for (var c: u32 = 0; c < ${A}; c++) {
            ${P.indicesSet("b_indices","0",`col * ${A} + c`)};
            var block_offset: u32 = 0;
            for (var block: u32 = 0; block < ${s}; block++) {
              // The scale and zero points are computed per block.
              let scale = ${H.getByOffset("scale_index")};
              // The default zero point is 8 for unsigned 4-bit quantization.
              let zero_point = ${fe}(${G?"extractBits(zero_point_word, zero_point_offset, 4)":8});
              ${P.indicesSet("b_indices","1","block")};
              var word_offset: u32 = block_offset;
              ${Be}
              scale_index++;
              ${Ke}
              block_offset += uniforms.block_size / ${w};
            }
            // Drop the trailing 4 bits if the zero_poit_offset is not a byte boundary to align with the next byte.
            ${G?`if (zero_point_offset % 8 > 0) {
                ${Ke}
              }`:""}
            }
            for (var k: u32 = 0u; k < ${y}u; k++) {
              ${N.indicesSet("output_indices",ne-2,`${y} * row + k`)};
              ${N.setByIndices("output_indices","output_values[k]")}
            }
        }`};return{name:T?"BlockwiseMatMulNBits":"MatMulNBits",shaderCache:{hint:`${e.cacheKey};${o};${c};${t.length}`,inputDependencies:Array(t.length).fill("rank")},getRunData:()=>({outputs:[{dims:B,dataType:c}],name:T?"BlockwiseMatMulNBits":"MatMulNBits",dispatchGroup:T?{x:1,y:Math.ceil(l/A),z:f}:{x:Math.ceil(R/64)},programUniforms:D}),getShaderSource:te}},um=(t,e)=>{Kd(t.inputs,e);let r=t.getMaxComputeWorkgroupSizes(),n=t.getMaxComputeWorkgroupStoragesize();t.compute(Yd(t.inputs,e,r,n))},lm=t=>qe(t)}),yt,Xd,dm,Us,Qd,_i,cm,vy=Z(()=>{$e(),Ae(),lt(),jo(),Wh(),Te(),$a(),yt=(t,e)=>t.length>e&&t[e].dims.length>0&&Y.size(t[e].dims)>0?t[e]:void 0,Xd=(t,e)=>{let r=t[0],n=yt(t,1),a=yt(t,2),i=yt(t,3),s=yt(t,4),o=yt(t,5),u=yt(t,6),l=yt(t,7);if(r.dims.length!==3&&r.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let p=!1,f=r.dims[0],m=r.dims[1],c=r.dims.length===3?p?r.dims[2]/3:r.dims[2]:e.numHeads*r.dims[4],y=m,w=0,v=0,k=Math.floor(c/e.numHeads);if(u&&l){if(u.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(u.dims[0]!==f||u.dims[1]!==e.numHeads||u.dims[3]!==k)throw new Error('Input "past_key" shape (batch_size, num_heads, past_sequence_length, head_size)');if(l.dims[0]!==f||l.dims[1]!==e.numHeads||l.dims[3]!==k)throw new Error('Input "past_value" shape (batch_size, num_heads, past_sequence_length, head_size)');if(u.dims[2]!==l.dims[2])throw new Error('Input "past_key" and "past_value" shall have same dim 2 (past_sequence_length)');if(l.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');w=u.dims[2],v=u.dims[2]}else if(u||l)throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let $;if(n){if(r.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(n.dims.length<3||n.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(r.dims[0]!==n.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(n.dims.length===3){if(n.dims[2]!==r.dims[2])throw new Error('Input "query" and "key" shall have same dim 2 (hidden_size)');$=2,y=n.dims[1]}else if(n.dims.length===5){if(n.dims[2]!==e.numHeads||n.dims[3]!==2||n.dims[4]!==k)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(a)throw new Error('Expect "value" be none when "key" has packed kv format.');$=5,y=n.dims[1]}else{if(n.dims[1]!==e.numHeads||n.dims[3]!==k)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');$=0,y=n.dims[2]}}else{if(r.dims.length!==3&&r.dims.length!==5)throw new Error('Input "query" is expected to have 3 or 5 dimensions when key is empty');if(r.dims.length===5&&(r.dims[2]!==e.numHeads||r.dims[3]!==3))throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');$=3}if(i){if(i.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimension');if(a&&r.dims.length===5&&r.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let C=0;if(s){C=8;let D=s.dims;throw D.length===1?D[0]===f?C=1:D[0]===3*f+2&&(C=3):D.length===2&&D[0]===f&&D[1]===y&&(C=5),C===8?new Error('Input "key_padding_mask" shape shall be (batch_size) or (batch_size, kv_sequence_length)'):new Error("Mask not supported")}let T=!1,A=c;if(a){if(a.dims.length!==3&&a.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==a.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(a.dims.length===3){if(y!==a.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');A=a.dims[2]}else{if(y!==a.dims[2])throw new Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');A=a.dims[1]*a.dims[3],T=!0}}let B=w+y,R=!1;if(s)throw new Error("Key padding mask is not supported");if(o){if(o.dims.length!==4)throw new Error('Input "relative_position_bias" is expected to have 4 dimensions');if(o.dims[0]!==f&&o.dims[0]!==1||o.dims[1]!==e.numHeads||o.dims[2]!==m||o.dims[3]!==B)throw new Error('Input "relative_position_bias" shape (batch_size, 1, sequence_length, kv_sequence_length)')}return{batchSize:f,sequenceLength:m,pastSequenceLength:w,kvSequenceLength:y,totalSequenceLength:B,maxSequenceLength:v,inputHiddenSize:0,hiddenSize:c,vHiddenSize:A,headSize:k,vHeadSize:Math.floor(A/e.numHeads),numHeads:e.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:e.maskFilterValue,maskType:C,scale:e.scale,broadcastResPosBias:R,passPastInKv:T,qkvFormat:$}},dm=t=>qe({...t}),Us=qe({perm:[0,2,1,3]}),Qd=(t,e,r,n,a,i,s)=>{let o=[n,a,i],u=Y.size(o),l=[{type:12,data:u},{type:12,data:s},{type:12,data:i}],p=f=>{let m=ge("qkv_with_bias",e.dataType,o),c=Q("qkv",e.dataType,o),y=Q("bias",r.dataType,o),w=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`
  ${f.registerUniforms(w).declareVariables(c,y,m)}
  ${f.mainStart()}
    ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
    let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset;

    qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx];
  }`};return t.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:o,dataType:e.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:l}),getShaderSource:p},{inputs:[e,r],outputs:[-1]})[0]},_i=(t,e,r,n,a,i,s,o)=>{let u=i;if(s){if(n===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return u=Qd(t,i,s,e,n,r*a,o),u=u.reshape([e,n,r,a]),t.compute(br(u,Us.perm),{inputs:[u],outputs:[-1]})[0]}else return i.dims.length===3&&(u=i.reshape([e,n,r,a])),t.compute(br(u,Us.perm),{inputs:[u],outputs:[-1]})[0]},cm=(t,e)=>{let r=Xd(t.inputs,e),n=t.inputs[0],a=yt(t.inputs,1),i=yt(t.inputs,2),s=yt(t.inputs,3),o=yt(t.inputs,4),u=yt(t.inputs,5),l=yt(t.inputs,6),p=yt(t.inputs,7);if(n.dims.length===5)throw new Error("Packed QKV is not implemented");if(a?.dims.length===5)throw new Error("Packed KV is not implemented");let f=a&&i&&a.dims.length===4&&i.dims.length===4,m=_i(t,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,n,s,0);if(f)return Wi(t,m,a,i,o,void 0,l,p,u,r,e);if(!a||!i)throw new Error("key and value must be provided");let c=_i(t,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,a,s,r.hiddenSize),y=_i(t,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,i,s,2*r.hiddenSize);Wi(t,m,c,y,o,void 0,l,p,u,r,e)}}),Jd,Zd,ec,tc,rc,nc,ac,ic,pm,$y=Z(()=>{$e(),Ae(),Te(),Jd=t=>{if(!t||t.length<1)throw new Error("Too few inputs");if(t[0].dataType!==1&&t[0].dataType!==10)throw new Error("Input type must be float or float16.");if(t.length>=2){let e=t[0].dims.length*2===t[1].dims[0];if(t.length===4&&(e=t[3].dims[0]*2===t[1].dims[0]),!e)throw new Error("The pads should be a 1D tensor of shape [2 * input_rank] or [2 * num_axes].")}},Zd=(t,e,r)=>{let n="";for(let a=e-1;a>=0;--a)n+=`
            k = i32(${t.indicesGet("indices",a)}) - ${Se("uniforms.pads",a,r)};
            if (k < 0) {
              break;
            }
            if (k >= i32(${Se("uniforms.x_shape",a,e)})) {
              break;
            }
            offset += k * i32(${Se("uniforms.x_strides",a,e)});
        `;return`
          value = ${t.type.value}(uniforms.constant_value);
          for (var i = 0; i < 1; i++) {
            var offset = 0;
            var k = 0;
            ${n}
            value = x[offset];
          }
      `},ec=(t,e,r)=>{let n="";for(let a=e-1;a>=0;--a)n+=`
                k = i32(${t.indicesGet("indices",a)}) - ${Se("uniforms.pads",a,r)};
                if (k < 0) {
                  k = -k;
                }
                {
                  let _2n_1 = 2 * (i32(${Se("uniforms.x_shape",a,e)}) - 1);
                  k = k % _2n_1;
                  if(k >= i32(${Se("uniforms.x_shape",a,e)})) {
                    k = _2n_1 - k;
                  }
                }
                offset += k * i32(${Se("uniforms.x_strides",a,e)});
            `;return`
              var offset = 0;
              var k = 0;
              ${n}
              value = x[offset];
          `},tc=(t,e,r)=>{let n="";for(let a=e-1;a>=0;--a)n+=`
                k = i32(${t.indicesGet("indices",a)}) - ${Se("uniforms.pads",a,r)};
                if (k < 0) {
                  k = 0;
                }
                if (k >= i32(${Se("uniforms.x_shape",a,e)})) {
                  k = i32(${Se("uniforms.x_shape",a,e)}) - 1;
                }
                offset += k * i32(${Se("uniforms.x_strides",a,e)});
            `;return`
              var offset = 0;
              var k = 0;
              ${n}
              value = x[offset];
          `},rc=(t,e,r)=>{let n="";for(let a=e-1;a>=0;--a)n+=`
                k = i32(${t.indicesGet("indices",a)}) - ${Se("uniforms.pads",a,r)};
                if (k < 0)  {
                  k += i32(${Se("uniforms.x_shape",a,e)}]);
                }
                if (k >= i32(${Se("uniforms.x_shape",a,e)})) {
                  k -= i32(${Se("uniforms.x_shape",a,e)});
                }
                offset += k * i32(${Se("uniforms.x_strides",a,e)});
            `;return`
              var offset = 0;
              var k = 0;
              ${n}
              value = x[offset];
          `},nc=(t,e,r)=>{switch(r.mode){case 0:return Zd(t,e,r.pads.length);case 1:return ec(t,e,r.pads.length);case 2:return tc(t,e,r.pads.length);case 3:return rc(t,e,r.pads.length);default:throw new Error("Invalid mode")}},ac=(t,e)=>{let r=Y.padShape(t[0].dims.slice(),e.pads),n=t[0].dims,a=Y.size(r),i=[{type:12,data:a},{type:6,data:e.pads}];e.mode===0&&i.push({type:t[0].dataType,data:e.value}),i.push(...ye(t[0].dims,r));let s=["rank"],o=u=>{let l=ge("output",t[0].dataType,r.length),p=Q("x",t[0].dataType,n.length),f=p.type.value,m=nc(l,n.length,e),c=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:e.pads.length}];return e.mode===0&&c.push({name:"constant_value",type:f}),`
            ${u.registerUniforms(c).declareVariables(p,l)}
            ${u.mainStart()}
            ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}

            let indices = ${l.offsetToIndices("global_idx")};

            var value = ${f}(0);
            ${m}
            output[global_idx] = value;
        }`};return{name:"Pad",shaderCache:{hint:`${e.mode}`,inputDependencies:s},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(Y.size(r)/64)},programUniforms:i}),getShaderSource:o}},ic=(t,e)=>{if(t.length>1){let r=t[1].getBigInt64Array(),n=t.length>=3&&t[2].data?t[2].getFloat32Array()[0]:0,a=t[0].dims.length,i=new Int32Array(2*a).fill(0);if(t.length>=4){let o=t[3].getBigInt64Array();for(let u=0;u<o.length;u++)i[Number(o[u])]=Number(r[u]),i[Number(o[u])+a]=Number(r[u+o.length])}else r.forEach((o,u)=>i[Number(u)]=Number(o));let s=[];return i.forEach(o=>s.push(o)),{mode:e.mode,value:n,pads:s}}else return e},pm=(t,e)=>{Jd(t.inputs);let r=ic(t.inputs,e);t.compute(ac(t.inputs,r),{inputs:[0]})}}),ea,Ws,Vs,Gs,Hs,sc,oc,js,qs,hm,fm,Ks,mm,gm,Ys,_m,ym,wm,bm,xy=Z(()=>{ar(),$e(),Ae(),Te(),ea=t=>{if(Ue.webgpu.validateInputContent&&(!t||t.length!==1))throw new Error("Pool ops requires 1 input.")},Ws=(t,e,r)=>{let n=e.format==="NHWC",a=t.dims.slice();n&&a.splice(1,0,a.pop());let i=Object.hasOwnProperty.call(e,"dilations"),s=e.kernelShape.slice(),o=e.strides.slice(),u=i?e.dilations.slice():[],l=e.pads.slice();Fi.adjustPoolAttributes(r,a,s,o,u,l);let p=Fi.computePoolOutputShape(r,a,o,u,s,l,e.autoPad),f=Object.assign({},e);i?Object.assign(f,{kernelShape:s,strides:o,pads:l,dilations:u,cacheKey:e.cacheKey}):Object.assign(f,{kernelShape:s,strides:o,pads:l,cacheKey:e.cacheKey});let m=p.slice();return m.push(m.splice(1,1)[0]),[f,n?m:p]},Vs=(t,e)=>{let r=e.format==="NHWC",n=Y.size(t),a=Y.size(e.kernelShape),i=[{type:12,data:n},{type:12,data:a}],s=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(e.kernelShape.length<=2){let o=e.kernelShape[e.kernelShape.length-1],u=e.strides[e.strides.length-1],l=e.pads[e.pads.length/2-1],p=e.pads[e.pads.length-1],f=!!(l+p);i.push({type:12,data:o},{type:12,data:u},{type:12,data:l},{type:12,data:p}),s.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let m=!1;if(e.kernelShape.length===2){let c=e.kernelShape[e.kernelShape.length-2],y=e.strides[e.strides.length-2],w=e.pads[e.pads.length/2-2],v=e.pads[e.pads.length-2];m=!!(w+v),i.push({type:12,data:c},{type:12,data:y},{type:12,data:w},{type:12,data:v}),s.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[i,s,!0,f,m]}else{if(r)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let o=Y.computeStrides(e.kernelShape);i.push({type:12,data:o},{type:12,data:e.pads},{type:12,data:e.strides}),s.push({name:"kernelStrides",type:"u32",length:o.length},{name:"pads",type:"u32",length:e.pads.length},{name:"strides",type:"u32",length:e.strides.length});let u=e.pads.reduce((l,p)=>l+p);return[i,s,!!u,!1,!1]}},Gs=(t,e,r,n,a,i,s,o,u,l,p,f)=>{let m=a.format==="NHWC",c=e.type.value,y=ge("output",e.type.tensor,n);if(a.kernelShape.length<=2){let w="",v="",k="",$=r-(m?2:1);if(p?w=`
                for (var i: u32 = 0u; i < uniforms.kw; i++) {
                  xIndices[${$}] = indices[${$}] * uniforms.sw - uniforms.pwStart + i;
                  if (xIndices[${$}] < 0 || xIndices[${$}]
                      >= uniforms.x_shape[${$}]) {
                    pad++;
                    continue;
                  }
                  let x_val = x[${e.indicesToOffset("xIndices")}];
                  ${i}
                }`:w=`
                for (var i: u32 = 0u; i < uniforms.kw; i++) {
                  xIndices[${$}] = indices[${$}] * uniforms.sw - uniforms.pwStart + i;
                  let x_val = x[${e.indicesToOffset("xIndices")}];
                  ${i}
                }`,a.kernelShape.length===2){let C=r-(m?3:2);f?v=`
                for (var j: u32 = 0u; j < uniforms.kh; j++) {
                  xIndices[${C}] = indices[${C}] * uniforms.sh - uniforms.phStart + j;
                  if (xIndices[${C}] < 0 || xIndices[${C}] >= uniforms.x_shape[${C}]) {
                    pad += i32(uniforms.kw);
                    continue;
                  }
              `:v=`
                for (var j: u32 = 0u; j < uniforms.kh; j++) {
                  xIndices[${C}] = indices[${C}] * uniforms.sh - uniforms.phStart + j;
                `,k=`
              }
            `}return`
            ${t.registerUniforms(u).declareVariables(e,y)}

            ${t.mainStart()}
              ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}

              let indices = ${y.offsetToIndices("global_idx")};
              var xIndices = ${y.offsetToIndices("global_idx")};

              var value = ${c}(${o});
              var pad = 0;
              ${v}
              ${w}
              ${k}
              ${s}

              output[global_idx] = value;
            }`}else{if(m)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let w=a.kernelShape.length,v=a.pads.length,k="";return l?k=`
                if (xIndices[j] >= uniforms.x_shape[j]) {
                  pad++;
                  isPad = true;
                  break;
                }
              }
              if (!isPad) {
                let x_val = x[${e.indicesToOffset("xIndices")}];
                ${i}
              }`:k=`
              }
              let x_val = x[${e.indicesToOffset("xIndices")}];
              ${i}
            `,`
            ${t.registerUniforms(u).declareVariables(e,y)}

            ${t.mainStart()}
              ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
              let indices = ${y.offsetToIndices("global_idx")};
              var xIndices = ${y.offsetToIndices("global_idx")};

              var offsets: array<u32, ${w}>;

              var value = ${c}(${o});
              var pad = 0;
              var isPad = false;

              for (var i: u32 = 0u; i < uniforms.kernelSize; i++) {
                var offset = i;
                for (var j = 0u; j < ${w-1}u; j++) {
                  offsets[j] = offset / ${Se("uniforms.kernelStrides","j",w)};
                  offset -= offsets[j] * ${Se("uniforms.kernelStrides","j",w)};
                }
                offsets[${w-1}] = offset;

                isPad = false;
                for (var j = ${r-w}u; j < ${r}u; j++) {
                  xIndices[j] = indices[j] * ${Se("uniforms.strides",`j - ${r-w}u`,w)}
                    + offsets[j - ${r-w}u] - ${Se("uniforms.pads","j - 2u",v)};
                  ${k}
              }
              ${s}

              output[global_idx] = value;
            }`}},Hs=t=>`${t.format};${t.ceilMode};${t.autoPad};${t.kernelShape.length}`,sc=t=>`${Hs(t)};${t.countIncludePad}`,oc=t=>`${Hs(t)};${t.storageOrder};${t.dilations}`,js=t=>({format:t.format,autoPad:["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][t.auto_pad],ceilMode:t.ceil_mode,kernelShape:t.kernel_shape,strides:t.strides,pads:t.pads}),qs=(t,e,r,n)=>{let[a,i]=Ws(e,n,r),s=Q("x",e.dataType,e.dims.length),o=s.type.value,u="value += x_val;",l="";a.countIncludePad?l+=`value /= ${o}(uniforms.kernelSize);`:l+=`value /= ${o}(i32(uniforms.kernelSize) - pad);`;let[p,f,m,c,y]=Vs(i,a);p.push(...ye(e.dims,i));let w=["rank"];return{name:t,shaderCache:{hint:`${n.cacheKey};${m};${c};${y}`,inputDependencies:w},getRunData:()=>({outputs:[{dims:i,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(Y.size(i)/64)},programUniforms:p}),getShaderSource:v=>Gs(v,s,e.dims.length,i.length,a,u,l,0,f,m,c,y)}},hm=t=>{let e=t.count_include_pad!==0,r=js(t);if(r.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for AveragePool");let n={countIncludePad:e,...r,cacheKey:""};return{...n,cacheKey:sc(n)}},fm=(t,e)=>{ea(t.inputs),t.compute(qs("AveragePool",t.inputs[0],!1,e))},Ks={autoPad:"",ceilMode:0,countIncludePad:!1,kernelShape:[],strides:[],pads:[],storageOrder:0,dilations:[]},mm=t=>{let e=t.format;return{format:e,...Ks,cacheKey:e}},gm=(t,e)=>{ea(t.inputs),t.compute(qs("GlobalAveragePool",t.inputs[0],!0,e))},Ys=(t,e,r,n)=>{let[a,i]=Ws(e,n,r),s=`
      value = max(x_val, value);
    `,o="",u=Q("x",e.dataType,e.dims.length),l=["rank"],[p,f,m,c,y]=Vs(i,a);return p.push(...ye(e.dims,i)),{name:t,shaderCache:{hint:`${n.cacheKey};${m};${c};${y}`,inputDependencies:l},getRunData:()=>({outputs:[{dims:i,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(Y.size(i)/64)},programUniforms:p}),getShaderSource:w=>Gs(w,u,e.dims.length,i.length,a,s,o,e.dataType===10?-65504:-1e5,f,m,c,y)}},_m=(t,e)=>{ea(t.inputs),t.compute(Ys("MaxPool",t.inputs[0],!1,e))},ym=t=>{let e=t.storage_order,r=t.dilations,n=js(t);if(e!==0)throw new Error("column major storage order is not yet supported for MaxPool");if(n.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for MaxPool");let a={storageOrder:e,dilations:r,...n,cacheKey:""};return{...a,cacheKey:oc(a)}},wm=t=>{let e=t.format;return{format:e,...Ks,cacheKey:e}},bm=(t,e)=>{ea(t.inputs),t.compute(Ys("GlobalMaxPool",t.inputs[0],!0,e))}}),uc,lc,vm,Sy=Z(()=>{ar(),$e(),Te(),uc=(t,e,r)=>{let n=t===e,a=t<e&&r<0,i=t>e&&r>0;if(n||a||i)throw new Error("Range these inputs' contents are invalid.")},lc=(t,e,r,n)=>{let a=Math.abs(Math.ceil((e-t)/r)),i=[a],s=a,o=[{type:12,data:s},{type:n,data:t},{type:n,data:r},...ye(i)],u=l=>{let p=ge("output",n,i.length),f=p.type.value,m=[{name:"outputSize",type:"u32"},{name:"start",type:f},{name:"delta",type:f}];return`
        ${l.registerUniforms(m).declareVariables(p)}
        ${l.mainStart()}
        ${l.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
        output[global_idx] = uniforms.start + ${f}(global_idx) * uniforms.delta;
      }`};return{name:"Range",shaderCache:{hint:`${n}`},getShaderSource:u,getRunData:()=>({outputs:[{dims:i,dataType:n}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:o})}},vm=t=>{let e=0,r=0,n=0;t.inputs[0].dataType===6?(e=t.inputs[0].getInt32Array()[0],r=t.inputs[1].getInt32Array()[0],n=t.inputs[2].getInt32Array()[0]):t.inputs[0].dataType===1&&(e=t.inputs[0].getFloat32Array()[0],r=t.inputs[1].getFloat32Array()[0],n=t.inputs[2].getFloat32Array()[0]),Ue.webgpu.validateInputContent&&uc(e,r,n),t.compute(lc(e,r,n,t.inputs[0].dataType),{inputs:[]})}}),dc,cc,pc,hc,fc,mc,gc,_c,yc,wc,bc,Xs,vc,$c,xc,Sc,kc,$m,xm,ky=Z(()=>{$e(),Ae(),lt(),Te(),dc=(t,e)=>{if(t.every(r=>r>0||(()=>{throw new Error("Resize requires scales input values to be positive")})),t.length>0){if(e.mode==="linear"){if(!(t.length===2||t.length===3||t.length===4&&t[0]===1&&t[1]===1||t.length===4&&t[0]===1&&t[3]===1||t.length===5&&t[0]===1&&t[1]===1))throw new Error(`For linear mode, Resize requires scales to be 2D, 3D, 4D with either two outermost or one innermost and
            one outermost scale values equal to 1, or 5D with two outermost scale values equal to 1`)}else if(e.mode==="cubic"&&!(t.length===2||t.length===4&&t[0]===1&&t[1]===1||t.length===4&&t[0]===1&&t[3]===1))throw new Error("Resize requires scales input size to be 2 or 4 for cubic mode")}},cc=(t,e,r)=>{e.every(a=>a>=0&&a<r||(()=>{throw new Error("Resize requires axes input values to be positive and less than rank")}));let n=new Array(r).fill(1);return e.forEach((a,i)=>n[a]=t[i]),n},pc=(t,e,r,n,a,i)=>{let[s,o,u]=r>10?[1,2,3]:[-1,t.length>1?1:-1,-1],l=t[0].dims.length;if(s>0&&t.length>s&&t[s].dims.length>0)t[s].getFloat32Array().forEach(p=>i.push(p));else if(e.coordinateTransformMode==="tf_crop_and_resize")throw new Error("Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize");if(o>0&&t.length>o&&t[o].dims.length>0){if(t[o].getFloat32Array().forEach(p=>n.push(p)),n.length!==0&&n.length!==l&&r>=18&&n.length!==e.axes.length)throw new Error("Resize requires scales input size to be same as input rank or axes size for opset 18 and up");dc(n,e),e.axes.length>0&&cc(n,e.axes,l).forEach((p,f)=>n[f]=p)}if(u>0&&t.length>u&&(t[u].getBigInt64Array().forEach(p=>a.push(Number(p))),a.length!==l||r>=18&&a.length===e.axes.length))throw new Error("Resize requires sizes input size to be same as input rank or axes size for opset 18 and up");if(e.axes.length>0){if(n.length!==e.axes.length)throw new Error('Resize requires "scales" input size to be of axes rank when axes attributes is specified');if(a.length!==e.axes.length)throw new Error('Resize requires "sizes" input size to be of rank axes rank when axes attributes is specified')}if(typeof n<"u"&&typeof a<"u"&&n.length>0&&a.length>l)throw new Error("Resize requires only of scales or sizes to be specified")},hc=(t,e)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32,
     lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${e} { `+(()=>{switch(t){case"asymmetric":return`return ${e}(xResized) / ${e}(xScale);`;case"pytorch_half_pixel":return`if (lengthResized > 1) {
                    return (${e}(xResized) + 0.5) / ${e}(xScale) - 0.5;
                  } else {
                    return 0.0;
                  }`;case"tf_half_pixel_for_nn":return`return (${e}(xResized) + 0.5) / ${e}(xScale);`;case"align_corners":return`if (lengthResized == 1) {
                    return 0.0;
                  } else {
                    // The whole part and the fractional part are calculated separately due to inaccuracy of floating
                    // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an
                    // offset-by-one error later in floor().
                    let whole = ${e}(xResized * (lengthOriginal - 1) / (lengthResized - 1));
                    let fract =
                        ${e}(xResized * (lengthOriginal - 1) % (lengthResized - 1)) / ${e}(lengthResized - 1);
                    return whole + fract;
                  }`;case"tf_crop_and_resize":return`if (lengthResized > 1) {
                    return ${e}(roiStart) * ${e}(lengthOriginal - 1) +
                        (${e}(xResized) * ${e}(roiEnd - roiStart) * ${e}(lengthOriginal - 1)) /
                        ${e}(lengthResized - 1);
                  } else {
                    return 0.5 * ${e}(roiStart + roiEnd) * ${e}(lengthOriginal - 1);
                  }`;case"half_pixel_symmetric":return`const outputWidth = ${e}xScale * ${e}(lengthResized);
                  const adjustment = ${e}(lengthResized) / outputWidth;
                  const center = ${e}(lengthOriginal) / 2;
                  const offset = center * (1 - adjustment);
                  return offset + ((${e}(xResized) + 0.5) / ${e}(xScale)) - 0.5;`;case"half_pixel":return`return ((${e}(xResized) + 0.5) / ${e}(xScale)) - 0.5;`;default:throw new Error(`Coordinate transform mode ${t} is not supported`)}})()+"}",fc=(t,e,r)=>`fn getNearestPixelFromOriginal(xOriginal: ${r}, isDownSample: bool) -> ${r} {`+(()=>{switch(t){case"round_prefer_ceil":return"if (fract(xOriginal) == 0.5) {             return ceil(xOriginal);           } else {             return round(xOriginal);           }";case"floor":return"return floor(xOriginal);";case"ceil":return"return ceil(xOriginal);";case"round_prefer_floor":return"if (fract(xOriginal) == 0.5) {                     return floor(xOriginal);                   } else {                     return round(xOriginal);                   }";case"simple":default:if(e<11)return"if (isDownSample)                     {                       return ceil(xOriginal);                     } else {                       return xOriginal;                     }";throw new Error(`Nearest mode ${t} is not supported`)}})()+"}",mc=(t,e,r)=>{let n=new Array(r).fill(0).concat(new Array(r).fill(1)),a=t.length===0?n:t.slice();return e.length>0?(e.forEach((i,s)=>{n[i]=a[s],n[s+r]=a[e.length+s]}),n):a},gc=(t,e,r,n)=>{let a=[];if(r.length>0)if(n.length>0){if(t.forEach(i=>a.push(i)),Math.max(...n)>t.length)throw new Error("axes is out of bound");n.forEach((i,s)=>a[i]=r[s])}else r.forEach(i=>a.push(i));else{if(e.length===0)throw new Error("Resize requires either scales or sizes.");a=t.map((i,s)=>Math.round(i*e[s]))}return a},_c=(t,e,r)=>{let n=(()=>{switch(r.keepAspectRatioPolicy){case"not_larger":return r.axes.length>0?Math.min(...r.axes.map(i=>e[i]),Number.MAX_VALUE):Math.min(...e,Number.MAX_VALUE);case"not_smaller":return r.axes.length>0?Math.max(...r.axes.map(i=>e[i]),Number.MIN_VALUE):Math.max(...e,Number.MIN_VALUE);default:throw new Error(`Keep aspect ratio policy ${r.keepAspectRatioPolicy} is not supported`)}})();e.fill(1,0,e.length);let a=t.slice();return r.axes.length>0?(r.axes.forEach(i=>e[i]=n),r.axes.forEach(i=>a[i]=Math.round(t[i]*e[i]))):(e.fill(n,0,e.length),a.forEach((i,s)=>a[s]=Math.round(i*e[s]))),a},yc=(t,e,r,n,a)=>`
    fn calculateOriginalIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> array<${t.type.value}, ${r.length}> {
      var original_indices: array<${t.type.value}, ${r.length}>;
      for (var i:u32 = 0; i < ${r.length}; i++) {
        var output_index = ${t.indicesGet("output_indices","i")};
        var scale = ${Se("uniforms.scales","i",n)};
        var roi_low = ${Se("uniforms.roi","i",a)};
        var roi_hi = ${Se("uniforms.roi",`i + ${e.length}`,a)};
        if (scale == 1.0) {
          original_indices[i] = ${t.type.value}(output_index);
        } else {
          var input_shape_i = ${Se("uniforms.input_shape","i",e.length)};
          var output_shape_i = ${Se("uniforms.output_shape","i",r.length)};
          original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
                                                                           input_shape_i, roi_low, roi_hi);
        }
      }
      return original_indices;
    }`,wc=(t,e,r,n,a,i,s)=>`
    fn calculateInputIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> ${t.type.indices} {
      var input_indices: ${t.type.indices};
      for (var i:u32 = 0; i < ${n.length}; i++) {
        var output_index = ${e.indicesGet("output_indices","i")};
        var input_index: u32;
        var scale = ${Se("uniforms.scales","i",a)};
        if (scale == 1.0) {
          input_index = output_index;
        } else {
          var roi_low = ${Se("uniforms.roi","i",i)};
          var roi_hi = ${Se("uniforms.roi",`i + ${r.length}`,i)};
          var input_shape_i = ${Se("uniforms.input_shape","i",r.length)};
          var output_shape_i = ${Se("uniforms.output_shape","i",n.length)};
          var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
                                                                        input_shape_i, roi_low, roi_hi);
          if (!${s} || (original_idx >= 0 && original_idx < ${e.type.value}(input_shape_i))) {
            if (original_idx < 0) {
              input_index = 0;
            } else if (original_idx > ${e.type.value}(input_shape_i - 1)) {
              input_index = input_shape_i - 1;
            } else {
              input_index = u32(getNearestPixelFromOriginal(original_idx, scale < 1));
            }
          } else {
            input_index = u32(original_idx);
          }
        }
        ${t.indicesSet("input_indices","i"," input_index")}
      }
      return input_indices;
    }`,bc=(t,e)=>`
    fn checkInputIndices(input_indices: ${t.type.indices}) -> bool {
      for (var i:u32 = 0; i < ${e.length}; i++) {
        var input_index = ${t.indicesGet("input_indices","i")};
        if (input_index < 0 || input_index >= ${Se("uniforms.input_shape","i",e.length)}) {
          return false;
        }
      }
      return true;
    }`,Xs=(t,e,r,n)=>t.rank>n?`
    ${t.indicesSet("input_indices",e,"channel")};
    ${t.indicesSet("input_indices",r,"batch")};
`:"",vc=(t,e,r,n,a)=>{let[i,s,o,u]=r.length===2?[-1,0,1,-1]:[0,2,3,1],l=t.type.value;return`
    fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${l} {
      var input_indices: ${t.type.indices};
      ${t.indicesSet("input_indices",s,`max(0, min(row, ${r[s]} - 1))`)};
      ${t.indicesSet("input_indices",o,`max(0, min(col, ${r[o]} - 1))`)};
      ${Xs(t,u,i,2)}
      return ${t.getByIndices("input_indices")};
    }

    fn bilinearInterpolation(output_indices: ${e.type.indices}) -> ${l} {
      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
      var row:${l} = originalIndices[${s}];
      var col:${l} = originalIndices[${o}];
      ${n?`if (row < 0 || row > (${r[s]} - 1) || col < 0 || col > (${r[o]} - 1)) {
        return ${a};
      }`:""};
      row = max(0, min(row, ${r[s]} - 1));
      col = max(0, min(col, ${r[o]} - 1));
      var row1: u32 = u32(row);
      var col1: u32 = u32(col);
      var row2: u32 = u32(row + 1);
      var col2: u32 = u32(col + 1);
      var channel: u32 = ${r.length>2?`u32(originalIndices[${u}])`:"0"};
      var batch: u32 =  ${r.length>2?`u32(originalIndices[${i}])`:"0"};
      var x11: ${l} = getInputValue(batch, channel, row1, col1);
      var x12: ${l} = getInputValue(batch, channel, row1, col2);
      var x21: ${l} = getInputValue(batch, channel, row2, col1);
      var x22: ${l} = getInputValue(batch, channel, row2, col2);
      var dx1: ${l} = abs(row - ${l}(row1));
      var dx2: ${l} = abs(${l}(row2) - row);
      var dy1: ${l} = abs(col - ${l}(col1));
      var dy2: ${l} = abs(${l}(col2) - col);
      if (row1 == row2) {
        dx1 = 0.5;
        dx2 = 0.5;
      }
      if (col1 == col2) {
        dy1 = 0.5;
        dy2 = 0.5;
      }
      return (x11 * dx2 * dy2 + x12 * dx2 * dy1 + x21 * dx1 * dy2 + x22 * dx1 * dy1);
    }`},$c=(t,e,r,n,a,i,s,o,u,l)=>{let p=r.length===2,[f,m]=p?[0,1]:[2,3],c=t.type.value,y=w=>{let v=w===f?"row":"col";return`
      fn ${v}CubicInterpolation(input_indices: ${t.type.indices}, output_indices: ${e.type.indices}) -> ${c} {
        var output_index = ${e.indicesGet("output_indices",w)};
        var originalIdx: ${c} = getOriginalCoordinateFromResizedCoordinate(output_index, ${a[w]},
        ${n[w]}, ${r[w]}, ${i[w]}, ${i[w]} + ${r.length});
        var fractOriginalIdx: ${c} = originalIdx - floor(originalIdx);
        var coefs = getCubicInterpolationCoefs(fractOriginalIdx);

        if (${o} && (originalIdx < 0 || originalIdx > (${r[w]} - 1))) {
          return ${u};
        }
        var data: array<${c}, 4> = array<${c}, 4>(0.0, 0.0, 0.0, 0.0);
        for (var i: i32 = -1; i < 3; i++) {
          var ${v}: ${c} = originalIdx + ${c}(i);
          if (${v} < 0 || ${v} >= ${r[w]}) {
            ${l?`coefs[i + 1] = 0.0;
                        continue;`:o?`return ${u};`:`${v} = max(0, min(${v}, ${r[w]} - 1));`};
          }
        var input_indices_copy: ${t.type.indices} = input_indices;
          ${t.indicesSet("input_indices_copy",w,`u32(${v})`)};
          data[i + 1] = ${w===f?t.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"};
        }
        return cubicInterpolation1D(data, coefs);
      }`};return`
    ${y(f)};
    ${y(m)};
  fn getCubicInterpolationCoefs(s: ${c}) -> array<${c}, 4> {
    var absS = abs(s);
    var coeffs: array<${c}, 4> = array<${c}, 4>(0.0, 0.0, 0.0, 0.0);
    var oneMinusAbsS: ${c} = 1.0 - absS;
    var twoMinusAbsS: ${c} = 2.0 - absS;
    var onePlusAbsS: ${c} = 1.0 + absS;
    coeffs[0] = ((${s} * onePlusAbsS - 5 * ${s}) * onePlusAbsS + 8 * ${s}) * onePlusAbsS - 4 * ${s};
    coeffs[1] = ((${s} + 2) * absS - (${s} + 3)) * absS * absS + 1;
    coeffs[2] = ((${s} + 2) * oneMinusAbsS - (${s} + 3)) * oneMinusAbsS * oneMinusAbsS + 1;
    coeffs[3] = ((${s} * twoMinusAbsS - 5 * ${s}) * twoMinusAbsS + 8 * ${s}) * twoMinusAbsS - 4 * ${s};
    return coeffs;
  }

  fn cubicInterpolation1D(x: array<${c}, 4>, coefs: array<${c}, 4>) -> ${c} {
    var coefsSum: ${c} = coefs[0] + coefs[1] + coefs[2] + coefs[3];
    return (x[0] * coefs[0] + x[1] * coefs[1]+ x[2] * coefs[2]+ x[3] * coefs[3]) / coefsSum;
  }

  fn bicubicInterpolation(output_indices: ${e.type.indices}) -> ${c} {
    var input_indices: ${t.type.indices} = output_indices;
    return colCubicInterpolation(input_indices, output_indices);
  }
    `},xc=(t,e,r,n,a)=>{let[i,s,o,u,l]=r.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],p=t.type.value;return`
    fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${p} {
      var input_indices: ${t.type.indices};
      ${t.indicesSet("input_indices",s,`max(0, min(depth, ${r[s]} - 1))`)};
      ${t.indicesSet("input_indices",o,`max(0, min(height, ${r[o]} - 1))`)};
      ${t.indicesSet("input_indices",u,`max(0, min(width, ${r[u]} - 1))`)};
      ${Xs(t,l,i,3)}
      return ${t.getByIndices("input_indices")};
    }

    fn trilinearInterpolation(output_indices: ${e.type.indices}) -> ${p} {
      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
      var depth:${p} = originalIndices[${s}];
      var height:${p} = originalIndices[${o}];
      var width:${p} = originalIndices[${u}];
      ${n?`if (depth < 0 || depth > (${r[s]} - 1) || height < 0 || height > (${r[o]} - 1) || width < 0 || (width > ${r[u]} - 1)) {
      return ${a};
        }`:""};

    depth = max(0, min(depth, ${r[s]} - 1));
      height = max(0, min(height, ${r[o]} - 1));
      width = max(0, min(width, ${r[u]} - 1));
      var depth1: u32 = u32(depth);
      var height1: u32 = u32(height);
      var width1: u32 = u32(width);
      var depth2: u32 = u32(depth + 1);
      var height2: u32 = u32(height + 1);
      var width2: u32 = u32(width + 1);
      var channel: u32 = ${r.length>3?`u32(originalIndices[${l}])`:"0"};
      var batch: u32 =  ${r.length>3?`u32(originalIndices[${i}])`:"0"};

      var x111: ${p} = getInputValue(batch, channel, depth1, height1, width1);
      var x112: ${p} = getInputValue(batch, channel, depth1, height1, width2);
      var x121: ${p} = getInputValue(batch, channel, depth1, height2, width1);
      var x122: ${p} = getInputValue(batch, channel, depth1, height2, width2);
      var x211: ${p} = getInputValue(batch, channel, depth2, height1, width1);
      var x212: ${p} = getInputValue(batch, channel, depth2, height1, width2);
      var x221: ${p} = getInputValue(batch, channel, depth2, height2, width1);
      var x222: ${p} = getInputValue(batch, channel, depth2, height2, width2);
      var dx1: ${p} = abs(depth - ${p}(depth1));
      var dx2: ${p} = abs(${p}(depth2) - depth);
      var dy1: ${p} = abs(height - ${p}(height1));
      var dy2: ${p} = abs(${p}(height2) - height);
      var dz1: ${p} = abs(width - ${p}(width1));
      var dz2: ${p} = abs(${p}(width2) - width);
      if (depth1 == depth2) {
        dx1 = 0.5;
        dx2 = 0.5;
      }
      if (height1 == height2) {
        dy1 = 0.5;
        dy2 = 0.5;
      }
      if (width1 == width2) {
        dz1 = 0.5;
        dz2 = 0.5;
      }
      return (x111 * dx2 * dy2 * dz2 + x112 * dx2 * dy2 * dz1 + x121 * dx2 * dy1 *dz2 + x122 * dx2 * dy1 * dz1 +
              x211 * dx1 * dy2 * dz2 + x212 * dx1 * dy2 * dz1 + x221 * dx1 * dy1 *dz2 + x222 * dx1 * dy1 * dz1);
    }`},Sc=(t,e,r,n,a,i)=>{let s=t.dims,o=mc(i,e.axes,s.length),u=gc(s,n,a,e.axes),l=n.slice();n.length===0&&(l=s.map(($,C)=>$===0?1:u[C]/$),e.keepAspectRatioPolicy!=="stretch"&&(u=_c(s,l,e)));let p=ge("output",t.dataType,u.length),f=Q("input",t.dataType,s.length),m=Y.size(u),c=s.length===u.length&&s.every(($,C)=>$===u[C]),y=e.coordinateTransformMode==="tf_crop_and_resize",w=e.extrapolationValue,v=f.type.value,k=$=>`
      ${c?"":`
      ${hc(e.coordinateTransformMode,v)};
      ${(()=>{switch(e.mode){case"nearest":return`
              ${bc(f,s)};
              ${fc(e.nearestMode,r,v)};
              ${wc(f,p,s,u,l.length,o.length,y)};
              `;case"linear":return`
              ${yc(p,s,u,l.length,o.length)};
              ${(()=>{if(s.length===2||s.length===4)return`${vc(f,p,s,y,w)}`;if(s.length===3||s.length===5)return`${xc(f,p,s,y,w)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()};
            `;case"cubic":return`
            ${(()=>{if(s.length===2||s.length===4)return`${$c(f,p,s,u,l,o,e.cubicCoeffA,y,e.extrapolationValue,e.excludeOutside)}`;throw Error("Cubic mode only supports input dims 2 and 4 are supported in linear mode.")})()};
            `;default:throw Error("Invalid resize mode")}})()};
      `}
      ${$.registerUniform("output_size","u32").registerUniform("scales","f32",l.length).registerUniform("roi","f32",o.length).declareVariables(f,p)}
      ${$.mainStart()}
        ${$.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
        ${c?"output[global_idx] = input[global_idx];":`
        let output_indices = ${p.offsetToIndices("global_idx")};
        var input_indices: ${f.type.indices};
        ${(()=>{switch(e.mode){case"nearest":return`input_indices = calculateInputIndicesFromOutputIndices(output_indices);
                if (checkInputIndices(input_indices)) {
                  output[global_idx] = ${f.getByIndices("input_indices")};
                } else {
                  output[global_idx] = ${e.extrapolationValue};
                }`;case"linear":return`output[global_idx] = ${s.length===2||s.length===4?"bilinearInterpolation":"trilinearInterpolation"}(output_indices);`;case"cubic":return"output[global_idx] = bicubicInterpolation(output_indices);";default:throw Error(`Unsupported resize mode: ${e.mode}`)}})()};
`}
      }`;return{name:"Resize",shaderCache:{hint:`${e.cacheKey}|${r}|${l.length>0?l:""}|${a.length>0?a:""}|${o.length>0?o:""}|${c}|${s}`,inputDependencies:["rank"]},getShaderSource:k,getRunData:()=>({outputs:[{dims:u,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:[{type:12,data:m},{type:1,data:l},{type:1,data:o},...ye(s,u)]})}},kc=t=>{let e=t.customDataBuffer;return new Uint32Array(e,e.byteOffset,1)[0]},$m=(t,e)=>{let r=[],n=[],a=[],i=kc(t);if(e.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");pc(t.inputs,e,i,r,n,a),t.compute(Sc(t.inputs[0],e,i,r,n,a),{inputs:[0]})},xm=t=>{let e=t.antialias,r=t.axes,n=t.coordinateTransformMode,a=t.cubicCoeffA,i=t.excludeOutside!==0,s=t.extrapolationValue,o=t.keepAspectRatioPolicy,u=t.mode,l=t.nearestMode===""?"simple":t.nearestMode;return qe({antialias:e,axes:r,coordinateTransformMode:n,cubicCoeffA:a,excludeOutside:i,extrapolationValue:s,keepAspectRatioPolicy:o,mode:u,nearestMode:l})}}),Ec,Cc,Sm,Ey=Z(()=>{$e(),Ae(),lt(),Te(),Ec=(t,e)=>{let[r,n,a,i]=t,{numHeads:s,rotaryEmbeddingDim:o}=e;if(r.dims.length!==3&&r.dims.length!==4)throw new Error(`Input 'x' is expected to have 3 or 4 dimensions, got ${r.dims.length}`);if(!Y.areEqual(n.dims,[])&&!Y.areEqual(n.dims,[1])&&n.dims.length!==2)throw new Error(`Input 'position_ids' is expected to have 0, 1, or 2 dimensions, got ${n.dims.length}`);if(a.dims.length!==2)throw new Error(`Input 'cos_cache' is expected to have 2 dimensions, got ${a.dims.length}`);if(i.dims.length!==2)throw new Error(`Input 'sin_cache' is expected to have 2 dimensions, got ${i.dims.length}`);if(!Y.areEqual(a.dims,i.dims))throw new Error("Inputs 'cos_cache' and 'sin_cache' are expected to have the same shape");if(o>0&&s===0)throw new Error("num_heads must be provided if rotary_embedding_dim is specified");let u=r.dims[0],l=r.dims[r.dims.length-2],p=a.dims[0],f=Y.sizeFromDimension(r.dims,1)/l,m=o===0?a.dims[1]*2:f/s;if(o>m)throw new Error("rotary_embedding_dim must be less than or equal to head_size");if(n.dims.length===2){if(u!==n.dims[0])throw new Error(`Input 'position_ids' dimension 0 should be of size batch_size, got ${n.dims[0]}`);if(l!==n.dims[1])throw new Error(`Input 'position_ids' dimension 1 should be of size sequence_length, got ${n.dims[1]}`)}if(m/2!==a.dims[1]&&o/2!==a.dims[1])throw new Error(`Input 'cos_cache' dimension 1 should be same as head_size / 2 or rotary_embedding_dim / 2, got ${a.dims[1]}`);if(l>p)throw new Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported")},Cc=(t,e)=>{let{interleaved:r,numHeads:n,rotaryEmbeddingDim:a,scale:i}=e,s=t[0].dims[0],o=Y.sizeFromDimension(t[0].dims,1),u=t[0].dims[t[0].dims.length-2],l=o/u,p=t[2].dims[1],f=a===0?p*2:l/n,m=new Array(s,u,l/f,f-p),c=Y.computeStrides(m),y=[{type:1,data:i},{type:12,data:m},{type:12,data:c},...t[0].dims.length===3?new Array({type:12,data:[o,l,f,1]}):[],...t[0].dims.length===4?new Array({type:12,data:[o,f,u*f,1]}):[],...ye(t[0].dims,t[1].dims,t[2].dims,t[3].dims,t[0].dims)],w=v=>{let k=Q("input",t[0].dataType,t[0].dims.length),$=Q("position_ids",t[1].dataType,t[1].dims.length),C=Q("cos_cache",t[2].dataType,t[2].dims.length),T=Q("sin_cache",t[3].dataType,t[3].dims.length),A=ge("output",t[0].dataType,t[0].dims.length);return v.registerUniforms([{name:"scale",type:"f32"},{name:"global_shape",type:"u32",length:m.length},{name:"global_strides",type:"u32",length:c.length},{name:"input_output_strides",type:"u32",length:c.length}]),`
        ${v.declareVariables(k,$,C,T,A)}

        ${v.mainStart(En)}
          let half_rotary_emb_dim = uniforms.${C.name}_shape[1];
          let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape;
          let size = uniforms.global_shape[0] * uniforms.global_strides[0];
          ${v.guardAgainstOutOfBoundsWorkgroupSizes("size")}

          if (bsnh[3] < half_rotary_emb_dim) {
            let position_ids_idx =
                ${$.broadcastedIndicesToOffset("bsnh.xy",ge("",$.type.tensor,2))};
            let position_id =
                u32(${$.getByOffset("position_ids_idx")}) + select(0, bsnh[1], position_ids_idx == 0);
            let i = dot(bsnh, uniforms.input_output_strides) + select(0, bsnh[3], ${r});
            let j = i + select(half_rotary_emb_dim, 1, ${r});
            let re = ${k.getByOffset("i")} * ${C.get("position_id","bsnh[3]")} -
                ${k.getByOffset("j")} * ${T.get("position_id","bsnh[3]")};
            ${A.setByOffset("i","re")}
            let im = ${k.getByOffset("i")} * ${T.get("position_id","bsnh[3]")} +
                ${k.getByOffset("j")} * ${C.get("position_id","bsnh[3]")};
            ${A.setByOffset("j","im")}
          } else {
            let k = dot(bsnh, uniforms.input_output_strides) + half_rotary_emb_dim;
            ${A.setByOffset("k",k.getByOffset("k"))}
          }
        }`};return{name:"RotaryEmbedding",shaderCache:{hint:qe({interleaved:r}).cacheKey,inputDependencies:["rank","rank","rank","rank"]},getShaderSource:w,getRunData:()=>({outputs:[{dims:t[0].dims,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(Y.size(m)/En)},programUniforms:y})}},Sm=(t,e)=>{Ec(t.inputs,e),t.compute(Cc(t.inputs,e))}}),Tc,Ic,km,Cy=Z(()=>{$e(),Ae(),Te(),Tc=t=>{if(!t||t.length<3)throw new Error("layerNorm requires at least 3 inputs.");let e=t[0],r=t[1],n=t[2];if(e.dataType!==r.dataType||e.dataType!==n.dataType)throw new Error("All inputs must have the same data type");if(e.dims.length!==3&&e.dims.length!==2)throw new Error("Input must be 2D or 3D");if(r.dims.length!==3&&r.dims.length!==2)throw new Error("Skip must be 2D or 3D");let a=e.dims[e.dims.length-1],i=e.dims[e.dims.length-2];if(r.dims[r.dims.length-1]!==a)throw new Error("Skip must have the same hidden size as input");if(r.dims[r.dims.length-2]!==i)throw new Error("Skip must have the same sequence length as input");if(n.dims.length!==1)throw new Error("Gamma must be 1D");if(n.dims[n.dims.length-1]!==a)throw new Error("Gamma must have the same hidden size as input");if(t.length>3){let s=t[3];if(s.dims.length!==1)throw new Error("Beta must be 1D");if(s.dims[s.dims.length-1]!==a)throw new Error("Beta must have the same hidden size as input")}if(t.length>4){let s=t[4];if(s.dims.length!==1)throw new Error("Bias must be 1D");if(s.dims[s.dims.length-1]!==a)throw new Error("Bias must have the same hidden size as input")}},Ic=(t,e,r,n)=>{let a=e.simplified,i=t[0].dims,s=Y.size(i),o=i,u=s,l=i.slice(-1)[0],p=n?i.slice(0,-1).concat(1):[],f=!a&&t.length>3,m=t.length>4,c=n&&r>1,y=n&&r>2,w=r>3,v=it(l),k=[{type:12,data:u},{type:12,data:v},{type:12,data:l},{type:1,data:e.epsilon}],$=T=>{let A=[{name:"output_size",type:"u32"},{name:"components",type:"u32"},{name:"hidden_size",type:"u32"},{name:"epsilon",type:"f32"}],B=[Q("x",t[0].dataType,t[0].dims,v),Q("skip",t[1].dataType,t[1].dims,v),Q("gamma",t[2].dataType,t[2].dims,v)];f&&B.push(Q("beta",t[3].dataType,t[3].dims,v)),m&&B.push(Q("bias",t[4].dataType,t[4].dims,v)),B.push(ge("output",t[0].dataType,o,v)),c&&B.push(ge("mean_output",1,p)),y&&B.push(ge("inv_std_output",1,p)),w&&B.push(ge("input_skip_bias_sum",t[0].dataType,o,v));let R=pt(t[0].dataType);return`

      ${T.registerUniforms(A).declareVariables(...B)}

      ${T.mainStart()}
        ${T.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size / uniforms.hidden_size")}
        let hidden_size_vectorized: u32 = uniforms.hidden_size / uniforms.components;
        let offset = global_idx * hidden_size_vectorized;
        var sum = ${yr("f32",v)};
        var squareSum = ${yr("f32",v)};
        for (var i: u32 = 0; i < hidden_size_vectorized; i++) {
          let skip_value = skip[offset + i];
          let bias_value = ${m?"bias[i]":R+"(0.0)"};
          let input_value = x[offset + i];
          let value = input_value + skip_value + bias_value;
          ${w?"input_skip_bias_sum[offset + i] = value;":""}
          output[offset + i] = value;
          let f32_value = ${xn(R,v,"value")};
          sum += f32_value;
          squareSum += f32_value * f32_value;
        }
        let mean = ${Mr("sum",v)} / f32(uniforms.hidden_size);
        let inv_std_dev = inverseSqrt(${Mr("squareSum",v)} / f32(uniforms.hidden_size) ${a?"":"- mean * mean"} + uniforms.epsilon);
        ${c?"mean_output[global_idx] = mean;":""}
        ${y?"inv_std_output[global_idx] = inv_std_dev;":""}
        for (var i: u32 = 0; i < hidden_size_vectorized; i++) {
          output[offset + i] = (output[offset + i] ${a?"":`- ${R}(mean)`}) * ${R}(inv_std_dev) * gamma[i] ${f?"+ beta[i]":""};
        }
      }`},C=[{dims:o,dataType:t[0].dataType}];return r>1&&C.push({dims:p,dataType:1}),r>2&&C.push({dims:p,dataType:1}),r>3&&C.push({dims:i,dataType:t[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:`${v};${c};${y};${w}`,inputDependencies:t.map((T,A)=>"type")},getShaderSource:$,getRunData:()=>({outputs:C,dispatchGroup:{x:Math.ceil(u/l/64)},programUniforms:k})}},km=(t,e)=>{Tc(t.inputs);let r=[0];t.outputCount>1&&r.push(-3),t.outputCount>2&&r.push(-3),t.outputCount>3&&r.push(3),t.compute(Ic(t.inputs,e,t.outputCount,!1),{outputs:r})}}),Ac,ta,Mc,Qs,Oc,zc,Em,Cm,Ty=Z(()=>{$e(),Ae(),lt(),Te(),Ac=(t,e)=>{if(!t||t.length<1)throw new Error("too few inputs");if(e.axes.length!==0){if(e.axes.length!==e.starts.length||e.axes.length!==e.ends.length)throw new Error("axes, starts and ends must have the same length")}else if(e.starts.length!==e.ends.length)throw new Error("starts and ends must have the same length");t.slice(1).forEach((r,n)=>{if(t[n+1].dataType!==6&&t[n+1].dataType!==7)throw new Error(`Input ${n} must be an array of int32 or int64`)})},ta=(t,e)=>{let r=[];if(t.length>e)if(t[e].dataType===7)t[e].getBigInt64Array().forEach(n=>r.push(Number(n)));else if(t[e].dataType===6)t[e].getInt32Array().forEach(n=>r.push(Number(n)));else throw new Error(`Input ${e} must be an array of int32 or int64`);return r},Mc=(t,e)=>{if(t.length>1){let r=ta(t,1),n=ta(t,2),a=ta(t,3);return a.length===0&&(a=[...Array(t[0].dims.length).keys()]),qe({starts:r,ends:n,axes:a})}else return e},Qs=(t,e,r,n,a)=>{let i=t;return t<0&&(i+=r[n[e]]),a[e]<0?Math.max(0,Math.min(i,r[n[e]]-1)):Math.max(0,Math.min(i,r[n[e]]))},Oc=(t,e,r)=>`fn calculateInputIndices(output_indices: ${e.type.indices}) -> ${t.type.indices} {
          var input_indices: ${t.type.indices};
          var carry = 0u;
          for (var i = ${r.length}; i >= 0; i--) {
            let input_shape_i = ${Se("uniforms.input_shape","i",r.length)};
            let steps_i = ${Se("uniforms.steps","i",r.length)};
            let signs_i = ${Se("uniforms.signs","i",r.length)};
            let starts_i = ${Se("uniforms.starts","i",r.length)};
            var output_index = ${e.indicesGet("output_indices","i")};
            var input_index = output_index * steps_i + starts_i + carry;
            carry = input_index / input_shape_i;
            input_index = input_index % input_shape_i;
            if (signs_i < 0) {
              input_index = input_shape_i - input_index - 1u + starts_i;
            }
            ${t.indicesSet("input_indices","i","input_index")};
          }
          return input_indices;
      }`,zc=(t,e)=>{let r=t[0].dims,n=Y.size(r),a=e.axes.length>0?Y.normalizeAxes(e.axes,r.length):[...Array(r.length).keys()],i=ta(t,4);i.forEach(k=>k!==0||(()=>{throw new Error("step cannot be 0")})),i.length===0&&(i=Array(a.length).fill(1));let s=e.starts.map((k,$)=>Qs(k,$,r,a,i)),o=e.ends.map((k,$)=>Qs(k,$,r,a,i));if(a.length!==s.length||a.length!==o.length)throw new Error("start, ends and axes should have the same number of elements");if(a.length!==r.length)for(let k=0;k<r.length;++k)a.includes(k)||(s.splice(k,0,0),o.splice(k,0,r[k]),i.splice(k,0,1));let u=i.map(k=>Math.sign(k));i.forEach((k,$,C)=>{if(k<0){let T=(o[$]-s[$])/k,A=s[$],B=A+T*i[$];s[$]=B,o[$]=A,C[$]=-k}});let l=r.slice(0);a.forEach((k,$)=>{l[k]=Math.ceil((o[k]-s[k])/i[k])});let p={dims:l,dataType:t[0].dataType},f=ge("output",t[0].dataType,l.length),m=Q("input",t[0].dataType,t[0].dims.length),c=Y.size(l),y=[{name:"outputSize",type:"u32"},{name:"starts",type:"u32",length:s.length},{name:"signs",type:"i32",length:u.length},{name:"steps",type:"u32",length:i.length}],w=[{type:12,data:c},{type:12,data:s},{type:6,data:u},{type:12,data:i},...ye(t[0].dims,l)],v=k=>`
      ${k.registerUniforms(y).declareVariables(m,f)}
        ${Oc(m,f,r)}
        ${k.mainStart()}
          ${k.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
          let output_indices = ${f.offsetToIndices("global_idx")};
          let input_indices = calculateInputIndices(output_indices);
          ${f.setByOffset("global_idx",m.getByIndices("input_indices"))}
      }`;return{name:"Slice",shaderCache:{hint:`${u.length}_${s.length}_${i.length}`,inputDependencies:["rank"]},getShaderSource:v,getRunData:()=>({outputs:[p],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:w})}},Em=(t,e)=>{Ac(t.inputs,e);let r=Mc(t.inputs,e);t.compute(zc(t.inputs,r),{inputs:[0]})},Cm=t=>{let e=t.starts,r=t.ends,n=t.axes;return qe({starts:e,ends:r,axes:n})}}),Pc,Rc,Tm,Im,Iy=Z(()=>{$e(),Ae(),lt(),Te(),Pc=t=>{if(!t||t.length!==1)throw new Error("Softmax op requires 1 input.")},Rc=(t,e)=>{let r=t.dims,n=Y.size(r),a=64,i=e.axis;if(i<0&&(i=r.length+i),i<r.length-1)throw new Error("softmax only supports last axis for now.");let s=r[i],o=n/s,u=it(s),l=s/u,p=(v,k)=>k===4?`max(max(${v}.x, ${v}.y), max(${v}.z, ${v}.w))`:k===2?`max(${v}.x, ${v}.y)`:k===3?`max(max(${v}.x, ${v}.y), ${v}.z)`:v,f=Q("x",t.dataType,t.dims,u),m=ge("result",t.dataType,t.dims,u),c=f.type.value,y=pt(t.dataType)==="f32"?`var threadMax = ${c}(-3.402823e+38f);`:`var threadMax = ${c}(-65504.0h);`,w=v=>`
      var<workgroup> rowMaxShared : ${c};
      var<workgroup> rowSumShared : ${c};
      var<workgroup> threadShared : array<${c}, ${a}>;

      fn getValue(row: i32, col: i32, row_stride: i32) -> ${c} {
        let index = row * row_stride + col;
        return x[index];
      }

      fn setValue(row: i32, col: i32, row_stride: i32, value: ${c}) {
        let index = row * row_stride + col;
        result[index] = value;
      }
      ${v.registerUniform("packedCols","i32").declareVariables(f,m)}
      ${v.mainStart()}
        let gindex = i32(global_idx);
        let lindex = i32(local_idx);
        const wg = ${a};
        let row = gindex / wg;
        let cols = uniforms.packedCols;
        let row_stride : i32 = uniforms.packedCols;

        // find the rows max
        ${y}
        for (var col = lindex; col < cols; col += wg) {
          let value = getValue(row, col, row_stride);
          threadMax = max(threadMax, value);
        }
        if (lindex < cols) {
          threadShared[lindex] = threadMax;
        }
        workgroupBarrier();

        var reduceSize = min(cols, wg);
        for (var currSize = reduceSize >> 1;  currSize > 0; currSize = reduceSize >> 1) {
          reduceSize = currSize + (reduceSize & 1);
          if (lindex < currSize) {
            threadShared[lindex] = max(threadShared[lindex], threadShared[lindex + reduceSize]);
          }
          workgroupBarrier();
        }
        if (lindex == 0) {
          rowMaxShared = ${c}(${p("threadShared[0]",u)});
        }
        workgroupBarrier();

        // find the rows sum
        var threadSum = ${c}(0.0);
        for (var col = lindex; col < cols; col += wg) {
          let subExp = exp(getValue(row, col, row_stride) - rowMaxShared);
          threadSum += subExp;
        }
        threadShared[lindex] = threadSum;
        workgroupBarrier();

        for (var currSize = wg >> 1;  currSize > 0; currSize = currSize >> 1) {
          if (lindex < currSize) {
            threadShared[lindex] = threadShared[lindex] + threadShared[lindex + currSize];
          }
          workgroupBarrier();
        }
        if (lindex == 0) {
          rowSumShared = ${c}(${Mr("threadShared[0]",u)});
        }
        workgroupBarrier();

        // calculate final value for each element in the row
        for (var col = lindex; col < cols; col += wg) {
          let value = exp(getValue(row, col, row_stride) - rowMaxShared) / rowSumShared;
          setValue(row, col, row_stride, value);
        }
      }`;return{name:"Softmax",shaderCache:{hint:`${u}`,inputDependencies:["type"]},getRunData:()=>({outputs:[{dims:r,dataType:t.dataType}],dispatchGroup:{x:o},programUniforms:[{type:6,data:l}]}),getShaderSource:w}},Tm=(t,e)=>{Pc(t.inputs),t.compute(Rc(t.inputs[0],e))},Im=t=>qe({axis:t.axis})}),Bc,Dc,Nc,Fc,Lc,Am,Mm,Ay=Z(()=>{$e(),Ae(),lt(),Te(),Bc=t=>{if(!t||t.length<1)throw new Error("too few inputs")},Dc=(t,e)=>{let r=[],n=e.numOutputs;return t[1].dims[0]>0&&(t[1].getBigInt64Array().forEach(a=>r.push(Number(a))),n=r.length),qe({numOutputs:n,axis:e.axis,splitSizes:r})},Nc=t=>`
fn calculateOutputIndex(index: u32) -> u32 {
    for (var i: u32 = 0u; i < ${t}u; i += 1u ) {
    if (index < ${Se("uniforms.size_in_split_axis","i",t)}) {
        return i;
    }
    }
    return ${t}u;
}`,Fc=t=>{let e=t.length,r=[];for(let n=0;n<e;++n){let a=t[n].setByIndices("indices","input[global_idx]");e===1?r.push(a):n===0?r.push(`if (output_number == ${n}u) { ${a} }`):n===e-1?r.push(`else { ${a} }`):r.push(`else if (output_number == ${n}) { ${a} }`)}return`
      fn writeBufferData(output_number: u32, indices: ${t[0].type.indices}, global_idx: u32) {
        ${r.join(`
`)}
      }`},Lc=(t,e)=>{let r=t[0].dims,n=Y.size(r),a=t[0].dataType,i=Y.normalizeAxis(e.axis,r.length),s=new Array(e.numOutputs),o=Q("input",a,r.length),u=new Array(e.numOutputs),l=[],p=[],f=0,m=[{type:12,data:n}];for(let y=0;y<e.numOutputs;y++){f+=e.splitSizes[y],u[y]=f;let w=r.slice();w[e.axis]=e.splitSizes[y],p.push(w),s[y]=ge(`output${y}`,a,w.length),l.push({dims:p[y],dataType:t[0].dataType})}m.push({type:12,data:u},...ye(r,...p));let c=y=>`
  ${y.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",u.length).declareVariables(o,...s)}
  ${Nc(u.length)}
  ${Fc(s)}

  ${y.mainStart()}
    ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size")}

    var indices = ${o.offsetToIndices("global_idx")};
    var index = ${o.indicesGet("indices",i)};
    let output_number = calculateOutputIndex(index);
    if (output_number != 0) {
      index -= ${Se("uniforms.size_in_split_axis","output_number - 1u",u.length)};
      ${o.indicesSet("indices",i,"index")};
    }
    writeBufferData(output_number, indices, global_idx);
  }`;return{name:"Split",shaderCache:{hint:e.cacheKey,inputDependencies:["rank"]},getShaderSource:c,getRunData:()=>({outputs:l,dispatchGroup:{x:Math.ceil(n/64)},programUniforms:m})}},Am=(t,e)=>{Bc(t.inputs);let r=t.inputs.length===1?e:Dc(t.inputs,e);t.compute(Lc(t.inputs,r),{inputs:[0]})},Mm=t=>{let e=t.axis,r=t.splitSizes,n=t.numOutputs<0?r.length:t.numOutputs;if(n!==r.length)throw new Error("numOutputs and splitSizes lengh must be equal");return qe({axis:e,numOutputs:n,splitSizes:r})}}),Js,Uc,Wc,Vc,Om,My=Z(()=>{$e(),Ae(),Te(),Js=t=>Array.from(t.getBigInt64Array(),Number),Uc=t=>{if(!t||t.length!==2)throw new Error("Tile requires 2 inputs.");if(t[0].dataType!==1&&t[0].dataType!==6&&t[0].dataType!==12)throw new Error("Tile only support float, int32, and uint32 data types");if(t[1].dataType!==7)throw new Error("Tile `repeats` input should be of int64 data type");if(t[1].dims.length!==1)throw new Error("Tile `repeats` input should be 1-D");if(Js(t[1]).length!==t[0].dims.length)throw new Error("Tile `repeats` input should have same number of elements as rank of input data tensor")},Wc=(t,e)=>{let r=[];for(let n=0;n<t.length;++n)r.push(t[n]*e[n]);return r},Vc=t=>{let e=t[0].dims,r=Js(t[1]),n=Wc(e,r),a=Y.size(n),i=t[0].dataType,s=Q("input",i,e.length),o=ge("output",i,n.length),u=l=>`
      const inputShape = ${s.indices(...e)};
      ${l.registerUniform("output_size","u32").declareVariables(s,o)}
      ${l.mainStart()}
      ${l.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
      let output_indices = ${o.offsetToIndices("global_idx")};
      var input_indices: ${s.type.indices};
      for (var i = 0; i < ${e.length}; i++) {
        let input_dim_i = ${s.indicesGet("uniforms.input_shape","i")};
        let input_dim_value = ${o.indicesGet("output_indices","i")}  % input_dim_i;

        ${s.indicesSet("input_indices","i","input_dim_value")}
      }
      ${o.setByOffset("global_idx",s.getByIndices("input_indices"))}
    }`;return{name:"Tile",shaderCache:{hint:`${r}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:n,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:[{type:12,data:a},...ye(t[0].dims,n)]}),getShaderSource:u}},Om=t=>{Uc(t.inputs),t.compute(Vc(t.inputs),{inputs:[0]})}}),Gc,Hc,zm,Oy=Z(()=>{$e(),Ae(),Te(),Gc=(t,e,r,n,a)=>{let i=ge("output_data",a,r.length,4),s=Q("a_data",e[1].dataType,e[1].dims.length,4),o=Q("b_data",e[2].dataType,e[2].dims.length,4),u=Q("c_data",e[0].dataType,e[0].dims.length,4),l,p=(f,m,c)=>`select(${m}, ${f}, ${c})`;if(!n)l=i.setByOffset("global_idx",p(s.getByOffset("global_idx"),o.getByOffset("global_idx"),u.getByOffset("global_idx")));else{let f=(m,c,y="")=>{let w=`a_data[index_a${c}][component_a${c}]`,v=`b_data[index_b${c}][component_b${c}]`,k=`bool(c_data[index_c${c}] & (0xffu << (component_c${c} * 8)))`;return`
            let output_indices${c} = ${i.offsetToIndices(`global_idx * 4u + ${c}u`)};
            let offset_a${c} = ${s.broadcastedIndicesToOffset(`output_indices${c}`,i)};
            let offset_b${c} = ${o.broadcastedIndicesToOffset(`output_indices${c}`,i)};
            let offset_c${c} = ${u.broadcastedIndicesToOffset(`output_indices${c}`,i)};
            let index_a${c} = offset_a${c} / 4u;
            let index_b${c} = offset_b${c} / 4u;
            let index_c${c} = offset_c${c} / 4u;
            let component_a${c} = offset_a${c} % 4u;
            let component_b${c} = offset_b${c} % 4u;
            let component_c${c} = offset_c${c} % 4u;
            ${m}[${c}] = ${y}(${p(w,v,k)});
          `};a===9?l=`
            var data = vec4<u32>(0);
            ${f("data",0,"u32")}
            ${f("data",1,"u32")}
            ${f("data",2,"u32")}
            ${f("data",3,"u32")}
            output_data[global_idx] = dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(data));`:l=`
            ${f("output_data[global_idx]",0)}
            ${f("output_data[global_idx]",1)}
            ${f("output_data[global_idx]",2)}
            ${f("output_data[global_idx]",3)}
          `}return`
        ${t.registerUniform("vec_size","u32").declareVariables(u,s,o,i)}
        ${t.mainStart()}
        ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
        ${l}
      }`},Hc=t=>{let e=t[1].dims,r=t[2].dims,n=t[0].dims,a=t[1].dataType,i=!(Y.areEqual(e,r)&&Y.areEqual(r,n)),s=e,o=Y.size(e);if(i){let l=kn.calcShape(kn.calcShape(e,r,!1),n,!1);if(!l)throw new Error("Can't perform where op on the given tensors");s=l,o=Y.size(s)}let u=Math.ceil(o/4);return{name:"Where",shaderCache:{inputDependencies:["rank","rank","rank"]},getShaderSource:l=>Gc(l,t,s,i,a),getRunData:()=>({outputs:[{dims:s,dataType:a}],dispatchGroup:{x:Math.ceil(o/64/4)},programUniforms:[{type:12,data:u},...ye(n,e,r,s)]})}},zm=t=>{t.compute(Hc(t.inputs))}}),Pm,zy=Z(()=>{ey(),Wh(),ty(),ry(),ny(),ay(),Lh(),Wf(),ly(),dy(),cy(),py(),hy(),fy(),my(),gy(),_y(),yy(),wy(),Uf(),by(),vy(),$y(),xy(),Sy(),Xo(),ky(),Ey(),Cy(),Ty(),Iy(),Ay(),My(),$a(),Qo(),Oy(),Pm=new Map([["Abs",[Hh]],["Acos",[jh]],["Acosh",[qh]],["Add",[Cf]],["ArgMax",[Dh,wo]],["ArgMin",[Bh,wo]],["Asin",[Kh]],["Asinh",[Yh]],["Atan",[Xh]],["Atanh",[Qh]],["Attention",[Uh]],["AveragePool",[fm,hm]],["BatchNormalization",[Vh]],["BiasAdd",[Gh]],["BiasSplitGelu",[Ef]],["Cast",[Zh,Jh]],["Ceil",[tf]],["Clip",[ef]],["Concat",[Nh,Fh]],["Conv",[ko,So]],["ConvTranspose",[Hf,Gf]],["Cos",[rf]],["Cosh",[nf]],["CumSum",[jf,qf]],["DepthToSpace",[Kf,Yf]],["Div",[Tf]],["Einsum",[Xf,Qf]],["Elu",[af,Ci]],["Equal",[If]],["Erf",[sf]],["Exp",[of]],["Expand",[Jf]],["FastGelu",[Zf]],["Floor",[uf]],["FusedConv",[ko,So]],["Gather",[tm,em]],["GatherElements",[nm,rm]],["Gelu",[lf]],["Gemm",[im,am]],["GlobalAveragePool",[gm,mm]],["GlobalMaxPool",[bm,wm]],["Greater",[zf]],["GreaterOrEqual",[Rf]],["HardSigmoid",[_f,gf]],["InstanceNormalization",[sm]],["LayerNormalization",[om]],["LeakyRelu",[df,Ci]],["Less",[Pf]],["LessOrEqual",[Bf]],["Log",[kf]],["MatMul",[Lf]],["MatMulNBits",[um,lm]],["MaxPool",[_m,ym]],["Mul",[Af]],["MultiHeadAttention",[cm,dm]],["Neg",[pf]],["Not",[cf]],["Pad",[pm]],["Pow",[Mf]],["Range",[vm]],["Reciprocal",[hf]],["ReduceMin",[Mh]],["ReduceMean",[Eh]],["ReduceMax",[Ah]],["ReduceSum",[zh]],["ReduceProd",[Oh]],["ReduceL1",[Ch]],["ReduceL2",[Th]],["ReduceLogSum",[Rh]],["ReduceLogSumExp",[Ih]],["ReduceSumSquare",[Ph]],["Relu",[ff]],["Resize",[$m,xm]],["RotaryEmbedding",[Sm]],["Sigmoid",[mf]],["Sin",[yf]],["Sinh",[wf]],["Slice",[Em,Cm]],["SkipLayerNormalization",[km]],["Split",[Am,Mm]],["Sqrt",[bf]],["Softmax",[Tm,Im]],["Sub",[Of]],["Tan",[vf]],["Tanh",[$f]],["ThresholdedRelu",[Sf,Ci]],["Tile",[Om]],["Transpose",[fh,mh]],["Where",[zm]]])}),Rm,Py=Z(()=>{ar(),rn(),Te(),Rm=class{constructor(t){this.backend=t,this.repo=new Map,this.attributesBound=!1}getArtifact(t){return this.repo.get(t)}setArtifact(t,e){this.repo.set(t,e)}run(t,e,r,n,a){nr(t.programInfo.name);let i=this.backend.device,s=this.backend.getComputePassEncoder();this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2);let o=[];for(let l of e)o.push({binding:o.length,resource:{buffer:l.buffer}});for(let l of r)o.push({binding:o.length,resource:{buffer:l.buffer}});a&&o.push({binding:o.length,resource:a});let u=i.createBindGroup({layout:t.computePipeline.getBindGroupLayout(0),entries:o,label:t.programInfo.name});if(this.backend.sessionStatus==="capturing"){let l={kernelId:this.backend.currentKernelId,computePipeline:t.computePipeline,bindGroup:u,dispatchGroup:n};this.backend.capturedCommandList.get(this.backend.currentSessionId).push(l)}s.setPipeline(t.computePipeline),s.setBindGroup(0,u),s.dispatchWorkgroups(...n),this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2+1),this.backend.pendingDispatchNumber++,(this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber||this.backend.queryType==="at-passes")&&this.backend.endComputePass(),this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber&&this.backend.flush(),Lt(t.programInfo.name)}dispose(){}build(t,e){nr(t.name);let r=this.backend.device,n=[];r.features.has("shader-f16")&&n.push("enable f16;");let a=hh(e,this.backend.device.limits),i=t.getShaderSource(a),s=`${n.join(`
`)}
${a.additionalImplementations}
${i}`,o=r.createShaderModule({code:s,label:t.name});rt("verbose",()=>`[WebGPU] ${t.name} shader code: ${s}`);let u=r.createComputePipeline({compute:{module:o,entryPoint:"main"},layout:"auto",label:t.name});return Lt(t.name),{programInfo:t,computePipeline:u,uniformVariablesInfo:a.variablesInfo}}normalizeDispatchGroupSize(t){let e=typeof t=="number"?t:t.x,r=typeof t=="number"?1:t.y||1,n=typeof t=="number"?1:t.z||1,a=this.backend.device.limits.maxComputeWorkgroupsPerDimension;if(e<=a&&r<=a&&n<=a)return[e,r,n];let i=e*r*n,s=Math.ceil(Math.sqrt(i));if(s>a){if(s=Math.ceil(Math.cbrt(i)),s>a)throw new Error("Total dispatch size exceeds WebGPU maximum.");return[s,s,s]}else return[s,s,1]}}}),jc,qc,Kc,Bm,Ry=Z(()=>{ar(),$e(),rn(),Q_(),J_(),zy(),Py(),jc=(t,e)=>{if(e.length!==t.length)throw new Error(`inputDependencies length ${e.length} is not equal to inputTensors length ${t.length}.`);let r=[];for(let n=0;n<t.length;++n){let a=t[n].dataType;switch(e[n]){case"none":{r.push("");break}case"type":{r.push(`${a}`);break}case"rank":{let i=t[n].dims.length;r.push(`${a};${i}`);break}case"dims":{let i=t[n].dims.join(",");r.push(`${a};${i}`);break}default:throw new Error(`unsupported input dependency: ${e[n]}`)}}return r.join("|")},qc=(t,e,r)=>{let n=t.name;return t.shaderCache?.hint&&(n+="["+t.shaderCache.hint+"]"),n+=":"+r+`:${jc(e,t.shaderCache?.inputDependencies??new Array(e.length).fill("dims"))}`,n},Kc=class{constructor(t){t&&(this.architecture=t.architecture,this.vendor=t.vendor)}isArchitecture(t){return this.architecture===t}isVendor(t){return this.vendor===t}},Bm=class{constructor(){this.currentSessionId=null,this.currentKernelId=null,this.commandEncoder=null,this.computePassEncoder=null,this.maxDispatchNumber=16,this.pendingDispatchNumber=0,this.pendingKernels=[],this.pendingQueries=new Map,this.sessionStatus="default",this.capturedCommandList=new Map,this.capturedPendingKernels=new Map,this.sessionExternalDataMapping=new Map}get currentKernelCustomData(){if(this.currentKernelId===null)throw new Error("currentKernelCustomData(): currentKernelId is null. (should not happen)");let t=this.kernelCustomData.get(this.currentKernelId);return t||(t={},this.kernelCustomData.set(this.currentKernelId,t)),t}async initialize(t,e){this.env=t;let r=[],n={requiredLimits:{maxComputeWorkgroupStorageSize:e.limits.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:e.limits.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:e.limits.maxStorageBufferBindingSize,maxBufferSize:e.limits.maxBufferSize,maxComputeInvocationsPerWorkgroup:e.limits.maxComputeInvocationsPerWorkgroup,maxComputeWorkgroupSizeX:e.limits.maxComputeWorkgroupSizeX,maxComputeWorkgroupSizeY:e.limits.maxComputeWorkgroupSizeY,maxComputeWorkgroupSizeZ:e.limits.maxComputeWorkgroupSizeZ},requiredFeatures:r};e.features.has("chromium-experimental-timestamp-query-inside-passes")?r.push("chromium-experimental-timestamp-query-inside-passes"):e.features.has("timestamp-query")&&r.push("timestamp-query"),e.features.has("shader-f16")&&r.push("shader-f16"),this.device=await e.requestDevice(n),this.adapterInfo=new Kc(await e.requestAdapterInfo()),this.gpuDataManager=ch(this),this.programManager=new Rm(this),this.kernels=new Map,this.kernelPersistentData=new Map,this.kernelCustomData=new Map,lh(t.logLevel,!!t.debug),this.device.onuncapturederror=a=>{a.error instanceof GPUValidationError&&console.error(`An uncaught WebGPU validation error was raised: ${a.error.message}`)},Object.defineProperty(this.env.webgpu,"device",{value:this.device,writable:!1,enumerable:!0,configurable:!1}),Object.defineProperty(this.env.webgpu,"adapter",{value:e,writable:!1,enumerable:!0,configurable:!1}),this.setQueryType()}dispose(){typeof this.querySet<"u"&&this.querySet.destroy(),this.gpuDataManager.dispose()}getCommandEncoder(){return this.commandEncoder||(this.commandEncoder=this.device.createCommandEncoder()),this.commandEncoder}getComputePassEncoder(){if(!this.computePassEncoder){let t=this.getCommandEncoder(),e={};this.queryType==="at-passes"&&(e.timestampWrites={querySet:this.querySet,beginningOfPassWriteIndex:this.pendingDispatchNumber*2,endOfPassWriteIndex:this.pendingDispatchNumber*2+1}),this.computePassEncoder=t.beginComputePass(e)}return this.computePassEncoder}endComputePass(){this.computePassEncoder&&(this.computePassEncoder.end(),this.computePassEncoder=null)}flush(){if(!this.commandEncoder)return;nr(),this.endComputePass();let t;this.queryType!=="none"&&(this.commandEncoder.resolveQuerySet(this.querySet,0,this.pendingDispatchNumber*2,this.queryResolveBuffer,0),t=this.device.createBuffer({size:this.pendingDispatchNumber*2*8,usage:GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST}),this.pendingQueries.set(t,this.pendingKernels),this.pendingKernels=[],this.commandEncoder.copyBufferToBuffer(this.queryResolveBuffer,0,t,0,this.pendingDispatchNumber*2*8)),this.device.queue.submit([this.commandEncoder.finish()]),this.gpuDataManager.refreshPendingBuffers(),this.commandEncoder=null,this.pendingDispatchNumber=0,this.queryType!=="none"&&t.mapAsync(GPUMapMode.READ).then(()=>{let e=new BigUint64Array(t.getMappedRange()),r=this.pendingQueries.get(t);for(let n=0;n<e.length/2;n++){let a=r[n],i=a.kernelId,s=this.kernels.get(i),o=s.kernelType,u=s.kernelName,l=a.programName,p=a.inputTensorViews,f=a.outputTensorViews,m=e[n*2],c=e[n*2+1];typeof this.queryTimeBase>"u"&&(this.queryTimeBase=m);let y=Number(m-this.queryTimeBase),w=Number(c-this.queryTimeBase);if(!Number.isSafeInteger(y)||!Number.isSafeInteger(w))throw new RangeError("incorrect timestamp range");if(this.env.webgpu.profiling?.ondata)this.env.webgpu.profiling.ondata({version:1,inputsMetadata:p.map(v=>({dims:v.dims,dataType:qr(v.dataType)})),outputsMetadata:f.map(v=>({dims:v.dims,dataType:qr(v.dataType)})),kernelId:i,kernelType:o,kernelName:u,programName:l,startTime:y,endTime:w});else{let v="";p.forEach(($,C)=>{v+=`input[${C}]: [${$.dims}] | ${qr($.dataType)}, `});let k="";f.forEach(($,C)=>{k+=`output[${C}]: [${$.dims}] | ${qr($.dataType)}, `}),console.log(`[profiling] kernel "${i}|${o}|${u}|${l}" ${v}${k}execution time: ${w-y} ns`)}ha("GPU",`${l}::${m}::${c}`)}t.unmap(),this.pendingQueries.delete(t)}),Lt()}run(t,e,r,n,a,i){nr(t.name);let s=[];for(let $=0;$<e.length;++$){let C=e[$].data;if(C===0)continue;let T=this.gpuDataManager.get(C);if(!T)throw new Error(`no GPU data for input: ${C}`);s.push(T)}let{outputs:o,dispatchGroup:u,programUniforms:l}=t.getRunData(e),p=r.length===0?o.map(($,C)=>C):r;if(p.length!==o.length)throw new Error(`Output size ${p.length} must be equal to ${o.length}.`);let f=[],m=[];for(let $=0;$<o.length;++$){if(!Number.isInteger(p[$])||p[$]<-3||p[$]>=i)throw new Error(`Invalid output index: ${p[$]}`);if(p[$]===-3)continue;let C=p[$]===-1,T=p[$]===-2,A=C||T?a(o[$].dataType,o[$].dims):n(p[$],o[$].dataType,o[$].dims);if(f.push(A),A.data===0)continue;let B=this.gpuDataManager.get(A.data);if(!B)throw new Error(`no GPU data for output: ${A.data}`);if(C&&this.temporaryData.push(B),T){let R=this.kernelPersistentData.get(this.currentKernelId);R||(R=[],this.kernelPersistentData.set(this.currentKernelId,R)),R.push(B)}m.push(B)}if(s.length!==e.length||m.length!==f.length){if(m.length===0)return Lt(t.name),f;throw new Error(`Program ${t.name} has zero-sized tensor(s) in inputs or outputs. This is not supported now.`)}let c;if(l){let $=0,C=[];l.forEach(R=>{let D=typeof R.data=="number"?[R.data]:R.data;if(D.length===0)return;let K=R.type===10?2:4,j,ie;R.type===10?(ie=D.length>4?16:D.length>2?8:D.length*K,j=D.length>4?16:K*D.length):(ie=D.length<=2?D.length*K:16,j=16),$=Math.ceil($/ie)*ie,C.push($);let te=R.type===10?8:4;$+=D.length>4?Math.ceil(D.length/te)*j:D.length*K});let T=16;$=Math.ceil($/T)*T;let A=new ArrayBuffer($);l.forEach((R,D)=>{let K=C[D],j=typeof R.data=="number"?[R.data]:R.data;if(R.type===6)new Int32Array(A,K,j.length).set(j);else if(R.type===12)new Uint32Array(A,K,j.length).set(j);else if(R.type===10)new Uint16Array(A,K,j.length).set(j);else if(R.type===1)new Float32Array(A,K,j.length).set(j);else throw new Error(`Unsupported uniform type: ${qr(R.type)}`)});let B=this.gpuDataManager.create($,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.device.queue.writeBuffer(B.buffer,0,A,0,$),this.gpuDataManager.release(B.id),c={offset:0,size:$,buffer:B.buffer}}let y=this.programManager.normalizeDispatchGroupSize(u),w=y[1]===1&&y[2]===1,v=qc(t,e,w),k=this.programManager.getArtifact(v);if(k||(k=this.programManager.build(t,y),this.programManager.setArtifact(v,k),rt("info",()=>`[artifact] key: ${v}, programName: ${t.name}`)),l&&k.uniformVariablesInfo){if(l.length!==k.uniformVariablesInfo.length)throw new Error(`Uniform variables count mismatch: expect ${k.uniformVariablesInfo.length}, got ${l.length} in program "${k.programInfo.name}".`);for(let $=0;$<l.length;$++){let C=l[$],T=C.type,A=typeof C.data=="number"?1:C.data.length,[B,R]=k.uniformVariablesInfo[$];if(T!==B||A!==R)throw new Error(`Uniform variable ${$} mismatch: expect type ${B} with size ${R}, got type ${T} with size ${A} in program "${k.programInfo.name}".`)}}if(rt("info",()=>`[ProgramManager] run "${t.name}" (key=${v}) with ${y[0]}x${y[1]}x${y[2]}`),this.queryType!=="none"||this.sessionStatus==="capturing"){let $={kernelId:this.currentKernelId,programName:k.programInfo.name,inputTensorViews:e,outputTensorViews:f};this.pendingKernels.push($),this.sessionStatus==="capturing"&&this.capturedPendingKernels.get(this.currentSessionId).push($)}return this.programManager.run(k,s,m,y,c),Lt(t.name),f}upload(t,e){this.gpuDataManager.upload(t,e)}memcpy(t,e){this.gpuDataManager.memcpy(t,e)}async download(t,e){await this.gpuDataManager.download(t,e)}alloc(t){return this.gpuDataManager.create(t).id}free(t){return this.gpuDataManager.release(t)}createKernel(t,e,r,n){let a=Pm.get(t);if(!a)throw new Error(`kernel not implemented: ${t}`);let i={kernelType:t,kernelName:n,kernelEntry:a[0],attributes:[a[1],r]};this.kernels.set(e,i)}releaseKernel(t){let e=this.kernelPersistentData.get(t);if(e){for(let r of e)this.gpuDataManager.release(r.id);this.kernelPersistentData.delete(t)}this.kernelCustomData.delete(t),this.kernels.delete(t)}computeKernel(t,e,r){let n=this.kernels.get(t);if(!n)throw new Error(`kernel not created: ${t}`);let a=n.kernelType,i=n.kernelName,s=n.kernelEntry,o=n.attributes;if(this.currentKernelId!==null)throw new Error(`kernel "[${a}] ${i}" is not allowed to be called recursively`);this.currentKernelId=t,o[0]&&(o[1]=o[0](o[1]),o[0]=void 0),rt("info",()=>`[WebGPU] Start to run kernel "[${a}] ${i}"...`);let u=this.env.debug;this.temporaryData=[];try{return u&&this.device.pushErrorScope("validation"),s(e,o[1]),0}catch(l){return r.push(Promise.resolve(`[WebGPU] Kernel "[${a}] ${i}" failed. ${l}`)),1}finally{u&&r.push(this.device.popErrorScope().then(l=>l?`GPU validation error for kernel "[${a}] ${i}": ${l.message}`:null));for(let l of this.temporaryData)this.gpuDataManager.release(l.id);this.temporaryData=[],this.currentKernelId=null}}registerBuffer(t,e,r,n){let a=this.sessionExternalDataMapping.get(t);a||(a=new Map,this.sessionExternalDataMapping.set(t,a));let i=a.get(e),s=this.gpuDataManager.registerExternalBuffer(r,n,i?.[1]);return a.set(e,[s,r]),s}unregisterBuffers(t){let e=this.sessionExternalDataMapping.get(t);e&&(e.forEach(r=>this.gpuDataManager.unregisterExternalBuffer(r[1])),this.sessionExternalDataMapping.delete(t))}getBuffer(t){let e=this.gpuDataManager.get(t);if(!e)throw new Error(`no GPU data for buffer: ${t}`);return e.buffer}createDownloader(t,e,r){return async()=>{let n=await _o(this,t,e);return dh(n.buffer,r)}}writeTimestamp(t){this.queryType==="inside-passes"&&this.computePassEncoder.writeTimestamp(this.querySet,t)}setQueryType(){this.queryType="none",(this.env.webgpu.profiling?.mode==="default"||(typeof this.env.trace>"u"?this.env.wasm.trace:this.env.trace))&&(this.device.features.has("chromium-experimental-timestamp-query-inside-passes")?this.queryType="inside-passes":this.device.features.has("timestamp-query")&&(this.queryType="at-passes"),this.queryType!=="none"&&typeof this.querySet>"u"&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:this.maxDispatchNumber*2}),this.queryResolveBuffer=this.device.createBuffer({size:this.maxDispatchNumber*2*8,usage:GPUBufferUsage.COPY_SRC|GPUBufferUsage.QUERY_RESOLVE})))}captureBegin(){rt("info","captureBegin"),this.capturedCommandList.get(this.currentSessionId)||this.capturedCommandList.set(this.currentSessionId,[]),this.capturedPendingKernels.get(this.currentSessionId)||this.capturedPendingKernels.set(this.currentSessionId,[]),this.flush(),this.sessionStatus="capturing"}captureEnd(){rt("info","captureEnd"),this.flush(),this.sessionStatus="default"}replay(){rt("info","replay"),this.sessionStatus="replaying";let t=this.capturedCommandList.get(this.currentSessionId),e=this.capturedPendingKernels.get(this.currentSessionId),r=t.length;this.pendingKernels=[];for(let n=0;n<r;n++){let a=this.getComputePassEncoder(),i=t[n];this.writeTimestamp(this.pendingDispatchNumber*2),a.setPipeline(i.computePipeline),a.setBindGroup(0,i.bindGroup),a.dispatchWorkgroups(...i.dispatchGroup),this.writeTimestamp(this.pendingDispatchNumber*2+1),this.pendingDispatchNumber++,this.queryType!=="none"&&this.pendingKernels.push(e[n]),(this.pendingDispatchNumber>=this.maxDispatchNumber||this.queryType==="at-passes")&&this.endComputePass(),this.pendingDispatchNumber>=this.maxDispatchNumber&&this.flush()}this.flush(),this.sessionStatus="default"}onReleaseSession(t){this.unregisterBuffers(t),this.capturedCommandList.has(t)&&this.capturedCommandList.delete(t),this.capturedPendingKernels.has(t)&&this.capturedPendingKernels.delete(t),this.gpuDataManager.onReleaseSession(t)}onRunStart(t){this.currentSessionId=t,this.setQueryType()}}}),Dm={};An(Dm,{init:()=>Nm});var yi,Yc,Nm,By=Z(()=>{$e(),Ry(),rn(),Ae(),yi=class Fm{constructor(e,r,n,a){this.module=e,this.dataType=r,this.data=n,this.dims=a}getFloat32Array(){if(this.dataType!==1)throw new Error("Invalid data type");let e=Y.size(this.dims);return e===0?new Float32Array:new Float32Array(this.module.HEAP8.buffer,this.data,e)}getBigInt64Array(){if(this.dataType!==7)throw new Error("Invalid data type");let e=Y.size(this.dims);return e===0?new BigInt64Array:new BigInt64Array(this.module.HEAP8.buffer,this.data,e)}getInt32Array(){if(this.dataType!==6)throw new Error("Invalid data type");let e=Y.size(this.dims);return e===0?new Int32Array:new Int32Array(this.module.HEAP8.buffer,this.data,e)}reshape(e){if(Y.size(e)!==Y.size(this.dims))throw new Error("Invalid new shape");return new Fm(this.module,this.dataType,this.data,e)}},Yc=class{constructor(t,e,r){this.module=t,this.backend=e,this.customDataOffset=0,this.customDataSize=0,this.adapterInfo=e.adapterInfo;let n=t.HEAPU32,a=r>>>2;this.opKernelContext=n[a++];let i=n[a++];this.outputCount=n[a++],this.customDataOffset=n[a++],this.customDataSize=n[a++];let s=[];for(let o=0;o<i;o++){let u=n[a++],l=n[a++],p=n[a++],f=[];for(let m=0;m<p;m++)f.push(n[a++]);s.push(new yi(t,u,l,f))}this.inputs=s}get kernelCustomData(){return this.backend.currentKernelCustomData}get customDataBuffer(){return this.module.HEAPU8.subarray(this.customDataOffset,this.customDataOffset+this.customDataSize)}getMaxComputeWorkgroupSizes(){return[this.backend.device.limits.maxComputeWorkgroupSizeX,this.backend.device.limits.maxComputeWorkgroupSizeY,this.backend.device.limits.maxComputeWorkgroupSizeZ]}getMaxComputeWorkgroupStoragesize(){return this.backend.device.limits.maxComputeWorkgroupStorageSize}compute(t,e){let r=e?.inputs?.map(s=>typeof s=="number"?this.inputs[s]:s)??this.inputs,n=e?.outputs??[],a=(s,o,u)=>new yi(this.module,o,this.output(s,u),u),i=(s,o)=>{let u=fa(s);if(!u)throw new Error(`Unsupported data type: ${s}`);let l=u*Y.size(o),p=l>0?this.backend.gpuDataManager.create(l).id:0;return new yi(this.module,s,p,o)};return this.backend.run(t,r,n,a,i,this.outputCount)}output(t,e){let r=this.module.stackSave();try{let n=this.module.stackAlloc((1+e.length)*4),a=n>>2;this.module.HEAPU32[a++]=e.length;for(let i=0;i<e.length;i++)this.module.HEAPU32[a++]=e[i];return this.module._JsepOutput(this.opKernelContext,t,n)}catch(n){throw new Error(`Failed to generate kernel's output[${t}] with dims [${e}]. If you are running with pre-allocated output, please make sure the output type/dims are correct. Error: ${n}`)}finally{this.module.stackRestore(r)}}},Nm=async(t,e,r,n)=>{let a=e.jsepInit;if(!a)throw new Error("Failed to initialize JSEP. The WebAssembly module is not built with JSEP support.");if(t==="webgpu"){let i=new Bm;await i.initialize(r,n),a("webgpu",[i,s=>i.alloc(s),s=>i.free(s),(s,o,u,l=!1)=>{if(l)rt("verbose",()=>`[WebGPU] jsepCopyGpuToGpu: src=${s}, dst=${o}, size=${u}`),i.memcpy(s,o);else{rt("verbose",()=>`[WebGPU] jsepCopyCpuToGpu: dataOffset=${s}, gpuDataId=${o}, size=${u}`);let p=e.HEAPU8.subarray(s>>>0,(s>>>0)+u);i.upload(o,p)}},async(s,o,u)=>{rt("verbose",()=>`[WebGPU] jsepCopyGpuToCpu: gpuDataId=${s}, dataOffset=${o}, size=${u}`),await i.download(s,()=>e.HEAPU8.subarray(o>>>0,(o>>>0)+u))},(s,o,u)=>i.createKernel(s,o,u,e.UTF8ToString(e._JsepGetNodeName(o))),s=>i.releaseKernel(s),(s,o,u,l)=>{rt("verbose",()=>`[WebGPU] jsepRun: sessionHandle=${u}, kernel=${s}, contextDataOffset=${o}`);let p=new Yc(e,i,o);return i.computeKernel(s,p,l)},()=>i.captureBegin(),()=>i.captureEnd(),()=>i.replay()])}else a("webnn")}}),Xc,Lm,Um,Cr,Qc,Co,Wm,Vm,Zs,Gm,Hm,jm,Dy=Z(()=>{Y_(),X_(),$e(),va(),Vo(),uh(),Xc=(t,e)=>{at()._OrtInit(t,e)!==0&&je("Can't initialize onnxruntime.")},Lm=async t=>{Xc(t.wasm.numThreads,Di(t.logLevel))},Um=async(t,e)=>{{let r=(By(),Qr(Dm)).init;if(e==="webgpu"){if(typeof navigator>"u"||!navigator.gpu)throw new Error("WebGPU is not supported in current environment");let n=t.webgpu.adapter;if(n){if(typeof n.limits!="object"||typeof n.features!="object"||typeof n.requestDevice!="function")throw new Error("Invalid GPU adapter set in `env.webgpu.adapter`. It must be a GPUAdapter object.")}else{let a=t.webgpu.powerPreference;if(a!==void 0&&a!=="low-power"&&a!=="high-performance")throw new Error(`Invalid powerPreference setting: "${a}"`);let i=t.webgpu.forceFallbackAdapter;if(i!==void 0&&typeof i!="boolean")throw new Error(`Invalid forceFallbackAdapter setting: "${i}"`);if(n=await navigator.gpu.requestAdapter({powerPreference:a,forceFallbackAdapter:i}),!n)throw new Error('Failed to get GPU adapter. You may need to enable flag "--enable-unsafe-webgpu" if you are using Chrome.')}if(!t.wasm.simd)throw new Error("Not supported for WebGPU=ON and SIMD=OFF. Please set `env.wasm.simd` to true when using `webgpu` EP");await r("webgpu",at(),t,n)}if(e==="webnn"){if(typeof navigator>"u"||!navigator.ml)throw new Error("WebNN is not supported in current environment");await r("webnn",at(),t)}}},Cr=new Map,Qc=t=>{let e=at(),r=e.stackSave();try{let n=e.stackAlloc(8);return e._OrtGetInputOutputCount(t,n,n+4)!==0&&je("Can't get session input/output count."),[e.HEAP32[n/4],e.HEAP32[n/4+1]]}finally{e.stackRestore(r)}},Co=t=>{let e=at(),r=e._malloc(t.byteLength);if(r===0)throw new Error(`Can't create a session. failed to allocate a buffer of size ${t.byteLength}.`);return e.HEAPU8.set(t,r),[r,t.byteLength]},Wm=async(t,e)=>{let r,n,a=at();Array.isArray(t)?[r,n]=t:t.buffer===a.HEAPU8.buffer?[r,n]=[t.byteOffset,t.byteLength]:[r,n]=Co(t);let i=0,s=0,o=0,u=[],l=[],p=[];try{if([s,u]=oh(e),e?.externalData&&a.mountExternalData){let $=[];for(let C of e.externalData){let T=typeof C=="string"?C:C.path;$.push(Ni(typeof C=="string"?C:C.data).then(A=>{a.mountExternalData(T,A)}))}await Promise.all($)}i=await a._OrtCreateSession(r,n,s),i===0&&je("Can't create a session.");let[f,m]=Qc(i),c=!!e?.enableGraphCapture,y=[],w=[],v=[];for(let $=0;$<f;$++){let C=a._OrtGetInputName(i,$);C===0&&je("Can't get an input name."),l.push(C),y.push(a.UTF8ToString(C))}for(let $=0;$<m;$++){let C=a._OrtGetOutputName(i,$);C===0&&je("Can't get an output name."),p.push(C);let T=a.UTF8ToString(C);w.push(T);{if(c&&e?.preferredOutputLocation===void 0){v.push("gpu-buffer");continue}let A=typeof e?.preferredOutputLocation=="string"?e.preferredOutputLocation:e?.preferredOutputLocation?.[T]??"cpu";if(A!=="cpu"&&A!=="cpu-pinned"&&A!=="gpu-buffer")throw new Error(`Not supported preferred output location: ${A}.`);if(c&&A!=="gpu-buffer")throw new Error(`Not supported preferred output location: ${A}. Only 'gpu-buffer' location is supported when enableGraphCapture is true.`);v.push(A)}}let k=null;return v.some($=>$==="gpu-buffer")&&(o=a._OrtCreateBinding(i),o===0&&je("Can't create IO binding."),k={handle:o,outputPreferredLocations:v,outputPreferredLocationsEncoded:v.map($=>go($))}),Cr.set(i,[i,l,p,k,c,!1]),[i,y,w]}catch(f){throw l.forEach(m=>a._OrtFree(m)),p.forEach(m=>a._OrtFree(m)),o!==0&&a._OrtReleaseBinding(o),i!==0&&a._OrtReleaseSession(i),f}finally{a._free(r),s!==0&&a._OrtReleaseSessionOptions(s),u.forEach(f=>a._free(f)),a.unmountExternalData?.()}},Vm=t=>{let e=at(),r=Cr.get(t);if(!r)throw new Error(`cannot release session. invalid session id: ${t}`);let[n,a,i,s,o]=r;s&&(o&&e._OrtClearBoundOutputs(s.handle),e._OrtReleaseBinding(s.handle)),e.jsepOnReleaseSession?.(t),a.forEach(u=>e._OrtFree(u)),i.forEach(u=>e._OrtFree(u)),e._OrtReleaseSession(n),Cr.delete(t)},Zs=(t,e,r,n,a,i=!1)=>{if(!t){e.push(0);return}let s=at(),o=t[0],u=t[1],l=t[3],p,f;if(o==="string"&&l==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");if(i&&l!=="gpu-buffer")throw new Error(`External buffer must be provided for input/output index ${a} when enableGraphCapture is true.`);if(l==="gpu-buffer"){let y=t[2].gpuBuffer,w=fa(mo(o));f=u.reduce((k,$)=>k*$,1)*w;let v=s.jsepRegisterBuffer;if(!v)throw new Error('Tensor location "gpu-buffer" is not supported without using WebGPU.');p=v(n,a,y,f)}else{let y=t[2];if(Array.isArray(y)){f=4*y.length,p=s._malloc(f),r.push(p);let w=p/4;for(let v=0;v<y.length;v++){if(typeof y[v]!="string")throw new TypeError(`tensor data at index ${v} is not a string`);s.HEAPU32[w++]=tt(y[v],r)}}else f=y.byteLength,p=s._malloc(f),r.push(p),s.HEAPU8.set(new Uint8Array(y.buffer,y.byteOffset,f),p)}let m=s.stackSave(),c=s.stackAlloc(4*u.length);try{let y=c/4;u.forEach(v=>s.HEAP32[y++]=v);let w=s._OrtCreateTensor(mo(o),p,f,c,u.length,go(l));w===0&&je(`Can't create tensor for input/output. session=${n}, index=${a}.`),e.push(w)}finally{s.stackRestore(m)}},Gm=async(t,e,r,n,a,i)=>{let s=at(),o=Cr.get(t);if(!o)throw new Error(`cannot run inference. invalid session id: ${t}`);let u=o[0],l=o[1],p=o[2],f=o[3],m=o[4],c=o[5],y=e.length,w=n.length,v=0,k=[],$=[],C=[],T=[],A=s.stackSave(),B=s.stackAlloc(y*4),R=s.stackAlloc(y*4),D=s.stackAlloc(w*4),K=s.stackAlloc(w*4);try{[v,k]=sh(i);for(let P=0;P<y;P++)Zs(r[P],$,T,t,e[P],m);for(let P=0;P<w;P++)Zs(a[P],C,T,t,y+n[P],m);let j=B/4,ie=R/4,te=D/4,oe=K/4;for(let P=0;P<y;P++)s.HEAPU32[j++]=$[P],s.HEAPU32[ie++]=l[e[P]];for(let P=0;P<w;P++)s.HEAPU32[te++]=C[P],s.HEAPU32[oe++]=p[n[P]];if(f&&!c){let{handle:P,outputPreferredLocations:H,outputPreferredLocationsEncoded:le}=f;if(l.length!==y)throw new Error(`input count from feeds (${y}) is expected to be always equal to model's input count (${l.length}).`);for(let G=0;G<y;G++){let ne=e[G];await s._OrtBindInput(P,l[ne],$[G])!==0&&je(`Can't bind input[${G}] for session=${t}.`)}for(let G=0;G<w;G++){let ne=n[G];a[G]?.[3]?s._OrtBindOutput(P,p[ne],C[G],0)!==0&&je(`Can't bind pre-allocated output[${G}] for session=${t}.`):s._OrtBindOutput(P,p[ne],0,le[ne])!==0&&je(`Can't bind output[${G}] to ${H[G]} for session=${t}.`)}Cr.set(t,[u,l,p,f,m,!0])}s.jsepOnRunStart?.(u);let re;f?re=await s._OrtRunWithBinding(u,f.handle,w,D,v):re=await s._OrtRun(u,R,B,y,K,w,D,v),re!==0&&je("failed to call OrtRun().");let M=[];for(let P=0;P<w;P++){let H=s.HEAPU32[D/4+P];if(H===C[P]){M.push(a[P]);continue}let le=s.stackSave(),G=s.stackAlloc(4*4),ne=!1,N,ae=0;try{s._OrtGetTensorData(H,G,G+4,G+8,G+12)!==0&&je(`Can't access output tensor data on index ${P}.`);let fe=G/4,Ce=s.HEAPU32[fe++];ae=s.HEAPU32[fe++];let Be=s.HEAPU32[fe++],Ke=s.HEAPU32[fe++],Qe=[];for(let Ve=0;Ve<Ke;Ve++)Qe.push(s.HEAPU32[Be/4+Ve]);s._OrtFree(Be);let We=Qe.reduce((Ve,Ne)=>Ve*Ne,1);N=qr(Ce);let dt=f?.outputPreferredLocations[n[P]];if(N==="string"){if(dt==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");let Ve=[],Ne=ae/4;for(let ht=0;ht<We;ht++){let Vt=s.HEAPU32[Ne++],st=ht===We-1?void 0:s.HEAPU32[Ne]-Vt;Ve.push(s.UTF8ToString(Vt,st))}M.push([N,Qe,Ve,"cpu"])}else if(dt==="gpu-buffer"&&We>0){let Ve=s.jsepGetBuffer;if(!Ve)throw new Error('preferredLocation "gpu-buffer" is not supported without using WebGPU.');let Ne=Ve(ae),ht=fa(Ce);if(ht===void 0||!Ho(N))throw new Error(`Unsupported data type: ${N}`);ne=!0,M.push([N,Qe,{gpuBuffer:Ne,download:s.jsepCreateDownloader(Ne,We*ht,N),dispose:()=>{s._OrtReleaseTensor(H)}},"gpu-buffer"])}else{let Ve=Go(N),Ne=new Ve(We);new Uint8Array(Ne.buffer,Ne.byteOffset,Ne.byteLength).set(s.HEAPU8.subarray(ae,ae+Ne.byteLength)),M.push([N,Qe,Ne,"cpu"])}}finally{s.stackRestore(le),N==="string"&&ae&&s._free(ae),ne||s._OrtReleaseTensor(H)}}return f&&!m&&(s._OrtClearBoundOutputs(f.handle),Cr.set(t,[u,l,p,f,m,!1])),M}finally{s.stackRestore(A),$.forEach(j=>s._OrtReleaseTensor(j)),C.forEach(j=>s._OrtReleaseTensor(j)),T.forEach(j=>s._free(j)),v!==0&&s._OrtReleaseRunOptions(v),k.forEach(j=>s._free(j))}},Hm=t=>{let e=at(),r=Cr.get(t);if(!r)throw new Error("invalid session id");let n=r[0],a=e._OrtEndProfiling(n);a===0&&je("Can't get an profile file name."),e._OrtFree(a)},jm=t=>{let e=[];for(let r of t){let n=r[2];!Array.isArray(n)&&"buffer"in n&&e.push(n.buffer)}return e}}),Ny=In((t,e)=>{e.exports='/*!\n * ONNX Runtime Web v1.18.0\n * Copyright (c) Microsoft Corporation. All rights reserved.\n * Licensed under the MIT License.\n */\n"use strict";(()=>{var ao=Object.defineProperty;var nl=Object.getOwnPropertyDescriptor;var ol=Object.getOwnPropertyNames;var il=Object.prototype.hasOwnProperty;var Y=(e,t)=>()=>(e&&(t=e(e=0)),t);var Wr=(e,t)=>()=>(t||e((t={exports:{}}).exports,t),t.exports),gn=(e,t)=>{for(var r in t)ao(e,r,{get:t[r],enumerable:!0})},al=(e,t,r,o)=>{if(t&&typeof t=="object"||typeof t=="function")for(let i of ol(t))!il.call(e,i)&&i!==r&&ao(e,i,{get:()=>t[i],enumerable:!(o=nl(t,i))||o.enumerable});return e};var wr=e=>al(ao({},"__esModule",{value:!0}),e);var so={};gn(so,{createReadStream:()=>Ai,readFile:()=>sl,readFileSync:()=>ul});var sl,ul,Ai,uo=Y(()=>{sl=void 0,ul=void 0,Ai=void 0});var lo={};gn(lo,{join:()=>dl});var dl,co=Y(()=>{dl=void 0});var Ei=Wr((Ti,po)=>{"use strict";var Ii=(()=>{var e=typeof document<"u"?document.currentScript?.src:void 0;return typeof __filename<"u"&&(e||=__filename),function(t={}){var r=t,o,i,u=new Promise((s,m)=>{o=s,i=m});r.mountExternalData=(s,m)=>{(r.eb||(r.eb=new Map)).set(s,m)},r.unmountExternalData=()=>{delete r.eb};let a=()=>{let s=(g,$,T)=>(...B)=>{let H=dt,q=$?.();B=g(...B);let te=$?.();return q!==te&&(g=te,T(q),$=T=null),dt!=H?tn():B},m=g=>async(...$)=>{try{if(r.cb)throw Error("Session already started");let T=r.cb={xb:$[0],errors:[]},B=await g(...$);if(r.cb!==T)throw Error("Session mismatch");r.kb?.flush();let H=T.errors;if(0<H.length){let q=await Promise.all(H);if(q=q.filter(te=>te),0<q.length)throw Error(q.join(`\n`))}return B}finally{r.cb=null}};r._OrtCreateSession=s(r._OrtCreateSession,()=>r._OrtCreateSession,g=>r._OrtCreateSession=g),r._OrtRun=m(s(r._OrtRun,()=>r._OrtRun,g=>r._OrtRun=g)),r._OrtRunWithBinding=m(s(r._OrtRunWithBinding,()=>r._OrtRunWithBinding,g=>r._OrtRunWithBinding=g)),r._OrtBindInput=s(r._OrtBindInput,()=>r._OrtBindInput,g=>r._OrtBindInput=g),a=void 0};r.jsepInit=(s,m)=>{if(a?.(),s==="webgpu"){[r.kb,r.pb,r.tb,r.lb,r.sb,r.Ra,r.ub,r.wb,r.qb,r.rb,r.vb]=m;let g=r.kb;r.jsepRegisterBuffer=($,T,B,H)=>g.registerBuffer($,T,B,H),r.jsepGetBuffer=$=>g.getBuffer($),r.jsepCreateDownloader=($,T,B)=>g.createDownloader($,T,B),r.jsepOnReleaseSession=$=>{g.onReleaseSession($)},r.jsepOnRunStart=$=>g.onRunStart($)}};var c=Object.assign({},r),p="./this.program",h=(s,m)=>{throw m},d=typeof window=="object",y=typeof importScripts=="function",w=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",_="",v,S,A;if(w){var I=(uo(),wr(so)),x=(co(),wr(lo));_=y?x.dirname(_)+"/":__dirname+"/",v=(s,m)=>(s=Ke(s)?new URL(s):x.normalize(s),I.readFileSync(s,m?void 0:"utf8")),A=s=>(s=v(s,!0),s.buffer||(s=new Uint8Array(s)),s),S=(s,m,g,$=!0)=>{s=Ke(s)?new URL(s):x.normalize(s),I.readFile(s,$?void 0:"utf8",(T,B)=>{T?g(T):m($?B.buffer:B)})},!r.thisProgram&&1<process.argv.length&&(p=process.argv[1].replace(/\\\\/g,"/")),process.argv.slice(2),h=(s,m)=>{throw process.exitCode=s,m}}else(d||y)&&(y?_=self.location.href:typeof document<"u"&&document.currentScript&&(_=document.currentScript.src),e&&(_=e),_.startsWith("blob:")?_="":_=_.substr(0,_.replace(/[?#].*/,"").lastIndexOf("/")+1),v=s=>{var m=new XMLHttpRequest;return m.open("GET",s,!1),m.send(null),m.responseText},y&&(A=s=>{var m=new XMLHttpRequest;return m.open("GET",s,!1),m.responseType="arraybuffer",m.send(null),new Uint8Array(m.response)}),S=(s,m,g)=>{var $=new XMLHttpRequest;$.open("GET",s,!0),$.responseType="arraybuffer",$.onload=()=>{$.status==200||$.status==0&&$.response?m($.response):g()},$.onerror=g,$.send(null)});var E=console.log.bind(console),P=console.error.bind(console);Object.assign(r,c),c=null;var O,R=!1,L,N,K,Q,he,W,se,Ce,We,ee,ae;function Ae(){var s=O.buffer;r.HEAP8=N=new Int8Array(s),r.HEAP16=Q=new Int16Array(s),r.HEAPU8=K=new Uint8Array(s),r.HEAPU16=he=new Uint16Array(s),r.HEAP32=W=new Int32Array(s),r.HEAPU32=se=new Uint32Array(s),r.HEAPF32=Ce=new Float32Array(s),r.HEAPF64=ae=new Float64Array(s),r.HEAP64=We=new BigInt64Array(s),r.HEAPU64=ee=new BigUint64Array(s)}var me=[],ie=[],ue=[],le=0,qe=null,G=null;function ne(s){throw s="Aborted("+s+")",P(s),R=!0,L=1,s=new WebAssembly.RuntimeError(s+". Build with -sASSERTIONS for more info."),i(s),s}var xe=s=>s.startsWith("data:application/octet-stream;base64,"),Ke=s=>s.startsWith("file://"),Be;if(Be="ort-wasm-simd.wasm",!xe(Be)){var Ge=Be;Be=r.locateFile?r.locateFile(Ge,_):_+Ge}function Ut(s){if(A)return A(s);throw"both async and sync fetching of the wasm failed"}function Ne(s){if(d||y){if(typeof fetch=="function"&&!Ke(s))return fetch(s,{credentials:"same-origin"}).then(m=>{if(!m.ok)throw`failed to load wasm binary file at \'${s}\'`;return m.arrayBuffer()}).catch(()=>Ut(s));if(S)return new Promise((m,g)=>{S(s,$=>m(new Uint8Array($)),g)})}return Promise.resolve().then(()=>Ut(s))}function Ye(s,m,g){return Ne(s).then($=>WebAssembly.instantiate($,m)).then(g,$=>{P(`failed to asynchronously prepare wasm: ${$}`),ne($)})}function mt(s,m){var g=Be;return typeof WebAssembly.instantiateStreaming!="function"||xe(g)||Ke(g)||w||typeof fetch!="function"?Ye(g,s,m):fetch(g,{credentials:"same-origin"}).then($=>WebAssembly.instantiateStreaming($,s).then(m,function(T){return P(`wasm streaming compile failed: ${T}`),P("falling back to ArrayBuffer instantiation"),Ye(g,s,m)}))}var Rt={824920:(s,m,g,$)=>{if(typeof r>"u"||!r.eb)return 1;if(s=je(s>>>0),s.startsWith("./")&&(s=s.substring(2)),s=r.eb.get(s),!s)return 2;if(m>>>=0,g>>>=0,m+g>s.byteLength)return 3;try{return K.set(s.subarray(m,m+g),$>>>0>>>0),0}catch{return 4}},825421:()=>{r.qb()},825452:()=>{r.rb()},825481:()=>{r.vb()},825506:s=>r.pb(s),825539:s=>r.tb(s),825571:(s,m,g)=>{r.lb(s,m,g,!0)},825610:(s,m,g)=>{r.lb(s,m,g)},825643:s=>{r.Ra("Abs",s,void 0)},825694:s=>{r.Ra("Neg",s,void 0)},825745:s=>{r.Ra("Floor",s,void 0)},825798:s=>{r.Ra("Ceil",s,void 0)},825850:s=>{r.Ra("Reciprocal",s,void 0)},825908:s=>{r.Ra("Sqrt",s,void 0)},825960:s=>{r.Ra("Exp",s,void 0)},826011:s=>{r.Ra("Erf",s,void 0)},826062:s=>{r.Ra("Sigmoid",s,void 0)},826117:(s,m,g)=>{r.Ra("HardSigmoid",s,{alpha:m,beta:g})},826196:s=>{r.Ra("Log",s,void 0)},826247:s=>{r.Ra("Sin",s,void 0)},826298:s=>{r.Ra("Cos",s,void 0)},826349:s=>{r.Ra("Tan",s,void 0)},826400:s=>{r.Ra("Asin",s,void 0)},826452:s=>{r.Ra("Acos",s,void 0)},826504:s=>{r.Ra("Atan",s,void 0)},826556:s=>{r.Ra("Sinh",s,void 0)},826608:s=>{r.Ra("Cosh",s,void 0)},826660:s=>{r.Ra("Asinh",s,void 0)},826713:s=>{r.Ra("Acosh",s,void 0)},826766:s=>{r.Ra("Atanh",s,void 0)},826819:s=>{r.Ra("Tanh",s,void 0)},826871:s=>{r.Ra("Not",s,void 0)},826922:(s,m,g)=>{r.Ra("Clip",s,{min:m,max:g})},826991:s=>{r.Ra("Clip",s,void 0)},827043:(s,m)=>{r.Ra("Elu",s,{alpha:m})},827101:s=>{r.Ra("Relu",s,void 0)},827153:(s,m)=>{r.Ra("LeakyRelu",s,{alpha:m})},827217:(s,m)=>{r.Ra("ThresholdedRelu",s,{alpha:m})},827287:(s,m)=>{r.Ra("Cast",s,{to:m})},827345:s=>{r.Ra("Add",s,void 0)},827396:s=>{r.Ra("Sub",s,void 0)},827447:s=>{r.Ra("Mul",s,void 0)},827498:s=>{r.Ra("Div",s,void 0)},827549:s=>{r.Ra("Pow",s,void 0)},827600:s=>{r.Ra("Equal",s,void 0)},827653:s=>{r.Ra("Greater",s,void 0)},827708:s=>{r.Ra("GreaterOrEqual",s,void 0)},827770:s=>{r.Ra("Less",s,void 0)},827822:s=>{r.Ra("LessOrEqual",s,void 0)},827881:(s,m,g,$,T)=>{r.Ra("ReduceMean",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828040:(s,m,g,$,T)=>{r.Ra("ReduceMax",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828198:(s,m,g,$,T)=>{r.Ra("ReduceMin",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828356:(s,m,g,$,T)=>{r.Ra("ReduceProd",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828515:(s,m,g,$,T)=>{r.Ra("ReduceSum",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828673:(s,m,g,$,T)=>{r.Ra("ReduceL1",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828830:(s,m,g,$,T)=>{r.Ra("ReduceL2",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828987:(s,m,g,$,T)=>{r.Ra("ReduceLogSum",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},829148:(s,m,g,$,T)=>{r.Ra("ReduceSumSquare",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},829312:(s,m,g,$,T)=>{r.Ra("ReduceLogSumExp",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},829476:s=>{r.Ra("Where",s,void 0)},829529:(s,m,g)=>{r.Ra("Transpose",s,{perm:m?Array.from(W.subarray(m>>>0,g>>>0)):[]})},829637:(s,m,g,$)=>{r.Ra("DepthToSpace",s,{blocksize:m,mode:je(g),format:$?"NHWC":"NCHW"})},829770:(s,m,g,$)=>{r.Ra("DepthToSpace",s,{blocksize:m,mode:je(g),format:$?"NHWC":"NCHW"})},829903:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be)=>{r.Ra("ConvTranspose",s,{format:te?"NHWC":"NCHW",autoPad:m,dilations:[g],group:$,kernelShape:[T],pads:[B,H],strides:[q],wIsConst:()=>!!N[X>>>0],outputPadding:de?Array.from(W.subarray(de>>>0,Ee>>>0)):[],outputShape:Oe?Array.from(W.subarray(Oe>>>0,D>>>0)):[],activation:je(be)})},830304:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D)=>{r.Ra("ConvTranspose",s,{format:q?"NHWC":"NCHW",autoPad:m,dilations:Array.from(W.subarray(g>>>0,(g>>>0)+2>>>0)),group:$,kernelShape:Array.from(W.subarray(T>>>0,(T>>>0)+2>>>0)),pads:Array.from(W.subarray(B>>>0,(B>>>0)+4>>>0)),strides:Array.from(W.subarray(H>>>0,(H>>>0)+2>>>0)),wIsConst:()=>!!N[te>>>0],outputPadding:X?Array.from(W.subarray(X>>>0,de>>>0)):[],outputShape:Ee?Array.from(W.subarray(Ee>>>0,Oe>>>0)):[],activation:je(D)})},830869:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be)=>{r.Ra("ConvTranspose",s,{format:te?"NHWC":"NCHW",autoPad:m,dilations:[g],group:$,kernelShape:[T],pads:[B,H],strides:[q],wIsConst:()=>!!N[X>>>0],outputPadding:de?Array.from(W.subarray(de>>>0,Ee>>>0)):[],outputShape:Oe?Array.from(W.subarray(Oe>>>0,D>>>0)):[],activation:je(be)})},831270:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D)=>{r.Ra("ConvTranspose",s,{format:q?"NHWC":"NCHW",autoPad:m,dilations:Array.from(W.subarray(g>>>0,(g>>>0)+2>>>0)),group:$,kernelShape:Array.from(W.subarray(T>>>0,(T>>>0)+2>>>0)),pads:Array.from(W.subarray(B>>>0,(B>>>0)+4>>>0)),strides:Array.from(W.subarray(H>>>0,(H>>>0)+2>>>0)),wIsConst:()=>!!N[te>>>0],outputPadding:X?Array.from(W.subarray(X>>>0,de>>>0)):[],outputShape:Ee?Array.from(W.subarray(Ee>>>0,Oe>>>0)):[],activation:je(D)})},831835:(s,m)=>{r.Ra("GlobalAveragePool",s,{format:m?"NHWC":"NCHW"})},831926:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be,Ie)=>{r.Ra("AveragePool",s,{format:Ie?"NHWC":"NCHW",auto_pad:m,ceil_mode:g,count_include_pad:$,storage_order:T,dilations:[B,H],kernel_shape:[q,te],pads:[X,de,Ee,Oe],strides:[D,be]})},832210:(s,m)=>{r.Ra("GlobalAveragePool",s,{format:m?"NHWC":"NCHW"})},832301:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be,Ie)=>{r.Ra("AveragePool",s,{format:Ie?"NHWC":"NCHW",auto_pad:m,ceil_mode:g,count_include_pad:$,storage_order:T,dilations:[B,H],kernel_shape:[q,te],pads:[X,de,Ee,Oe],strides:[D,be]})},832585:(s,m)=>{r.Ra("GlobalMaxPool",s,{format:m?"NHWC":"NCHW"})},832672:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be,Ie)=>{r.Ra("MaxPool",s,{format:Ie?"NHWC":"NCHW",auto_pad:m,ceil_mode:g,count_include_pad:$,storage_order:T,dilations:[B,H],kernel_shape:[q,te],pads:[X,de,Ee,Oe],strides:[D,be]})},832952:(s,m)=>{r.Ra("GlobalMaxPool",s,{format:m?"NHWC":"NCHW"})},833039:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be,Ie)=>{r.Ra("MaxPool",s,{format:Ie?"NHWC":"NCHW",auto_pad:m,ceil_mode:g,count_include_pad:$,storage_order:T,dilations:[B,H],kernel_shape:[q,te],pads:[X,de,Ee,Oe],strides:[D,be]})},833319:(s,m,g,$,T)=>{r.Ra("Gemm",s,{alpha:m,beta:g,transA:$,transB:T})},833423:s=>{r.Ra("MatMul",s,void 0)},833477:(s,m,g,$)=>{r.Ra("ArgMax",s,{keepDims:!!m,selectLastIndex:!!g,axis:$})},833585:(s,m,g,$)=>{r.Ra("ArgMin",s,{keepDims:!!m,selectLastIndex:!!g,axis:$})},833693:(s,m)=>{r.Ra("Softmax",s,{axis:m})},833756:(s,m)=>{r.Ra("Concat",s,{axis:m})},833816:(s,m,g,$,T)=>{r.Ra("Split",s,{axis:m,numOutputs:g,splitSizes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},833956:s=>{r.Ra("Expand",s,void 0)},834010:(s,m)=>{r.Ra("Gather",s,{axis:Number(m)})},834081:(s,m)=>{r.Ra("GatherElements",s,{axis:Number(m)})},834160:(s,m,g,$,T,B,H,q,te,X,de)=>{r.Ra("Resize",s,{antialias:m,axes:g?Array.from(W.subarray(g>>>0,$>>>0)):[],coordinateTransformMode:je(T),cubicCoeffA:B,excludeOutside:H,extrapolationValue:q,keepAspectRatioPolicy:je(te),mode:je(X),nearestMode:je(de)})},834506:(s,m,g,$,T,B,H)=>{r.Ra("Slice",s,{starts:m?Array.from(W.subarray(m>>>0,g>>>0)):[],ends:$?Array.from(W.subarray($>>>0,T>>>0)):[],axes:B?Array.from(W.subarray(B>>>0,H>>>0)):[]})},834722:s=>{r.Ra("Tile",s,void 0)},834774:(s,m,g,$)=>{r.Ra("LayerNormalization",s,{axis:m,epsilon:g,simplified:!!$})},834885:(s,m,g)=>{r.Ra("InstanceNormalization",s,{epsilon:m,format:g?"NHWC":"NCHW"})},834999:(s,m,g)=>{r.Ra("InstanceNormalization",s,{epsilon:m,format:g?"NHWC":"NCHW"})},835113:s=>{r.Ra("Range",s,void 0)},835166:(s,m)=>{r.Ra("Einsum",s,{equation:je(m)})},835247:(s,m,g,$,T)=>{r.Ra("Pad",s,{mode:m,value:g,pads:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},835374:(s,m,g,$,T,B)=>{r.Ra("BatchNormalization",s,{epsilon:m,momentum:g,spatial:!!T,trainingMode:!!$,format:B?"NHWC":"NCHW"})},835543:(s,m,g,$,T,B)=>{r.Ra("BatchNormalization",s,{epsilon:m,momentum:g,spatial:!!T,trainingMode:!!$,format:B?"NHWC":"NCHW"})},835712:(s,m,g)=>{r.Ra("CumSum",s,{exclusive:Number(m),reverse:Number(g)})},835809:(s,m,g,$,T,B,H,q,te)=>{r.Ra("Attention",s,{numHeads:m,isUnidirectional:g,maskFilterValue:$,scale:T,doRotary:B,qkvHiddenSizes:H?Array.from(W.subarray(Number(q)>>>0,Number(q)+H>>>0)):[],pastPresentShareBuffer:!!te})},836081:s=>{r.Ra("BiasAdd",s,void 0)},836136:s=>{r.Ra("BiasSplitGelu",s,void 0)},836197:s=>{r.Ra("FastGelu",s,void 0)},836253:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe)=>{r.Ra("Conv",s,{format:te?"NHWC":"NCHW",auto_pad:m,dilations:[g],group:$,kernel_shape:[T],pads:B?Array.from(W.subarray(B>>>0,H>>>0)):[],strides:[q],w_is_const:()=>!!N[X>>>0],activation:je(de),activation_params:Ee?Array.from(Ce.subarray(Ee>>>0,Oe>>>0)):[]})},836623:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be,Ie)=>{r.Ra("Conv",s,{format:Ee?"NHWC":"NCHW",auto_pad:m,dilations:[g,$],group:T,kernel_shape:[B,H],pads:q?Array.from(W.subarray(q>>>0,te>>>0)):[],strides:[X,de],w_is_const:()=>!!N[Oe>>>0],activation:je(D),activation_params:be?Array.from(Ce.subarray(be>>>0,Ie>>>0)):[]})},837014:s=>{r.Ra("Gelu",s,void 0)},837066:(s,m,g,$,T,B)=>{r.Ra("MatMulNBits",s,{k:m,n:g,accuracyLevel:$,bits:T,blockSize:B})},837193:(s,m,g,$,T,B)=>{r.Ra("MultiHeadAttention",s,{numHeads:m,isUnidirectional:g,maskFilterValue:$,scale:T,doRotary:B})},837352:(s,m,g,$,T)=>{r.Ra("RotaryEmbedding",s,{interleaved:!!m,numHeads:g,rotaryEmbeddingDim:$,scale:T})},837491:(s,m,g)=>{r.Ra("SkipLayerNormalization",s,{epsilon:m,simplified:!!g})},837593:(s,m,g)=>{r.Ra("SkipLayerNormalization",s,{epsilon:m,simplified:!!g})},837695:(s,m,g,$)=>{r.Ra("LayerNormalization",s,{axis:m,epsilon:g,simplified:!!$})},837806:s=>{r.ub(s)},837840:(s,m)=>r.wb(s,m,r.cb.xb,r.cb.errors)};function qt(s){this.name="ExitStatus",this.message=`Program terminated with exit(${s})`,this.status=s}class Vt{constructor(m){this.hb=m-24}}var sr=0,jt=0,xr=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,Kt=(s,m,g)=>{m>>>=0;var $=m+g;for(g=m;s[g]&&!(g>=$);)++g;if(16<g-m&&s.buffer&&xr)return xr.decode(s.subarray(m,g));for($="";m<g;){var T=s[m++];if(T&128){var B=s[m++]&63;if((T&224)==192)$+=String.fromCharCode((T&31)<<6|B);else{var H=s[m++]&63;T=(T&240)==224?(T&15)<<12|B<<6|H:(T&7)<<18|B<<12|H<<6|s[m++]&63,65536>T?$+=String.fromCharCode(T):(T-=65536,$+=String.fromCharCode(55296|T>>10,56320|T&1023))}}else $+=String.fromCharCode(T)}return $},je=(s,m)=>(s>>>=0)?Kt(K,s,m):"",Wt=s=>{for(var m=0,g=0;g<s.length;++g){var $=s.charCodeAt(g);127>=$?m++:2047>=$?m+=2:55296<=$&&57343>=$?(m+=4,++g):m+=3}return m},at=(s,m,g,$)=>{if(g>>>=0,!(0<$))return 0;var T=g;$=g+$-1;for(var B=0;B<s.length;++B){var H=s.charCodeAt(B);if(55296<=H&&57343>=H){var q=s.charCodeAt(++B);H=65536+((H&1023)<<10)|q&1023}if(127>=H){if(g>=$)break;m[g++>>>0]=H}else{if(2047>=H){if(g+1>=$)break;m[g++>>>0]=192|H>>6}else{if(65535>=H){if(g+2>=$)break;m[g++>>>0]=224|H>>12}else{if(g+3>=$)break;m[g++>>>0]=240|H>>18,m[g++>>>0]=128|H>>12&63}m[g++>>>0]=128|H>>6&63}m[g++>>>0]=128|H&63}}return m[g>>>0]=0,g-T},Cr,ft=s=>{for(var m="";K[s>>>0];)m+=Cr[K[s++>>>0]];return m},ur={},Ar={},Ir={},st;function Tr(s,m,g={}){var $=m.name;if(!s)throw new st(`type "${$}" must have a positive integer typeid pointer`);if(Ar.hasOwnProperty(s)){if(g.nb)return;throw new st(`Cannot register type \'${$}\' twice`)}Ar[s]=m,delete Ir[s],ur.hasOwnProperty(s)&&(m=ur[s],delete ur[s],m.forEach(T=>T()))}function ht(s,m,g={}){if(!("argPackAdvance"in m))throw new TypeError("registerType registeredInstance requires argPackAdvance");return Tr(s,m,g)}var bt=(s,m,g)=>{switch(m){case 1:return g?$=>N[$>>>0]:$=>K[$>>>0];case 2:return g?$=>Q[$>>>1>>>0]:$=>he[$>>>1>>>0];case 4:return g?$=>W[$>>>2>>>0]:$=>se[$>>>2>>>0];case 8:return g?$=>We[$>>>3]:$=>ee[$>>>3];default:throw new TypeError(`invalid integer width (${m}): ${s}`)}},Er=[],Te=[];function dr(s){s>>>=0,9<s&&--Te[s+1]===0&&(Te[s]=void 0,Er.push(s))}var Xe=s=>{if(!s)throw new st("Cannot use deleted val. handle = "+s);return Te[s]},ut=s=>{switch(s){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let m=Er.pop()||Te.length;return Te[m]=s,Te[m+1]=1,m}};function lr(s){return this.fromWireType(se[s>>>2>>>0])}var Fn={name:"emscripten::val",fromWireType:s=>{var m=Xe(s);return dr(s),m},toWireType:(s,m)=>ut(m),argPackAdvance:8,readValueFromPointer:lr,bb:null},ge=(s,m)=>{switch(m){case 4:return function(g){return this.fromWireType(Ce[g>>>2>>>0])};case 8:return function(g){return this.fromWireType(ae[g>>>3>>>0])};default:throw new TypeError(`invalid float width (${m}): ${s}`)}},Yt=typeof TextDecoder<"u"?new TextDecoder("utf-16le"):void 0,jr=(s,m)=>{for(var g=s>>1,$=g+m/2;!(g>=$)&&he[g>>>0];)++g;if(g<<=1,32<g-s&&Yt)return Yt.decode(K.subarray(s>>>0,g>>>0));for(g="",$=0;!($>=m/2);++$){var T=Q[s+2*$>>>1>>>0];if(T==0)break;g+=String.fromCharCode(T)}return g},qn=(s,m,g)=>{if(g??=2147483647,2>g)return 0;g-=2;var $=m;g=g<2*s.length?g/2:s.length;for(var T=0;T<g;++T)Q[m>>>1>>>0]=s.charCodeAt(T),m+=2;return Q[m>>>1>>>0]=0,m-$},Kr=s=>2*s.length,jn=(s,m)=>{for(var g=0,$="";!(g>=m/4);){var T=W[s+4*g>>>2>>>0];if(T==0)break;++g,65536<=T?(T-=65536,$+=String.fromCharCode(55296|T>>10,56320|T&1023)):$+=String.fromCharCode(T)}return $},Yr=(s,m,g)=>{if(m>>>=0,g??=2147483647,4>g)return 0;var $=m;g=$+g-4;for(var T=0;T<s.length;++T){var B=s.charCodeAt(T);if(55296<=B&&57343>=B){var H=s.charCodeAt(++T);B=65536+((B&1023)<<10)|H&1023}if(W[m>>>2>>>0]=B,m+=4,m+4>g)break}return W[m>>>2>>>0]=0,m-$},Zr=s=>{for(var m=0,g=0;g<s.length;++g){var $=s.charCodeAt(g);55296<=$&&57343>=$&&++g,m+=4}return m},cr=(s,m)=>{var g=Ar[s];if(g===void 0)throw s=Mr(s),g=ft(s),rt(s),new st(`${m} has unknown type ${g}`);return g},Pr=(s,m,g)=>{var $=[];return s=s.toWireType($,g),$.length&&(se[m>>>2>>>0]=ut($)),s},He=s=>{try{s()}catch(m){ne(m)}},Xr=s=>{if(!R)try{s();try{L=L=s=L,r.onExit?.(s),R=!0,h(s,new qt(s))}catch(m){m instanceof qt||m=="unwind"||h(1,m)}}catch(m){m instanceof qt||m=="unwind"||h(1,m)}};function Qr(){var s=pe,m={};for(let[g,$]of Object.entries(s))m[g]=typeof $=="function"?(...T)=>{Zt.push(g);try{return $(...T)}finally{R||(Zt.pop(),dt&&gt===1&&Zt.length===0&&(gt=0,He(dn),typeof Fibers<"u"&&Fibers.Db()))}}:$;return m}var gt=0,dt=null,Bt=0,Zt=[],kr={},Or={},Jr=0,pr=null,en=[];function tn(){return new Promise((s,m)=>{pr={resolve:s,reject:m}})}function rn(){var s=Nt(65548),m=s+12;se[s>>>2>>>0]=m,se[s+4>>>2>>>0]=m+65536,m=Zt[0];var g=kr[m];return g===void 0&&(g=Jr++,kr[m]=g,Or[g]=m),W[s+8>>>2>>>0]=g,s}function nn(s){if(!R){if(gt===0){var m=!1,g=!1;s(($=0)=>{if(!R&&(Bt=$,m=!0,g)){gt=2,He(()=>vt(dt)),typeof Browser<"u"&&Browser.ib.mb&&Browser.ib.resume(),$=!1;try{var T=(0,pe[Or[W[dt+8>>>2>>>0]]])()}catch(q){T=q,$=!0}var B=!1;if(!dt){var H=pr;H&&(pr=null,($?H.reject:H.resolve)(T),B=!0)}if($&&!B)throw T}}),g=!0,m||(gt=1,dt=rn(),typeof Browser<"u"&&Browser.ib.mb&&Browser.ib.pause(),He(()=>Jt(dt)))}else gt===2?(gt=0,He(ct),rt(dt),dt=null,en.forEach(Xr)):ne(`invalid state: ${gt}`);return Bt}}function Rr(s){return nn(m=>{s().then(m)})}var Xt=[],on={},nt=s=>{var m=on[s];return m===void 0?ft(s):m},mr=()=>typeof globalThis=="object"?globalThis:Function("return this")(),Br=s=>{var m=Xt.length;return Xt.push(s),m},Kn=(s,m)=>{for(var g=Array(s),$=0;$<s;++$)g[$]=cr(se[m+4*$>>>2>>>0],"parameter "+$);return g},wt=(s,m)=>Object.defineProperty(m,"name",{value:s});function Yn(s){var m=Function;if(!(m instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof m} which is not a function`);var g=wt(m.name||"unknownFunctionName",function(){});return g.prototype=m.prototype,g=new g,s=m.apply(g,s),s instanceof Object?s:g}var Je=s=>s%4===0&&(s%100!==0||s%400===0),Dr=[0,31,60,91,121,152,182,213,244,274,305,335],fr=[0,31,59,90,120,151,181,212,243,273,304,334],lt=[],hr=(s,m)=>{lt.length=0;for(var g;g=K[s++>>>0];){var $=g!=105;$&=g!=112,m+=$&&m%8?4:0,lt.push(g==112?se[m>>>2>>>0]:g==106?We[m>>>3]:g==105?W[m>>>2>>>0]:ae[m>>>3>>>0]),m+=$?8:4}return lt},Qe={},ot=()=>{if(!Qt){var s={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(typeof navigator=="object"&&navigator.languages&&navigator.languages[0]||"C").replace("-","_")+".UTF-8",_:p||"./this.program"},m;for(m in Qe)Qe[m]===void 0?delete s[m]:s[m]=Qe[m];var g=[];for(m in s)g.push(`${m}=${s[m]}`);Qt=g}return Qt},Qt,Zn=[null,[],[]],an=[31,29,31,30,31,30,31,31,30,31,30,31],zr=[31,28,31,30,31,30,31,31,30,31,30,31];function Xn(s){var m=Array(Wt(s)+1);return at(s,m,0,m.length),m}function sn(s,m,g,$){function T(D,be,Ie){for(D=typeof D=="number"?D.toString():D||"";D.length<be;)D=Ie[0]+D;return D}function B(D,be){return T(D,be,"0")}function H(D,be){function Ie(Mt){return 0>Mt?-1:0<Mt?1:0}var zt;return(zt=Ie(D.getFullYear()-be.getFullYear()))===0&&(zt=Ie(D.getMonth()-be.getMonth()))===0&&(zt=Ie(D.getDate()-be.getDate())),zt}function q(D){switch(D.getDay()){case 0:return new Date(D.getFullYear()-1,11,29);case 1:return D;case 2:return new Date(D.getFullYear(),0,3);case 3:return new Date(D.getFullYear(),0,2);case 4:return new Date(D.getFullYear(),0,1);case 5:return new Date(D.getFullYear()-1,11,31);case 6:return new Date(D.getFullYear()-1,11,30)}}function te(D){var be=D.$a;for(D=new Date(new Date(D.ab+1900,0,1).getTime());0<be;){var Ie=D.getMonth(),zt=(Je(D.getFullYear())?an:zr)[Ie];if(be>zt-D.getDate())be-=zt-D.getDate()+1,D.setDate(1),11>Ie?D.setMonth(Ie+1):(D.setMonth(0),D.setFullYear(D.getFullYear()+1));else{D.setDate(D.getDate()+be);break}}return Ie=new Date(D.getFullYear()+1,0,4),be=q(new Date(D.getFullYear(),0,4)),Ie=q(Ie),0>=H(be,D)?0>=H(Ie,D)?D.getFullYear()+1:D.getFullYear():D.getFullYear()-1}s>>>=0,m>>>=0,g>>>=0,$>>>=0;var X=se[$+40>>>2>>>0];$={Ab:W[$>>>2>>>0],zb:W[$+4>>>2>>>0],fb:W[$+8>>>2>>>0],jb:W[$+12>>>2>>>0],gb:W[$+16>>>2>>>0],ab:W[$+20>>>2>>>0],Va:W[$+24>>>2>>>0],$a:W[$+28>>>2>>>0],Cb:W[$+32>>>2>>>0],yb:W[$+36>>>2>>>0],Bb:X?je(X):""},g=je(g),X={"%c":"%a %b %d %H:%M:%S %Y","%D":"%m/%d/%y","%F":"%Y-%m-%d","%h":"%b","%r":"%I:%M:%S %p","%R":"%H:%M","%T":"%H:%M:%S","%x":"%m/%d/%y","%X":"%H:%M:%S","%Ec":"%c","%EC":"%C","%Ex":"%m/%d/%y","%EX":"%H:%M:%S","%Ey":"%y","%EY":"%Y","%Od":"%d","%Oe":"%e","%OH":"%H","%OI":"%I","%Om":"%m","%OM":"%M","%OS":"%S","%Ou":"%u","%OU":"%U","%OV":"%V","%Ow":"%w","%OW":"%W","%Oy":"%y"};for(var de in X)g=g.replace(new RegExp(de,"g"),X[de]);var Ee="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),Oe="January February March April May June July August September October November December".split(" ");X={"%a":D=>Ee[D.Va].substring(0,3),"%A":D=>Ee[D.Va],"%b":D=>Oe[D.gb].substring(0,3),"%B":D=>Oe[D.gb],"%C":D=>B((D.ab+1900)/100|0,2),"%d":D=>B(D.jb,2),"%e":D=>T(D.jb,2," "),"%g":D=>te(D).toString().substring(2),"%G":te,"%H":D=>B(D.fb,2),"%I":D=>(D=D.fb,D==0?D=12:12<D&&(D-=12),B(D,2)),"%j":D=>{for(var be=0,Ie=0;Ie<=D.gb-1;be+=(Je(D.ab+1900)?an:zr)[Ie++]);return B(D.jb+be,3)},"%m":D=>B(D.gb+1,2),"%M":D=>B(D.zb,2),"%n":()=>`\n`,"%p":D=>0<=D.fb&&12>D.fb?"AM":"PM","%S":D=>B(D.Ab,2),"%t":()=>"	","%u":D=>D.Va||7,"%U":D=>B(Math.floor((D.$a+7-D.Va)/7),2),"%V":D=>{var be=Math.floor((D.$a+7-(D.Va+6)%7)/7);if(2>=(D.Va+371-D.$a-2)%7&&be++,be)be==53&&(Ie=(D.Va+371-D.$a)%7,Ie==4||Ie==3&&Je(D.ab)||(be=1));else{be=52;var Ie=(D.Va+7-D.$a-1)%7;(Ie==4||Ie==5&&Je(D.ab%400-1))&&be++}return B(be,2)},"%w":D=>D.Va,"%W":D=>B(Math.floor((D.$a+7-(D.Va+6)%7)/7),2),"%y":D=>(D.ab+1900).toString().substring(2),"%Y":D=>D.ab+1900,"%z":D=>{D=D.yb;var be=0<=D;return D=Math.abs(D)/60,(be?"+":"-")+("0000"+(D/60*100+D%60)).slice(-4)},"%Z":D=>D.Bb,"%%":()=>"%"},g=g.replace(/%%/g,"\\0\\0");for(de in X)g.includes(de)&&(g=g.replace(new RegExp(de,"g"),X[de]($)));return g=g.replace(/\\0\\0/g,"%"),de=Xn(g),de.length>m?0:(N.set(de,s>>>0),de.length-1)}for(var un=Array(256),gr=0;256>gr;++gr)un[gr]=String.fromCharCode(gr);Cr=un,st=r.BindingError=class extends Error{constructor(s){super(s),this.name="BindingError"}},r.InternalError=class extends Error{constructor(s){super(s),this.name="InternalError"}},Te.push(0,1,void 0,1,null,1,!0,1,!1,1),r.count_emval_handles=()=>Te.length/2-5-Er.length;var Qn={ia:function(s,m,g){return Rr(async()=>{await r.sb(s,m,g)})},a:function(s,m,g){s>>>=0;var $=new Vt(s);throw se[$.hb+16>>>2>>>0]=0,se[$.hb+4>>>2>>>0]=m>>>0,se[$.hb+8>>>2>>>0]=g>>>0,sr=s,jt++,sr},y:function(){return 0},ea:function(){},R:function(){},T:function(){},ga:function(){return 0},ca:function(){},Z:function(){},ba:function(){},G:function(){},S:function(){},P:function(){},da:function(){},Q:function(){},C:function(s,m,g){m=ft(m>>>0),ht(s>>>0,{name:m,fromWireType:$=>$,toWireType:function($,T){if(typeof T!="bigint"&&typeof T!="number")throw T===null?T="null":($=typeof T,T=$==="object"||$==="array"||$==="function"?T.toString():""+T),new TypeError(`Cannot convert "${T}" to ${this.name}`);return typeof T=="number"&&(T=BigInt(T)),T},argPackAdvance:8,readValueFromPointer:bt(m,g>>>0,m.indexOf("u")==-1),bb:null})},K:function(s,m,g,$){m=ft(m>>>0),ht(s>>>0,{name:m,fromWireType:function(T){return!!T},toWireType:function(T,B){return B?g:$},argPackAdvance:8,readValueFromPointer:function(T){return this.fromWireType(K[T>>>0])},bb:null})},J:function(s){return ht(s>>>0,Fn)},B:function(s,m,g){m=ft(m>>>0),ht(s>>>0,{name:m,fromWireType:$=>$,toWireType:($,T)=>T,argPackAdvance:8,readValueFromPointer:ge(m,g>>>0),bb:null})},s:function(s,m,g,$,T){if(s>>>=0,g>>>=0,m=ft(m>>>0),T===-1&&(T=4294967295),T=q=>q,$===0){var B=32-8*g;T=q=>q<<B>>>B}var H=m.includes("unsigned")?function(q,te){return te>>>0}:function(q,te){return te};ht(s,{name:m,fromWireType:T,toWireType:H,argPackAdvance:8,readValueFromPointer:bt(m,g,$!==0),bb:null})},o:function(s,m,g){function $(B){return new T(N.buffer,se[B+4>>>2>>>0],se[B>>>2>>>0])}var T=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][m];g=ft(g>>>0),ht(s>>>0,{name:g,fromWireType:$,argPackAdvance:8,readValueFromPointer:$},{nb:!0})},D:function(s,m){m=ft(m>>>0);var g=m==="std::string";ht(s>>>0,{name:m,fromWireType:function($){var T=se[$>>>2>>>0],B=$+4;if(g)for(var H=B,q=0;q<=T;++q){var te=B+q;if(q==T||K[te>>>0]==0){if(H=je(H,te-H),X===void 0)var X=H;else X+=String.fromCharCode(0),X+=H;H=te+1}}else{for(X=Array(T),q=0;q<T;++q)X[q]=String.fromCharCode(K[B+q>>>0]);X=X.join("")}return rt($),X},toWireType:function($,T){T instanceof ArrayBuffer&&(T=new Uint8Array(T));var B=typeof T=="string";if(!(B||T instanceof Uint8Array||T instanceof Uint8ClampedArray||T instanceof Int8Array))throw new st("Cannot pass non-string to std::string");var H=g&&B?Wt(T):T.length,q=Nt(4+H+1),te=q+4;if(se[q>>>2>>>0]=H,g&&B)at(T,K,te,H+1);else if(B)for(B=0;B<H;++B){var X=T.charCodeAt(B);if(255<X)throw rt(te),new st("String has UTF-16 code units that do not fit in 8 bits");K[te+B>>>0]=X}else for(B=0;B<H;++B)K[te+B>>>0]=T[B];return $!==null&&$.push(rt,q),q},argPackAdvance:8,readValueFromPointer:lr,bb($){rt($)}})},x:function(s,m,g){if(m>>>=0,g>>>=0,g=ft(g),m===2)var $=jr,T=qn,B=Kr,H=q=>he[q>>>1>>>0];else m===4&&($=jn,T=Yr,B=Zr,H=q=>se[q>>>2>>>0]);ht(s>>>0,{name:g,fromWireType:q=>{for(var te=se[q>>>2>>>0],X,de=q+4,Ee=0;Ee<=te;++Ee){var Oe=q+4+Ee*m;(Ee==te||H(Oe)==0)&&(de=$(de,Oe-de),X===void 0?X=de:(X+=String.fromCharCode(0),X+=de),de=Oe+m)}return rt(q),X},toWireType:(q,te)=>{if(typeof te!="string")throw new st(`Cannot pass non-string to C++ string type ${g}`);var X=B(te),de=Nt(4+X+m);return se[de>>>2>>>0]=X/m,T(te,de+4,X+m),q!==null&&q.push(rt,de),de},argPackAdvance:8,readValueFromPointer:lr,bb(q){rt(q)}})},L:function(s,m){m=ft(m>>>0),ht(s>>>0,{ob:!0,name:m,argPackAdvance:0,fromWireType:()=>{},toWireType:()=>{}})},ha:()=>1,u:function(s,m,g){return m>>>=0,g>>>=0,s=Xe(s>>>0),m=cr(m,"emval::as"),Pr(m,g,s)},w:function(s){return s>>>=0,Rr(()=>(s=Xe(s),s.then(ut)))},n:function(s,m,g,$){return g>>>=0,$>>>=0,s=Xt[s>>>0],m=Xe(m>>>0),s(null,m,g,$)},j:function(s,m,g,$,T){return g>>>=0,$>>>=0,T>>>=0,s=Xt[s>>>0],m=Xe(m>>>0),g=nt(g),s(m,m[g],$,T)},b:dr,A:function(s,m){return m>>>=0,s=Xe(s>>>0),m=Xe(m),s==m},m:function(s){return s>>>=0,s===0?ut(mr()):(s=nt(s),ut(mr()[s]))},i:function(s,m,g){m=Kn(s,m>>>0);var $=m.shift();s--;var T=`return function (obj, func, destructorsRef, args) {\n`,B=0,H=[];g===0&&H.push("obj");for(var q=["retType"],te=[$],X=0;X<s;++X)H.push("arg"+X),q.push("argType"+X),te.push(m[X]),T+=`  var arg${X} = argType${X}.readValueFromPointer(args${B?"+"+B:""});\n`,B+=m[X].argPackAdvance;return T+=`  var rv = ${g===1?"new func":"func.call"}(${H.join(", ")});\n`,$.ob||(q.push("emval_returnValue"),te.push(Pr),T+=`  return emval_returnValue(retType, destructorsRef, rv);\n`),q.push(T+`};\n`),s=Yn(q)(...te),g=`methodCaller<(${m.map(de=>de.name).join(", ")}) => ${$.name}>`,Br(wt(g,s))},r:function(s,m){return m>>>=0,s=Xe(s>>>0),m=Xe(m),ut(s[m])},e:function(s){s>>>=0,9<s&&(Te[s+1]+=1)},t:function(){return ut([])},l:function(s){s=Xe(s>>>0);for(var m=Array(s.length),g=0;g<s.length;g++)m[g]=s[g];return ut(m)},f:function(s){return ut(nt(s>>>0))},k:function(){return ut({})},h:function(s){s>>>=0;for(var m=Xe(s);m.length;){var g=m.pop();m.pop()(g)}dr(s)},g:function(s,m,g){m>>>=0,g>>>=0,s=Xe(s>>>0),m=Xe(m),g=Xe(g),s[m]=g},c:function(s,m){return m>>>=0,s=cr(s>>>0,"_emval_take_value"),s=s.readValueFromPointer(m),ut(s)},W:function(s,m){s=-9007199254740992>s||9007199254740992<s?NaN:Number(s),m>>>=0,s=new Date(1e3*s),W[m>>>2>>>0]=s.getUTCSeconds(),W[m+4>>>2>>>0]=s.getUTCMinutes(),W[m+8>>>2>>>0]=s.getUTCHours(),W[m+12>>>2>>>0]=s.getUTCDate(),W[m+16>>>2>>>0]=s.getUTCMonth(),W[m+20>>>2>>>0]=s.getUTCFullYear()-1900,W[m+24>>>2>>>0]=s.getUTCDay(),W[m+28>>>2>>>0]=(s.getTime()-Date.UTC(s.getUTCFullYear(),0,1,0,0,0,0))/864e5|0},X:function(s,m){s=-9007199254740992>s||9007199254740992<s?NaN:Number(s),m>>>=0,s=new Date(1e3*s),W[m>>>2>>>0]=s.getSeconds(),W[m+4>>>2>>>0]=s.getMinutes(),W[m+8>>>2>>>0]=s.getHours(),W[m+12>>>2>>>0]=s.getDate(),W[m+16>>>2>>>0]=s.getMonth(),W[m+20>>>2>>>0]=s.getFullYear()-1900,W[m+24>>>2>>>0]=s.getDay(),W[m+28>>>2>>>0]=(Je(s.getFullYear())?Dr:fr)[s.getMonth()]+s.getDate()-1|0,W[m+36>>>2>>>0]=-(60*s.getTimezoneOffset());var g=new Date(s.getFullYear(),6,1).getTimezoneOffset(),$=new Date(s.getFullYear(),0,1).getTimezoneOffset();W[m+32>>>2>>>0]=(g!=$&&s.getTimezoneOffset()==Math.min($,g))|0},Y:function(s){s>>>=0;var m=new Date(W[s+20>>>2>>>0]+1900,W[s+16>>>2>>>0],W[s+12>>>2>>>0],W[s+8>>>2>>>0],W[s+4>>>2>>>0],W[s>>>2>>>0],0),g=W[s+32>>>2>>>0],$=m.getTimezoneOffset(),T=new Date(m.getFullYear(),6,1).getTimezoneOffset(),B=new Date(m.getFullYear(),0,1).getTimezoneOffset(),H=Math.min(B,T);return 0>g?W[s+32>>>2>>>0]=+(T!=B&&H==$):0<g!=(H==$)&&(T=Math.max(B,T),m.setTime(m.getTime()+6e4*((0<g?H:T)-$))),W[s+24>>>2>>>0]=m.getDay(),W[s+28>>>2>>>0]=(Je(m.getFullYear())?Dr:fr)[m.getMonth()]+m.getDate()-1|0,W[s>>>2>>>0]=m.getSeconds(),W[s+4>>>2>>>0]=m.getMinutes(),W[s+8>>>2>>>0]=m.getHours(),W[s+12>>>2>>>0]=m.getDate(),W[s+16>>>2>>>0]=m.getMonth(),W[s+20>>>2>>>0]=m.getYear(),s=m.getTime(),BigInt(isNaN(s)?-1:s/1e3)},U:function(){return-52},V:function(){},N:function(s,m,g,$){g>>>=0,$>>>=0;var T=new Date().getFullYear(),B=new Date(T,0,1),H=new Date(T,6,1);T=B.getTimezoneOffset();var q=H.getTimezoneOffset();se[s>>>0>>>2>>>0]=60*Math.max(T,q),W[m>>>0>>>2>>>0]=+(T!=q),s=te=>te.toLocaleTimeString(void 0,{hour12:!1,timeZoneName:"short"}).split(" ")[1],B=s(B),H=s(H),q<T?(at(B,K,g,17),at(H,K,$,17)):(at(B,K,$,17),at(H,K,g,17))},v:()=>{ne("")},d:function(s,m,g){return s>>>=0,m=hr(m>>>0,g>>>0),Rt[s](...m)},I:function(s,m,g){return s>>>=0,m=hr(m>>>0,g>>>0),Rt[s](...m)},H:()=>Date.now(),O:function(){return 4294901760},q:()=>performance.now(),M:function(s){s>>>=0;var m=K.length;if(4294901760<s)return!1;for(var g=1;4>=g;g*=2){var $=m*(1+.2/g);$=Math.min($,s+100663296);var T=Math;$=Math.max(s,$);e:{T=(T.min.call(T,4294901760,$+(65536-$%65536)%65536)-O.buffer.byteLength+65535)/65536;try{O.grow(T),Ae();var B=1;break e}catch{}B=void 0}if(B)return!0}return!1},$:function(s,m){s>>>=0,m>>>=0;var g=0;return ot().forEach(($,T)=>{var B=m+g;for(T=se[s+4*T>>>2>>>0]=B,B=0;B<$.length;++B)N[T++>>>0]=$.charCodeAt(B);N[T>>>0]=0,g+=$.length+1}),0},aa:function(s,m){s>>>=0,m>>>=0;var g=ot();se[s>>>2>>>0]=g.length;var $=0;return g.forEach(T=>$+=T.length+1),se[m>>>2>>>0]=$,0},z:()=>52,F:function(){return 52},_:function(){return 70},E:function(s,m,g,$){m>>>=0,g>>>=0,$>>>=0;for(var T=0,B=0;B<g;B++){var H=se[m>>>2>>>0],q=se[m+4>>>2>>>0];m+=8;for(var te=0;te<q;te++){var X=K[H+te>>>0],de=Zn[s];X===0||X===10?((s===1?E:P)(Kt(de,0)),de.length=0):de.push(X)}T+=q}return se[$>>>2>>>0]=T,0},fa:sn,p:function(s,m,g,$){return sn(s>>>0,m>>>0,g>>>0,$>>>0)}},pe=function(){function s(g){return pe=g.exports,pe=Qr(),pe=ln(),O=pe.ja,Ae(),ie.unshift(pe.ka),le--,le==0&&(qe!==null&&(clearInterval(qe),qe=null),G&&(g=G,G=null,g())),pe}var m={a:Qn};if(le++,r.instantiateWasm)try{return r.instantiateWasm(m,s)}catch(g){P(`Module.instantiateWasm callback failed with error: ${g}`),i(g)}return mt(m,function(g){s(g.instance)}).catch(i),{}}(),Mr=s=>(Mr=pe.la)(s);r._OrtInit=(s,m)=>(r._OrtInit=pe.ma)(s,m),r._OrtGetLastError=(s,m)=>(r._OrtGetLastError=pe.na)(s,m),r._OrtCreateSessionOptions=(s,m,g,$,T,B,H,q,te,X)=>(r._OrtCreateSessionOptions=pe.oa)(s,m,g,$,T,B,H,q,te,X),r._OrtAppendExecutionProvider=(s,m)=>(r._OrtAppendExecutionProvider=pe.pa)(s,m),r._OrtAddFreeDimensionOverride=(s,m,g)=>(r._OrtAddFreeDimensionOverride=pe.qa)(s,m,g),r._OrtAddSessionConfigEntry=(s,m,g)=>(r._OrtAddSessionConfigEntry=pe.ra)(s,m,g),r._OrtReleaseSessionOptions=s=>(r._OrtReleaseSessionOptions=pe.sa)(s),r._OrtCreateSession=(s,m,g)=>(r._OrtCreateSession=pe.ta)(s,m,g),r._OrtReleaseSession=s=>(r._OrtReleaseSession=pe.ua)(s),r._OrtGetInputOutputCount=(s,m,g)=>(r._OrtGetInputOutputCount=pe.va)(s,m,g),r._OrtGetInputName=(s,m)=>(r._OrtGetInputName=pe.wa)(s,m),r._OrtGetOutputName=(s,m)=>(r._OrtGetOutputName=pe.xa)(s,m),r._OrtFree=s=>(r._OrtFree=pe.ya)(s),r._OrtCreateTensor=(s,m,g,$,T,B)=>(r._OrtCreateTensor=pe.za)(s,m,g,$,T,B),r._OrtGetTensorData=(s,m,g,$,T)=>(r._OrtGetTensorData=pe.Aa)(s,m,g,$,T),r._OrtReleaseTensor=s=>(r._OrtReleaseTensor=pe.Ba)(s),r._OrtCreateRunOptions=(s,m,g,$)=>(r._OrtCreateRunOptions=pe.Ca)(s,m,g,$),r._OrtAddRunConfigEntry=(s,m,g)=>(r._OrtAddRunConfigEntry=pe.Da)(s,m,g),r._OrtReleaseRunOptions=s=>(r._OrtReleaseRunOptions=pe.Ea)(s),r._OrtCreateBinding=s=>(r._OrtCreateBinding=pe.Fa)(s),r._OrtBindInput=(s,m,g)=>(r._OrtBindInput=pe.Ga)(s,m,g),r._OrtBindOutput=(s,m,g,$)=>(r._OrtBindOutput=pe.Ha)(s,m,g,$),r._OrtClearBoundOutputs=s=>(r._OrtClearBoundOutputs=pe.Ia)(s),r._OrtReleaseBinding=s=>(r._OrtReleaseBinding=pe.Ja)(s),r._OrtRunWithBinding=(s,m,g,$,T)=>(r._OrtRunWithBinding=pe.Ka)(s,m,g,$,T),r._OrtRun=(s,m,g,$,T,B,H,q)=>(r._OrtRun=pe.La)(s,m,g,$,T,B,H,q),r._OrtEndProfiling=s=>(r._OrtEndProfiling=pe.Ma)(s),r._JsepOutput=(s,m,g)=>(r._JsepOutput=pe.Na)(s,m,g),r._JsepGetNodeName=s=>(r._JsepGetNodeName=pe.Oa)(s);var Nt=r._malloc=s=>(Nt=r._malloc=pe.Pa)(s),rt=r._free=s=>(rt=r._free=pe.Qa)(s),yr=s=>(yr=pe.Sa)(s),br=s=>(br=pe.Ta)(s),Ur=()=>(Ur=pe.Ua)(),Jt=s=>(Jt=pe.Wa)(s),dn=()=>(dn=pe.Xa)(),vt=s=>(vt=pe.Ya)(s),ct=()=>(ct=pe.Za)();r.___start_em_js=837952,r.___stop_em_js=838113;function ln(){var s=pe;s=Object.assign({},s);var m=g=>$=>g($)>>>0;return s.la=m(s.la),s.Pa=m(s.Pa),s.Ta=m(s.Ta),s.Ua=(g=>()=>g()>>>0)(s.Ua),s}r.stackSave=()=>Ur(),r.stackRestore=s=>yr(s),r.stackAlloc=s=>br(s),r.UTF8ToString=je,r.stringToUTF8=(s,m,g)=>at(s,K,m,g),r.lengthBytesUTF8=Wt;var Dt;G=function s(){Dt||Vr(),Dt||(G=s)};function Vr(){if(!(0<le)){if(r.preRun)for(typeof r.preRun=="function"&&(r.preRun=[r.preRun]);r.preRun.length;){var s=r.preRun.shift();me.unshift(s)}for(;0<me.length;)me.shift()(r);if(!(0<le||Dt||(Dt=!0,r.calledRun=!0,R))){for(;0<ie.length;)ie.shift()(r);for(o(r);0<ue.length;)ue.shift()(r)}}}return Vr(),u}})();typeof Ti=="object"&&typeof po=="object"?po.exports=Ii:typeof define=="function"&&define.amd&&define([],()=>Ii)});var Pi=Wr(()=>{});var ki=Wr(()=>{});var Oi={};gn(Oi,{cpus:()=>ll});var ll,Ri=Y(()=>{ll=void 0});var zi=Wr((Di,mo)=>{"use strict";var Bi=(()=>{var e=typeof document<"u"?document.currentScript?.src:void 0;return typeof __filename<"u"&&(e||=__filename),function(t={}){function r(){return ae.buffer!=ue.buffer&&Ne(),ue}function o(){return ae.buffer!=ue.buffer&&Ne(),le}function i(){return ae.buffer!=ue.buffer&&Ne(),qe}function u(){return ae.buffer!=ue.buffer&&Ne(),G}function a(){return ae.buffer!=ue.buffer&&Ne(),ne}function c(){return ae.buffer!=ue.buffer&&Ne(),xe}function p(){return ae.buffer!=ue.buffer&&Ne(),Ke}function h(){return ae.buffer!=ue.buffer&&Ne(),Ut}var d=t,y,w,_=new Promise((n,l)=>{y=n,w=l});d.mountExternalData=(n,l)=>{(d.Db||(d.Db=new Map)).set(n,l)},d.unmountExternalData=()=>{delete d.Db};let v=()=>{let n=(f,b,C)=>(...k)=>{let V=ct,F=b?.();k=f(...k);let re=b?.();return F!==re&&(f=re,C(F),b=C=null),ct!=V?T():k},l=f=>async(...b)=>{try{if(d.Cb)throw Error("Session already started");let C=d.Cb={ec:b[0],errors:[]},k=await f(...b);if(d.Cb!==C)throw Error("Session mismatch");d.Kb?.flush();let V=C.errors;if(0<V.length){let F=await Promise.all(V);if(F=F.filter(re=>re),0<F.length)throw Error(F.join(`\n`))}return k}finally{d.Cb=null}};d._OrtCreateSession=n(d._OrtCreateSession,()=>d._OrtCreateSession,f=>d._OrtCreateSession=f),d._OrtRun=l(n(d._OrtRun,()=>d._OrtRun,f=>d._OrtRun=f)),d._OrtRunWithBinding=l(n(d._OrtRunWithBinding,()=>d._OrtRunWithBinding,f=>d._OrtRunWithBinding=f)),d._OrtBindInput=n(d._OrtBindInput,()=>d._OrtBindInput,f=>d._OrtBindInput=f),v=void 0};d.jsepInit=(n,l)=>{if(v?.(),n==="webgpu"){[d.Kb,d.Wb,d.$b,d.Lb,d.Zb,d.ob,d.ac,d.cc,d.Xb,d.Yb,d.bc]=l;let f=d.Kb;d.jsepRegisterBuffer=(b,C,k,V)=>f.registerBuffer(b,C,k,V),d.jsepGetBuffer=b=>f.getBuffer(b),d.jsepCreateDownloader=(b,C,k)=>f.createDownloader(b,C,k),d.jsepOnReleaseSession=b=>{f.onReleaseSession(b)},d.jsepOnRunStart=b=>f.onRunStart(b)}};var S=Object.assign({},d),A="./this.program",I=(n,l)=>{throw l},x=typeof window=="object",E=typeof importScripts=="function",P=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",O=d.ENVIRONMENT_IS_PTHREAD||!1,R="";function L(n){return d.locateFile?d.locateFile(n,R):R+n}var N,K,Q;if(P){var he=(uo(),wr(so)),W=(co(),wr(lo));R=E?W.dirname(R)+"/":__dirname+"/",N=(n,l)=>(n=Wt(n)?new URL(n):W.normalize(n),he.readFileSync(n,l?void 0:"utf8")),Q=n=>(n=N(n,!0),n.buffer||(n=new Uint8Array(n)),n),K=(n,l,f,b=!0)=>{n=Wt(n)?new URL(n):W.normalize(n),he.readFile(n,b?void 0:"utf8",(C,k)=>{C?f(C):l(b?k.buffer:k)})},!d.thisProgram&&1<process.argv.length&&(A=process.argv[1].replace(/\\\\/g,"/")),process.argv.slice(2),I=(n,l)=>{throw process.exitCode=n,l},global.Worker=Pi().Worker}else(x||E)&&(E?R=self.location.href:typeof document<"u"&&document.currentScript&&(R=document.currentScript.src),typeof e<"u"&&e&&(R=e),R.startsWith("blob:")?R="":R=R.substr(0,R.replace(/[?#].*/,"").lastIndexOf("/")+1),P||(N=n=>{var l=new XMLHttpRequest;return l.open("GET",n,!1),l.send(null),l.responseText},E&&(Q=n=>{var l=new XMLHttpRequest;return l.open("GET",n,!1),l.responseType="arraybuffer",l.send(null),new Uint8Array(l.response)}),K=(n,l,f)=>{var b=new XMLHttpRequest;b.open("GET",n,!0),b.responseType="arraybuffer",b.onload=()=>{b.status==200||b.status==0&&b.response?l(b.response):f()},b.onerror=f,b.send(null)}));P&&typeof performance>"u"&&(global.performance=ki().performance);var se=console.log.bind(console),Ce=console.error.bind(console);P&&(se=(...n)=>he.writeSync(1,n.join(" ")+`\n`),Ce=(...n)=>he.writeSync(2,n.join(" ")+`\n`));var We=se,ee=Ce;Object.assign(d,S),S=null;var ae,Ae,me=!1,ie,ue,le,qe,G,ne,xe,Ke,Be,Ge,Ut;function Ne(){var n=ae.buffer;d.HEAP8=ue=new Int8Array(n),d.HEAP16=qe=new Int16Array(n),d.HEAPU8=le=new Uint8Array(n),d.HEAPU16=G=new Uint16Array(n),d.HEAP32=ne=new Int32Array(n),d.HEAPU32=xe=new Uint32Array(n),d.HEAPF32=Ke=new Float32Array(n),d.HEAPF64=Ut=new Float64Array(n),d.HEAP64=Be=new BigInt64Array(n),d.HEAPU64=Ge=new BigUint64Array(n)}var Ye=16777216;if(O)ae=d.wasmMemory;else if(d.wasmMemory)ae=d.wasmMemory;else if(ae=new WebAssembly.Memory({initial:Ye/65536,maximum:65536,shared:!0}),!(ae.buffer instanceof SharedArrayBuffer))throw ee("requested a shared WebAssembly.Memory but the returned buffer is not a SharedArrayBuffer, indicating that while the browser has SharedArrayBuffer it does not have WebAssembly threads support - you may need to set a flag"),P&&ee("(on node you may need: --experimental-wasm-threads --experimental-wasm-bulk-memory and/or recent version)"),Error("bad memory");Ne(),Ye=ae.buffer.byteLength;var mt=[],Rt=[],qt=[],Vt=0,sr=null,jt=null;function xr(){if(Vt--,Vt==0&&(sr!==null&&(clearInterval(sr),sr=null),jt)){var n=jt;jt=null,n()}}function Kt(n){throw n="Aborted("+n+")",ee(n),me=!0,ie=1,n=new WebAssembly.RuntimeError(n+". Build with -sASSERTIONS for more info."),w(n),n}var je=n=>n.startsWith("data:application/octet-stream;base64,"),Wt=n=>n.startsWith("file://"),at;at="ort-wasm-simd-threaded.wasm",je(at)||(at=L(at));function Cr(n){if(Q)return Q(n);throw"both async and sync fetching of the wasm failed"}function ft(n){if(x||E){if(typeof fetch=="function"&&!Wt(n))return fetch(n,{credentials:"same-origin"}).then(l=>{if(!l.ok)throw`failed to load wasm binary file at \'${n}\'`;return l.arrayBuffer()}).catch(()=>Cr(n));if(K)return new Promise((l,f)=>{K(n,b=>l(new Uint8Array(b)),f)})}return Promise.resolve().then(()=>Cr(n))}function ur(n,l,f){return ft(n).then(b=>WebAssembly.instantiate(b,l)).then(f,b=>{ee(`failed to asynchronously prepare wasm: ${b}`),Kt(b)})}function Ar(n,l){var f=at;return typeof WebAssembly.instantiateStreaming!="function"||je(f)||Wt(f)||P||typeof fetch!="function"?ur(f,n,l):fetch(f,{credentials:"same-origin"}).then(b=>WebAssembly.instantiateStreaming(b,n).then(l,function(C){return ee(`wasm streaming compile failed: ${C}`),ee("falling back to ArrayBuffer instantiation"),ur(f,n,l)}))}var Ir={826468:(n,l,f,b)=>{if(typeof d>"u"||!d.Db)return 1;if(n=He(n>>>0),n.startsWith("./")&&(n=n.substring(2)),n=d.Db.get(n),!n)return 2;if(l>>>=0,f>>>=0,b>>>=0,l+f>n.byteLength)return 3;try{return o().set(n.subarray(l,l+f),b>>>0),0}catch{return 4}},826969:()=>{d.Xb()},827e3:()=>{d.Yb()},827029:()=>{d.bc()},827054:n=>d.Wb(n),827087:n=>d.$b(n),827119:(n,l,f)=>{d.Lb(n,l,f,!0)},827158:(n,l,f)=>{d.Lb(n,l,f)},827191:n=>{d.ob("Abs",n,void 0)},827242:n=>{d.ob("Neg",n,void 0)},827293:n=>{d.ob("Floor",n,void 0)},827346:n=>{d.ob("Ceil",n,void 0)},827398:n=>{d.ob("Reciprocal",n,void 0)},827456:n=>{d.ob("Sqrt",n,void 0)},827508:n=>{d.ob("Exp",n,void 0)},827559:n=>{d.ob("Erf",n,void 0)},827610:n=>{d.ob("Sigmoid",n,void 0)},827665:(n,l,f)=>{d.ob("HardSigmoid",n,{alpha:l,beta:f})},827744:n=>{d.ob("Log",n,void 0)},827795:n=>{d.ob("Sin",n,void 0)},827846:n=>{d.ob("Cos",n,void 0)},827897:n=>{d.ob("Tan",n,void 0)},827948:n=>{d.ob("Asin",n,void 0)},828e3:n=>{d.ob("Acos",n,void 0)},828052:n=>{d.ob("Atan",n,void 0)},828104:n=>{d.ob("Sinh",n,void 0)},828156:n=>{d.ob("Cosh",n,void 0)},828208:n=>{d.ob("Asinh",n,void 0)},828261:n=>{d.ob("Acosh",n,void 0)},828314:n=>{d.ob("Atanh",n,void 0)},828367:n=>{d.ob("Tanh",n,void 0)},828419:n=>{d.ob("Not",n,void 0)},828470:(n,l,f)=>{d.ob("Clip",n,{min:l,max:f})},828539:n=>{d.ob("Clip",n,void 0)},828591:(n,l)=>{d.ob("Elu",n,{alpha:l})},828649:n=>{d.ob("Relu",n,void 0)},828701:(n,l)=>{d.ob("LeakyRelu",n,{alpha:l})},828765:(n,l)=>{d.ob("ThresholdedRelu",n,{alpha:l})},828835:(n,l)=>{d.ob("Cast",n,{to:l})},828893:n=>{d.ob("Add",n,void 0)},828944:n=>{d.ob("Sub",n,void 0)},828995:n=>{d.ob("Mul",n,void 0)},829046:n=>{d.ob("Div",n,void 0)},829097:n=>{d.ob("Pow",n,void 0)},829148:n=>{d.ob("Equal",n,void 0)},829201:n=>{d.ob("Greater",n,void 0)},829256:n=>{d.ob("GreaterOrEqual",n,void 0)},829318:n=>{d.ob("Less",n,void 0)},829370:n=>{d.ob("LessOrEqual",n,void 0)},829429:(n,l,f,b,C)=>{d.ob("ReduceMean",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},829588:(n,l,f,b,C)=>{d.ob("ReduceMax",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},829746:(n,l,f,b,C)=>{d.ob("ReduceMin",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},829904:(n,l,f,b,C)=>{d.ob("ReduceProd",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830063:(n,l,f,b,C)=>{d.ob("ReduceSum",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830221:(n,l,f,b,C)=>{d.ob("ReduceL1",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830378:(n,l,f,b,C)=>{d.ob("ReduceL2",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830535:(n,l,f,b,C)=>{d.ob("ReduceLogSum",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830696:(n,l,f,b,C)=>{d.ob("ReduceSumSquare",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830860:(n,l,f,b,C)=>{d.ob("ReduceLogSumExp",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},831024:n=>{d.ob("Where",n,void 0)},831077:(n,l,f)=>{d.ob("Transpose",n,{perm:l?Array.from(a().subarray(l>>>0,f>>>0)):[]})},831185:(n,l,f,b)=>{d.ob("DepthToSpace",n,{blocksize:l,mode:He(f),format:b?"NHWC":"NCHW"})},831318:(n,l,f,b)=>{d.ob("DepthToSpace",n,{blocksize:l,mode:He(f),format:b?"NHWC":"NCHW"})},831451:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we)=>{d.ob("ConvTranspose",n,{format:re?"NHWC":"NCHW",autoPad:l,dilations:[f],group:b,kernelShape:[C],pads:[k,V],strides:[F],wIsConst:()=>!!r()[J>>>0],outputPadding:ce?Array.from(a().subarray(ce>>>0,Re>>>0)):[],outputShape:ze?Array.from(a().subarray(ze>>>0,z>>>0)):[],activation:He(we)})},831852:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z)=>{d.ob("ConvTranspose",n,{format:F?"NHWC":"NCHW",autoPad:l,dilations:Array.from(a().subarray(f>>>0,(f>>>0)+2>>>0)),group:b,kernelShape:Array.from(a().subarray(C>>>0,(C>>>0)+2>>>0)),pads:Array.from(a().subarray(k>>>0,(k>>>0)+4>>>0)),strides:Array.from(a().subarray(V>>>0,(V>>>0)+2>>>0)),wIsConst:()=>!!r()[re>>>0],outputPadding:J?Array.from(a().subarray(J>>>0,ce>>>0)):[],outputShape:Re?Array.from(a().subarray(Re>>>0,ze>>>0)):[],activation:He(z)})},832417:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we)=>{d.ob("ConvTranspose",n,{format:re?"NHWC":"NCHW",autoPad:l,dilations:[f],group:b,kernelShape:[C],pads:[k,V],strides:[F],wIsConst:()=>!!r()[J>>>0],outputPadding:ce?Array.from(a().subarray(ce>>>0,Re>>>0)):[],outputShape:ze?Array.from(a().subarray(ze>>>0,z>>>0)):[],activation:He(we)})},832818:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z)=>{d.ob("ConvTranspose",n,{format:F?"NHWC":"NCHW",autoPad:l,dilations:Array.from(a().subarray(f>>>0,(f>>>0)+2>>>0)),group:b,kernelShape:Array.from(a().subarray(C>>>0,(C>>>0)+2>>>0)),pads:Array.from(a().subarray(k>>>0,(k>>>0)+4>>>0)),strides:Array.from(a().subarray(V>>>0,(V>>>0)+2>>>0)),wIsConst:()=>!!r()[re>>>0],outputPadding:J?Array.from(a().subarray(J>>>0,ce>>>0)):[],outputShape:Re?Array.from(a().subarray(Re>>>0,ze>>>0)):[],activation:He(z)})},833383:(n,l)=>{d.ob("GlobalAveragePool",n,{format:l?"NHWC":"NCHW"})},833474:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we,Pe)=>{d.ob("AveragePool",n,{format:Pe?"NHWC":"NCHW",auto_pad:l,ceil_mode:f,count_include_pad:b,storage_order:C,dilations:[k,V],kernel_shape:[F,re],pads:[J,ce,Re,ze],strides:[z,we]})},833758:(n,l)=>{d.ob("GlobalAveragePool",n,{format:l?"NHWC":"NCHW"})},833849:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we,Pe)=>{d.ob("AveragePool",n,{format:Pe?"NHWC":"NCHW",auto_pad:l,ceil_mode:f,count_include_pad:b,storage_order:C,dilations:[k,V],kernel_shape:[F,re],pads:[J,ce,Re,ze],strides:[z,we]})},834133:(n,l)=>{d.ob("GlobalMaxPool",n,{format:l?"NHWC":"NCHW"})},834220:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we,Pe)=>{d.ob("MaxPool",n,{format:Pe?"NHWC":"NCHW",auto_pad:l,ceil_mode:f,count_include_pad:b,storage_order:C,dilations:[k,V],kernel_shape:[F,re],pads:[J,ce,Re,ze],strides:[z,we]})},834500:(n,l)=>{d.ob("GlobalMaxPool",n,{format:l?"NHWC":"NCHW"})},834587:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we,Pe)=>{d.ob("MaxPool",n,{format:Pe?"NHWC":"NCHW",auto_pad:l,ceil_mode:f,count_include_pad:b,storage_order:C,dilations:[k,V],kernel_shape:[F,re],pads:[J,ce,Re,ze],strides:[z,we]})},834867:(n,l,f,b,C)=>{d.ob("Gemm",n,{alpha:l,beta:f,transA:b,transB:C})},834971:n=>{d.ob("MatMul",n,void 0)},835025:(n,l,f,b)=>{d.ob("ArgMax",n,{keepDims:!!l,selectLastIndex:!!f,axis:b})},835133:(n,l,f,b)=>{d.ob("ArgMin",n,{keepDims:!!l,selectLastIndex:!!f,axis:b})},835241:(n,l)=>{d.ob("Softmax",n,{axis:l})},835304:(n,l)=>{d.ob("Concat",n,{axis:l})},835364:(n,l,f,b,C)=>{d.ob("Split",n,{axis:l,numOutputs:f,splitSizes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},835504:n=>{d.ob("Expand",n,void 0)},835558:(n,l)=>{d.ob("Gather",n,{axis:Number(l)})},835629:(n,l)=>{d.ob("GatherElements",n,{axis:Number(l)})},835708:(n,l,f,b,C,k,V,F,re,J,ce)=>{d.ob("Resize",n,{antialias:l,axes:f?Array.from(a().subarray(f>>>0,b>>>0)):[],coordinateTransformMode:He(C),cubicCoeffA:k,excludeOutside:V,extrapolationValue:F,keepAspectRatioPolicy:He(re),mode:He(J),nearestMode:He(ce)})},836054:(n,l,f,b,C,k,V)=>{d.ob("Slice",n,{starts:l?Array.from(a().subarray(l>>>0,f>>>0)):[],ends:b?Array.from(a().subarray(b>>>0,C>>>0)):[],axes:k?Array.from(a().subarray(k>>>0,V>>>0)):[]})},836270:n=>{d.ob("Tile",n,void 0)},836322:(n,l,f,b)=>{d.ob("LayerNormalization",n,{axis:l,epsilon:f,simplified:!!b})},836433:(n,l,f)=>{d.ob("InstanceNormalization",n,{epsilon:l,format:f?"NHWC":"NCHW"})},836547:(n,l,f)=>{d.ob("InstanceNormalization",n,{epsilon:l,format:f?"NHWC":"NCHW"})},836661:n=>{d.ob("Range",n,void 0)},836714:(n,l)=>{d.ob("Einsum",n,{equation:He(l)})},836795:(n,l,f,b,C)=>{d.ob("Pad",n,{mode:l,value:f,pads:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},836922:(n,l,f,b,C,k)=>{d.ob("BatchNormalization",n,{epsilon:l,momentum:f,spatial:!!C,trainingMode:!!b,format:k?"NHWC":"NCHW"})},837091:(n,l,f,b,C,k)=>{d.ob("BatchNormalization",n,{epsilon:l,momentum:f,spatial:!!C,trainingMode:!!b,format:k?"NHWC":"NCHW"})},837260:(n,l,f)=>{d.ob("CumSum",n,{exclusive:Number(l),reverse:Number(f)})},837357:(n,l,f,b,C,k,V,F,re)=>{d.ob("Attention",n,{numHeads:l,isUnidirectional:f,maskFilterValue:b,scale:C,doRotary:k,qkvHiddenSizes:V?Array.from(a().subarray(Number(F)>>>0,Number(F)+V>>>0)):[],pastPresentShareBuffer:!!re})},837629:n=>{d.ob("BiasAdd",n,void 0)},837684:n=>{d.ob("BiasSplitGelu",n,void 0)},837745:n=>{d.ob("FastGelu",n,void 0)},837801:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze)=>{d.ob("Conv",n,{format:re?"NHWC":"NCHW",auto_pad:l,dilations:[f],group:b,kernel_shape:[C],pads:k?Array.from(a().subarray(k>>>0,V>>>0)):[],strides:[F],w_is_const:()=>!!r()[J>>>0],activation:He(ce),activation_params:Re?Array.from(p().subarray(Re>>>0,ze>>>0)):[]})},838171:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we,Pe)=>{d.ob("Conv",n,{format:Re?"NHWC":"NCHW",auto_pad:l,dilations:[f,b],group:C,kernel_shape:[k,V],pads:F?Array.from(a().subarray(F>>>0,re>>>0)):[],strides:[J,ce],w_is_const:()=>!!r()[ze>>>0],activation:He(z),activation_params:we?Array.from(p().subarray(we>>>0,Pe>>>0)):[]})},838562:n=>{d.ob("Gelu",n,void 0)},838614:(n,l,f,b,C,k)=>{d.ob("MatMulNBits",n,{k:l,n:f,accuracyLevel:b,bits:C,blockSize:k})},838741:(n,l,f,b,C,k)=>{d.ob("MultiHeadAttention",n,{numHeads:l,isUnidirectional:f,maskFilterValue:b,scale:C,doRotary:k})},838900:(n,l,f,b,C)=>{d.ob("RotaryEmbedding",n,{interleaved:!!l,numHeads:f,rotaryEmbeddingDim:b,scale:C})},839039:(n,l,f)=>{d.ob("SkipLayerNormalization",n,{epsilon:l,simplified:!!f})},839141:(n,l,f)=>{d.ob("SkipLayerNormalization",n,{epsilon:l,simplified:!!f})},839243:(n,l,f,b)=>{d.ob("LayerNormalization",n,{axis:l,epsilon:f,simplified:!!b})},839354:n=>{d.ac(n)},839388:(n,l)=>d.cc(n,l,d.Cb.ec,d.Cb.errors)};function st(n){this.name="ExitStatus",this.message=`Program terminated with exit(${n})`,this.status=n}var Tr=n=>{n.terminate(),n.onmessage=()=>{}},ht=n=>{ge.xb.length==0&&(lr(),ge.Mb(ge.xb[0]));var l=ge.xb.pop();if(!l)return 6;ge.yb.push(l),ge.ub[n.wb]=l,l.wb=n.wb;var f={cmd:"run",start_routine:n.fc,arg:n.Rb,pthread_ptr:n.wb};return P&&l.unref(),l.postMessage(f,n.mc),0},bt=0,Er=n=>{var l=io();return n=n(),fn(l),n},Te=(n,l,...f)=>Er(()=>{for(var b=2*f.length,C=oo(8*b),k=C>>>3,V=0;V<f.length;V++){var F=f[V];typeof F=="bigint"?(Be[k+2*V]=1n,Be[k+2*V+1]=F):(Be[k+2*V]=0n,h()[k+2*V+1>>>0]=F)}return gi(n,0,b,C,l)});function dr(n){if(O)return Te(0,1,n);ie=n,0<bt||(ge.hc(),d.onExit?.(n),me=!0),I(n,new st(n))}var Xe=n=>{if(ie=n,O)throw jr(n),"unwind";dr(n)};function ut(){for(var n=d.numThreads;n--;)lr();mt.unshift(()=>{Vt++,Fn(()=>xr())})}function lr(){var n=L("ort-wasm-simd-threaded.worker.js");n=new Worker(n),ge.xb.push(n)}function Fn(n){O?n():Promise.all(ge.xb.map(ge.Mb)).then(n)}var ge={xb:[],yb:[],Qb:[],ub:{},Gb(){O?(ge.receiveObjectTransfer=ge.dc,ge.threadInitTLS=ge.Pb,ge.setExitStatus=ge.Ob):ut()},Ob:n=>ie=n,pc:["$terminateWorker"],hc:()=>{for(var n of ge.yb)Tr(n);for(n of ge.xb)Tr(n);ge.xb=[],ge.yb=[],ge.ub=[]},Nb:n=>{var l=n.wb;delete ge.ub[l],ge.xb.push(n),ge.yb.splice(ge.yb.indexOf(n),1),n.wb=0,ro(l)},dc(){},Pb(){ge.Qb.forEach(n=>n())},Mb:n=>new Promise(l=>{n.onmessage=k=>{k=k.data;var V=k.cmd;if(k.targetThread&&k.targetThread!=pn()){var F=ge.ub[k.targetThread];F?F.postMessage(k,k.transferList):ee(`Internal error! Worker sent a message "${V}" to target pthread ${k.targetThread}, but that thread no longer exists!`)}else V==="checkMailbox"?rt():V==="spawnThread"?ht(k):V==="cleanupThread"?ge.Nb(ge.ub[k.thread]):V==="killThread"?(k=k.thread,V=ge.ub[k],delete ge.ub[k],Tr(V),ro(k),ge.yb.splice(ge.yb.indexOf(V),1),V.wb=0):V==="cancelThread"?ge.ub[k.thread].postMessage({cmd:"cancel"}):V==="loaded"?(n.loaded=!0,P&&!n.wb&&n.unref(),l(n)):V==="alert"?alert(`Thread ${k.threadId}: ${k.text}`):k.target==="setimmediate"?n.postMessage(k):V==="callHandler"?d[k.handler](...k.args):V&&ee(`worker sent an unknown command ${V}`)},n.onerror=k=>{throw ee(`worker sent an error! ${k.filename}:${k.lineno}: ${k.message}`),k},P&&(n.on("message",k=>n.onmessage({data:k})),n.on("error",k=>n.onerror(k)));var f=[],b=["onExit"],C;for(C of b)d.hasOwnProperty(C)&&f.push(C);n.postMessage({cmd:"load",handlers:f,urlOrBlob:d.mainScriptUrlOrBlob||e,wasmMemory:ae,wasmModule:Ae})})};d.PThread=ge;var Yt=n=>{for(;0<n.length;)n.shift()(d)};d.establishStackSpace=()=>{var n=pn(),l=c()[n+52>>>2>>>0];n=c()[n+56>>>2>>>0],bi(l,l-n),fn(l)};function jr(n){if(O)return Te(1,0,n);Xe(n)}d.invokeEntryPoint=(n,l)=>{bt=0,n=wi(n,l),0<bt?ge.Ob(n):no(n)};class qn{constructor(l){this.Ib=l-24}Gb(l,f){c()[this.Ib+16>>>2>>>0]=0,c()[this.Ib+4>>>2>>>0]=l,c()[this.Ib+8>>>2>>>0]=f}}var Kr=0,jn=0;function Yr(n,l,f,b){return O?Te(2,1,n,l,f,b):Zr(n,l,f,b)}function Zr(n,l,f,b){if(n>>>=0,l>>>=0,f>>>=0,b>>>=0,typeof SharedArrayBuffer>"u")return ee("Current environment does not support SharedArrayBuffer, pthreads are not available!"),6;var C=[];return O&&C.length===0?Yr(n,l,f,b):(n={fc:f,wb:n,Rb:b,mc:C},O?(n.oc="spawnThread",postMessage(n,C),0):ht(n))}var cr=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,Pr=(n,l,f)=>{l>>>=0;var b=l+f;for(f=l;n[f]&&!(f>=b);)++f;if(16<f-l&&n.buffer&&cr)return cr.decode(n.buffer instanceof SharedArrayBuffer?n.slice(l,f):n.subarray(l,f));for(b="";l<f;){var C=n[l++];if(C&128){var k=n[l++]&63;if((C&224)==192)b+=String.fromCharCode((C&31)<<6|k);else{var V=n[l++]&63;C=(C&240)==224?(C&15)<<12|k<<6|V:(C&7)<<18|k<<12|V<<6|n[l++]&63,65536>C?b+=String.fromCharCode(C):(C-=65536,b+=String.fromCharCode(55296|C>>10,56320|C&1023))}}else b+=String.fromCharCode(C)}return b},He=(n,l)=>(n>>>=0)?Pr(o(),n,l):"";function Xr(n,l,f){return O?Te(3,1,n,l,f):0}function Qr(n,l){if(O)return Te(4,1,n,l)}var gt=n=>{for(var l=0,f=0;f<n.length;++f){var b=n.charCodeAt(f);127>=b?l++:2047>=b?l+=2:55296<=b&&57343>=b?(l+=4,++f):l+=3}return l},dt=(n,l,f,b)=>{if(f>>>=0,!(0<b))return 0;var C=f;b=f+b-1;for(var k=0;k<n.length;++k){var V=n.charCodeAt(k);if(55296<=V&&57343>=V){var F=n.charCodeAt(++k);V=65536+((V&1023)<<10)|F&1023}if(127>=V){if(f>=b)break;l[f++>>>0]=V}else{if(2047>=V){if(f+1>=b)break;l[f++>>>0]=192|V>>6}else{if(65535>=V){if(f+2>=b)break;l[f++>>>0]=224|V>>12}else{if(f+3>=b)break;l[f++>>>0]=240|V>>18,l[f++>>>0]=128|V>>12&63}l[f++>>>0]=128|V>>6&63}l[f++>>>0]=128|V&63}}return l[f>>>0]=0,f-C},Bt=(n,l,f)=>dt(n,o(),l,f);function Zt(n,l){if(O)return Te(5,1,n,l)}function kr(n,l,f){if(O)return Te(6,1,n,l,f)}function Or(n,l,f){return O?Te(7,1,n,l,f):0}function Jr(n,l){if(O)return Te(8,1,n,l)}function pr(n,l,f){if(O)return Te(9,1,n,l,f)}function en(n,l,f,b){if(O)return Te(10,1,n,l,f,b)}function tn(n,l,f,b){if(O)return Te(11,1,n,l,f,b)}function rn(n,l,f,b){if(O)return Te(12,1,n,l,f,b)}function nn(n){if(O)return Te(13,1,n)}function Rr(n,l){if(O)return Te(14,1,n,l)}function Xt(n,l,f){if(O)return Te(15,1,n,l,f)}var on,nt=n=>{for(var l="";o()[n>>>0];)l+=on[o()[n++>>>0]];return l},mr={},Br={},Kn={},wt;function Yn(n,l,f={}){var b=l.name;if(!n)throw new wt(`type "${b}" must have a positive integer typeid pointer`);if(Br.hasOwnProperty(n)){if(f.Ub)return;throw new wt(`Cannot register type \'${b}\' twice`)}Br[n]=l,delete Kn[n],mr.hasOwnProperty(n)&&(l=mr[n],delete mr[n],l.forEach(C=>C()))}function Je(n,l,f={}){if(!("argPackAdvance"in l))throw new TypeError("registerType registeredInstance requires argPackAdvance");return Yn(n,l,f)}var Dr=(n,l,f)=>{switch(l){case 1:return f?b=>r()[b>>>0]:b=>o()[b>>>0];case 2:return f?b=>i()[b>>>1>>>0]:b=>u()[b>>>1>>>0];case 4:return f?b=>a()[b>>>2>>>0]:b=>c()[b>>>2>>>0];case 8:return f?b=>Be[b>>>3]:b=>Ge[b>>>3];default:throw new TypeError(`invalid integer width (${l}): ${n}`)}},fr=[],lt=[];function hr(n){n>>>=0,9<n&&--lt[n+1]===0&&(lt[n]=void 0,fr.push(n))}var Qe=n=>{if(!n)throw new wt("Cannot use deleted val. handle = "+n);return lt[n]},ot=n=>{switch(n){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let l=fr.pop()||lt.length;return lt[l]=n,lt[l+1]=1,l}};function Qt(n){return this.fromWireType(c()[n>>>2>>>0])}var Zn={name:"emscripten::val",fromWireType:n=>{var l=Qe(n);return hr(n),l},toWireType:(n,l)=>ot(l),argPackAdvance:8,readValueFromPointer:Qt,Bb:null},an=(n,l)=>{switch(l){case 4:return function(f){return this.fromWireType(p()[f>>>2>>>0])};case 8:return function(f){return this.fromWireType(h()[f>>>3>>>0])};default:throw new TypeError(`invalid float width (${l}): ${n}`)}},zr=typeof TextDecoder<"u"?new TextDecoder("utf-16le"):void 0,Xn=(n,l)=>{for(var f=n>>1,b=f+l/2;!(f>=b)&&u()[f>>>0];)++f;if(f<<=1,32<f-n&&zr)return zr.decode(o().slice(n,f));for(f="",b=0;!(b>=l/2);++b){var C=i()[n+2*b>>>1>>>0];if(C==0)break;f+=String.fromCharCode(C)}return f},sn=(n,l,f)=>{if(f??=2147483647,2>f)return 0;f-=2;var b=l;f=f<2*n.length?f/2:n.length;for(var C=0;C<f;++C){var k=n.charCodeAt(C);i()[l>>>1>>>0]=k,l+=2}return i()[l>>>1>>>0]=0,l-b},un=n=>2*n.length,gr=(n,l)=>{for(var f=0,b="";!(f>=l/4);){var C=a()[n+4*f>>>2>>>0];if(C==0)break;++f,65536<=C?(C-=65536,b+=String.fromCharCode(55296|C>>10,56320|C&1023)):b+=String.fromCharCode(C)}return b},Qn=(n,l,f)=>{if(l>>>=0,f??=2147483647,4>f)return 0;var b=l;f=b+f-4;for(var C=0;C<n.length;++C){var k=n.charCodeAt(C);if(55296<=k&&57343>=k){var V=n.charCodeAt(++C);k=65536+((k&1023)<<10)|V&1023}if(a()[l>>>2>>>0]=k,l+=4,l+4>f)break}return a()[l>>>2>>>0]=0,l-b},pe=n=>{for(var l=0,f=0;f<n.length;++f){var b=n.charCodeAt(f);55296<=b&&57343>=b&&++f,l+=4}return l},Mr=n=>{if(!me)try{if(n(),!(0<bt))try{O?no(ie):Xe(ie)}catch(l){l instanceof st||l=="unwind"||I(1,l)}}catch(l){l instanceof st||l=="unwind"||I(1,l)}};function Nt(n){n>>>=0,typeof Atomics.nc=="function"&&(Atomics.nc(a(),n>>>2,n).value.then(rt),n+=128,Atomics.store(a(),n>>>2,1))}d.__emscripten_thread_mailbox_await=Nt;var rt=()=>{var n=pn();n&&(Nt(n),Mr(yi))};d.checkMailbox=rt;var yr=[],br=(n,l)=>{var f=Br[n];if(f===void 0)throw n=fi(n),f=nt(n),At(n),new wt(`${l} has unknown type ${f}`);return f},Ur=(n,l,f)=>{var b=[];return n=n.toWireType(b,f),b.length&&(c()[l>>>2>>>0]=ot(b)),n},Jt=n=>{try{n()}catch(l){Kt(l)}};function dn(){var n=oe,l={};for(let[f,b]of Object.entries(n))l[f]=typeof b=="function"?(...C)=>{Dt.push(f);try{return b(...C)}finally{me||(Dt.pop(),ct&&vt===1&&Dt.length===0&&(vt=0,bt+=1,Jt($i),typeof Fibers<"u"&&Fibers.rc()))}}:b;return l}var vt=0,ct=null,ln=0,Dt=[],Vr={},s={},m=0,g=null,$=[];function T(){return new Promise((n,l)=>{g={resolve:n,reject:l}})}function B(){var n=mn(65548),l=n+12;c()[n>>>2>>>0]=l,c()[n+4>>>2>>>0]=l+65536,l=Dt[0];var f=Vr[l];return f===void 0&&(f=m++,Vr[l]=f,s[f]=l),l=f,a()[n+8>>>2>>>0]=l,n}function H(){var n=a()[ct+8>>>2>>>0];return n=oe[s[n]],--bt,n()}function q(n){if(!me){if(vt===0){var l=!1,f=!1;n((b=0)=>{if(!me&&(ln=b,l=!0,f)){vt=2,Jt(()=>_i(ct)),typeof Browser<"u"&&Browser.Hb.Tb&&Browser.Hb.resume(),b=!1;try{var C=H()}catch(F){C=F,b=!0}var k=!1;if(!ct){var V=g;V&&(g=null,(b?V.reject:V.resolve)(C),k=!0)}if(b&&!k)throw C}}),f=!0,l||(vt=1,ct=B(),typeof Browser<"u"&&Browser.Hb.Tb&&Browser.Hb.pause(),Jt(()=>vi(ct)))}else vt===2?(vt=0,Jt(Si),At(ct),ct=null,$.forEach(Mr)):Kt(`invalid state: ${vt}`);return ln}}function te(n){return q(l=>{n().then(l)})}var X=[],de={},Ee=n=>{var l=de[n];return l===void 0?nt(n):l},Oe=()=>typeof globalThis=="object"?globalThis:Function("return this")(),D=n=>{var l=X.length;return X.push(n),l},be=(n,l)=>{for(var f=Array(n),b=0;b<n;++b)f[b]=br(c()[l+4*b>>>2>>>0],"parameter "+b);return f},Ie=(n,l)=>Object.defineProperty(l,"name",{value:n});function zt(n){var l=Function;if(!(l instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof l} which is not a function`);var f=Ie(l.name||"unknownFunctionName",function(){});return f.prototype=l.prototype,f=new f,n=l.apply(f,n),n instanceof Object?n:f}var Mt=n=>n%4===0&&(n%100!==0||n%400===0),Qo=[0,31,60,91,121,152,182,213,244,274,305,335],Jo=[0,31,59,90,120,151,181,212,243,273,304,334];function ei(n,l,f,b,C,k,V){return O?Te(16,1,n,l,f,b,C,k,V):-52}function ti(n,l,f,b,C,k){if(O)return Te(17,1,n,l,f,b,C,k)}var Jn=[],ri=(n,l)=>{Jn.length=0;for(var f;f=o()[n++>>>0];){var b=f!=105;b&=f!=112,l+=b&&l%8?4:0,Jn.push(f==112?c()[l>>>2>>>0]:f==106?Be[l>>>3]:f==105?a()[l>>>2>>>0]:h()[l>>>3>>>0]),l+=b?8:4}return Jn},eo={},ni=()=>{if(!to){var n={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(typeof navigator=="object"&&navigator.languages&&navigator.languages[0]||"C").replace("-","_")+".UTF-8",_:A||"./this.program"},l;for(l in eo)eo[l]===void 0?delete n[l]:n[l]=eo[l];var f=[];for(l in n)f.push(`${l}=${n[l]}`);to=f}return to},to;function oi(n,l){if(O)return Te(18,1,n,l);n>>>=0,l>>>=0;var f=0;return ni().forEach((b,C)=>{var k=l+f;for(C=c()[n+4*C>>>2>>>0]=k,k=0;k<b.length;++k)r()[C++>>>0]=b.charCodeAt(k);r()[C>>>0]=0,f+=b.length+1}),0}function ii(n,l){if(O)return Te(19,1,n,l);n>>>=0,l>>>=0;var f=ni();c()[n>>>2>>>0]=f.length;var b=0;return f.forEach(C=>b+=C.length+1),c()[l>>>2>>>0]=b,0}function ai(n){return O?Te(20,1,n):52}function si(n,l,f,b){return O?Te(21,1,n,l,f,b):52}function ui(n,l,f,b){return O?Te(22,1,n,l,f,b):70}var Xd=[null,[],[]];function di(n,l,f,b){if(O)return Te(23,1,n,l,f,b);l>>>=0,f>>>=0,b>>>=0;for(var C=0,k=0;k<f;k++){var V=c()[l>>>2>>>0],F=c()[l+4>>>2>>>0];l+=8;for(var re=0;re<F;re++){var J=o()[V+re>>>0],ce=Xd[n];J===0||J===10?((n===1?We:ee)(Pr(ce,0)),ce.length=0):ce.push(J)}C+=F}return c()[b>>>2>>>0]=C,0}var li=[31,29,31,30,31,30,31,31,30,31,30,31],ci=[31,28,31,30,31,30,31,31,30,31,30,31];function Qd(n){var l=Array(gt(n)+1);return dt(n,l,0,l.length),l}var Jd=(n,l)=>{r().set(n,l>>>0)};function pi(n,l,f,b){function C(z,we,Pe){for(z=typeof z=="number"?z.toString():z||"";z.length<we;)z=Pe[0]+z;return z}function k(z,we){return C(z,we,"0")}function V(z,we){function Pe(Ci){return 0>Ci?-1:0<Ci?1:0}var er;return(er=Pe(z.getFullYear()-we.getFullYear()))===0&&(er=Pe(z.getMonth()-we.getMonth()))===0&&(er=Pe(z.getDate()-we.getDate())),er}function F(z){switch(z.getDay()){case 0:return new Date(z.getFullYear()-1,11,29);case 1:return z;case 2:return new Date(z.getFullYear(),0,3);case 3:return new Date(z.getFullYear(),0,2);case 4:return new Date(z.getFullYear(),0,1);case 5:return new Date(z.getFullYear()-1,11,31);case 6:return new Date(z.getFullYear()-1,11,30)}}function re(z){var we=z.zb;for(z=new Date(new Date(z.Ab+1900,0,1).getTime());0<we;){var Pe=z.getMonth(),er=(Mt(z.getFullYear())?li:ci)[Pe];if(we>er-z.getDate())we-=er-z.getDate()+1,z.setDate(1),11>Pe?z.setMonth(Pe+1):(z.setMonth(0),z.setFullYear(z.getFullYear()+1));else{z.setDate(z.getDate()+we);break}}return Pe=new Date(z.getFullYear()+1,0,4),we=F(new Date(z.getFullYear(),0,4)),Pe=F(Pe),0>=V(we,z)?0>=V(Pe,z)?z.getFullYear()+1:z.getFullYear():z.getFullYear()-1}n>>>=0,l>>>=0,f>>>=0,b>>>=0;var J=c()[b+40>>>2>>>0];b={kc:a()[b>>>2>>>0],jc:a()[b+4>>>2>>>0],Eb:a()[b+8>>>2>>>0],Jb:a()[b+12>>>2>>>0],Fb:a()[b+16>>>2>>>0],Ab:a()[b+20>>>2>>>0],vb:a()[b+24>>>2>>>0],zb:a()[b+28>>>2>>>0],qc:a()[b+32>>>2>>>0],ic:a()[b+36>>>2>>>0],lc:J?He(J):""},f=He(f),J={"%c":"%a %b %d %H:%M:%S %Y","%D":"%m/%d/%y","%F":"%Y-%m-%d","%h":"%b","%r":"%I:%M:%S %p","%R":"%H:%M","%T":"%H:%M:%S","%x":"%m/%d/%y","%X":"%H:%M:%S","%Ec":"%c","%EC":"%C","%Ex":"%m/%d/%y","%EX":"%H:%M:%S","%Ey":"%y","%EY":"%Y","%Od":"%d","%Oe":"%e","%OH":"%H","%OI":"%I","%Om":"%m","%OM":"%M","%OS":"%S","%Ou":"%u","%OU":"%U","%OV":"%V","%Ow":"%w","%OW":"%W","%Oy":"%y"};for(var ce in J)f=f.replace(new RegExp(ce,"g"),J[ce]);var Re="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),ze="January February March April May June July August September October November December".split(" ");J={"%a":z=>Re[z.vb].substring(0,3),"%A":z=>Re[z.vb],"%b":z=>ze[z.Fb].substring(0,3),"%B":z=>ze[z.Fb],"%C":z=>k((z.Ab+1900)/100|0,2),"%d":z=>k(z.Jb,2),"%e":z=>C(z.Jb,2," "),"%g":z=>re(z).toString().substring(2),"%G":re,"%H":z=>k(z.Eb,2),"%I":z=>(z=z.Eb,z==0?z=12:12<z&&(z-=12),k(z,2)),"%j":z=>{for(var we=0,Pe=0;Pe<=z.Fb-1;we+=(Mt(z.Ab+1900)?li:ci)[Pe++]);return k(z.Jb+we,3)},"%m":z=>k(z.Fb+1,2),"%M":z=>k(z.jc,2),"%n":()=>`\n`,"%p":z=>0<=z.Eb&&12>z.Eb?"AM":"PM","%S":z=>k(z.kc,2),"%t":()=>"	","%u":z=>z.vb||7,"%U":z=>k(Math.floor((z.zb+7-z.vb)/7),2),"%V":z=>{var we=Math.floor((z.zb+7-(z.vb+6)%7)/7);if(2>=(z.vb+371-z.zb-2)%7&&we++,we)we==53&&(Pe=(z.vb+371-z.zb)%7,Pe==4||Pe==3&&Mt(z.Ab)||(we=1));else{we=52;var Pe=(z.vb+7-z.zb-1)%7;(Pe==4||Pe==5&&Mt(z.Ab%400-1))&&we++}return k(we,2)},"%w":z=>z.vb,"%W":z=>k(Math.floor((z.zb+7-(z.vb+6)%7)/7),2),"%y":z=>(z.Ab+1900).toString().substring(2),"%Y":z=>z.Ab+1900,"%z":z=>{z=z.ic;var we=0<=z;return z=Math.abs(z)/60,(we?"+":"-")+("0000"+(z/60*100+z%60)).slice(-4)},"%Z":z=>z.lc,"%%":()=>"%"},f=f.replace(/%%/g,"\\0\\0");for(ce in J)f.includes(ce)&&(f=f.replace(new RegExp(ce,"g"),J[ce](b)));return f=f.replace(/\\0\\0/g,"%"),ce=Qd(f),ce.length>l?0:(Jd(ce,n),ce.length-1)}ge.Gb();for(var mi=Array(256),cn=0;256>cn;++cn)mi[cn]=String.fromCharCode(cn);on=mi,wt=d.BindingError=class extends Error{constructor(n){super(n),this.name="BindingError"}},d.InternalError=class extends Error{constructor(n){super(n),this.name="InternalError"}},lt.push(0,1,void 0,1,null,1,!0,1,!1,1),d.count_emval_handles=()=>lt.length/2-5-fr.length;var el=[dr,jr,Yr,Xr,Qr,Zt,kr,Or,Jr,pr,en,tn,rn,nn,Rr,Xt,ei,ti,oi,ii,ai,si,ui,di],tl={ua:function(n,l,f){return te(async()=>{await d.Zb(n,l,f)})},b:function(n,l,f){throw n>>>=0,new qn(n).Gb(l>>>0,f>>>0),Kr=n,jn++,Kr},ia:function(n){hi(n>>>0,!E,1,!x,131072,!1),ge.Pb()},G:function(n){n>>>=0,O?postMessage({cmd:"cleanupThread",thread:n}):ge.Nb(ge.ub[n])},_:Zr,A:Xr,pa:Qr,W:Zt,Y:kr,qa:Or,na:Jr,fa:pr,ma:en,K:tn,X:rn,U:nn,oa:Rr,V:Xt,E:function(n,l,f){n>>>=0,f>>>=0,l=nt(l>>>0),Je(n,{name:l,fromWireType:b=>b,toWireType:function(b,C){if(typeof C!="bigint"&&typeof C!="number")throw C===null?C="null":(b=typeof C,C=b==="object"||b==="array"||b==="function"?C.toString():""+C),new TypeError(`Cannot convert "${C}" to ${this.name}`);return typeof C=="number"&&(C=BigInt(C)),C},argPackAdvance:8,readValueFromPointer:Dr(l,f,l.indexOf("u")==-1),Bb:null})},O:function(n,l,f,b){n>>>=0,l=nt(l>>>0),Je(n,{name:l,fromWireType:function(C){return!!C},toWireType:function(C,k){return k?f:b},argPackAdvance:8,readValueFromPointer:function(C){return this.fromWireType(o()[C>>>0])},Bb:null})},N:function(n){return Je(n>>>0,Zn)},D:function(n,l,f){n>>>=0,f>>>=0,l=nt(l>>>0),Je(n,{name:l,fromWireType:b=>b,toWireType:(b,C)=>C,argPackAdvance:8,readValueFromPointer:an(l,f),Bb:null})},t:function(n,l,f,b,C){if(n>>>=0,f>>>=0,l=nt(l>>>0),C===-1&&(C=4294967295),C=F=>F,b===0){var k=32-8*f;C=F=>F<<k>>>k}var V=l.includes("unsigned")?function(F,re){return re>>>0}:function(F,re){return re};Je(n,{name:l,fromWireType:C,toWireType:V,argPackAdvance:8,readValueFromPointer:Dr(l,f,b!==0),Bb:null})},p:function(n,l,f){function b(k){var V=c()[k>>>2>>>0];return k=c()[k+4>>>2>>>0],new C(r().buffer,k,V)}n>>>=0;var C=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][l];f=nt(f>>>0),Je(n,{name:f,fromWireType:b,argPackAdvance:8,readValueFromPointer:b},{Ub:!0})},F:function(n,l){n>>>=0,l=nt(l>>>0);var f=l==="std::string";Je(n,{name:l,fromWireType:function(b){var C=c()[b>>>2>>>0],k=b+4;if(f)for(var V=k,F=0;F<=C;++F){var re=k+F;if(F==C||o()[re>>>0]==0){if(V=He(V,re-V),J===void 0)var J=V;else J+=String.fromCharCode(0),J+=V;V=re+1}}else{for(J=Array(C),F=0;F<C;++F)J[F]=String.fromCharCode(o()[k+F>>>0]);J=J.join("")}return At(b),J},toWireType:function(b,C){C instanceof ArrayBuffer&&(C=new Uint8Array(C));var k=typeof C=="string";if(!(k||C instanceof Uint8Array||C instanceof Uint8ClampedArray||C instanceof Int8Array))throw new wt("Cannot pass non-string to std::string");var V=f&&k?gt(C):C.length,F=mn(4+V+1),re=F+4;if(c()[F>>>2>>>0]=V,f&&k)Bt(C,re,V+1);else if(k)for(k=0;k<V;++k){var J=C.charCodeAt(k);if(255<J)throw At(re),new wt("String has UTF-16 code units that do not fit in 8 bits");o()[re+k>>>0]=J}else for(k=0;k<V;++k)o()[re+k>>>0]=C[k];return b!==null&&b.push(At,F),F},argPackAdvance:8,readValueFromPointer:Qt,Bb(b){At(b)}})},z:function(n,l,f){if(n>>>=0,l>>>=0,f>>>=0,f=nt(f),l===2)var b=Xn,C=sn,k=un,V=F=>u()[F>>>1>>>0];else l===4&&(b=gr,C=Qn,k=pe,V=F=>c()[F>>>2>>>0]);Je(n,{name:f,fromWireType:F=>{for(var re=c()[F>>>2>>>0],J,ce=F+4,Re=0;Re<=re;++Re){var ze=F+4+Re*l;(Re==re||V(ze)==0)&&(ce=b(ce,ze-ce),J===void 0?J=ce:(J+=String.fromCharCode(0),J+=ce),ce=ze+l)}return At(F),J},toWireType:(F,re)=>{if(typeof re!="string")throw new wt(`Cannot pass non-string to C++ string type ${f}`);var J=k(re),ce=mn(4+J+l);return c()[ce>>>2>>>0]=J/l,C(re,ce+4,J+l),F!==null&&F.push(At,ce),ce},argPackAdvance:8,readValueFromPointer:Qt,Bb(F){At(F)}})},P:function(n,l){n>>>=0,l=nt(l>>>0),Je(n,{Vb:!0,name:l,argPackAdvance:0,fromWireType:()=>{},toWireType:()=>{}})},ta:()=>1,S:function(n,l){n>>>=0,n==l>>>0?setTimeout(rt):O?postMessage({targetThread:n,cmd:"checkMailbox"}):(n=ge.ub[n])&&n.postMessage({cmd:"checkMailbox"})},$:function(n,l,f,b,C){l>>>=0,f>>>=0,b/=2,yr.length=b,C=C>>>0>>>3;for(var k=0;k<b;k++)yr[k]=Be[C+2*k]?Be[C+2*k+1]:h()[C+2*k+1>>>0];return n=l?Ir[l]:el[n],ge.Sb=f,f=n(...yr),ge.Sb=0,f},ha:Nt,sa:function(n){P&&ge.ub[n>>>0].ref()},w:function(n,l,f){return l>>>=0,f>>>=0,n=Qe(n>>>0),l=br(l,"emval::as"),Ur(l,f,n)},y:function(n){return n>>>=0,te(()=>(n=Qe(n),n.then(ot)))},o:function(n,l,f,b){return f>>>=0,b>>>=0,n=X[n>>>0],l=Qe(l>>>0),n(null,l,f,b)},k:function(n,l,f,b,C){return f>>>=0,b>>>=0,C>>>=0,n=X[n>>>0],l=Qe(l>>>0),f=Ee(f),n(l,l[f],b,C)},c:hr,C:function(n,l){return l>>>=0,n=Qe(n>>>0),l=Qe(l),n==l},n:function(n){return n>>>=0,n===0?ot(Oe()):(n=Ee(n),ot(Oe()[n]))},j:function(n,l,f){l=be(n,l>>>0);var b=l.shift();n--;var C=`return function (obj, func, destructorsRef, args) {\n`,k=0,V=[];f===0&&V.push("obj");for(var F=["retType"],re=[b],J=0;J<n;++J)V.push("arg"+J),F.push("argType"+J),re.push(l[J]),C+=`  var arg${J} = argType${J}.readValueFromPointer(args${k?"+"+k:""});\n`,k+=l[J].argPackAdvance;return C+=`  var rv = ${f===1?"new func":"func.call"}(${V.join(", ")});\n`,b.Vb||(F.push("emval_returnValue"),re.push(Ur),C+=`  return emval_returnValue(retType, destructorsRef, rv);\n`),F.push(C+`};\n`),n=zt(F)(...re),f=`methodCaller<(${l.map(ce=>ce.name).join(", ")}) => ${b.name}>`,D(Ie(f,n))},s:function(n,l){return l>>>=0,n=Qe(n>>>0),l=Qe(l),ot(n[l])},d:function(n){n>>>=0,9<n&&(lt[n+1]+=1)},v:function(){return ot([])},m:function(n){n=Qe(n>>>0);for(var l=Array(n.length),f=0;f<n.length;f++)l[f]=n[f];return ot(l)},g:function(n){return ot(Ee(n>>>0))},l:function(){return ot({})},i:function(n){n>>>=0;for(var l=Qe(n);l.length;){var f=l.pop();l.pop()(f)}hr(n)},h:function(n,l,f){l>>>=0,f>>>=0,n=Qe(n>>>0),l=Qe(l),f=Qe(f),n[l]=f},e:function(n,l){return l>>>=0,n=br(n>>>0,"_emval_take_value"),n=n.readValueFromPointer(l),ot(n)},ca:function(n,l){n=-9007199254740992>n||9007199254740992<n?NaN:Number(n),l>>>=0,n=new Date(1e3*n),a()[l>>>2>>>0]=n.getUTCSeconds(),a()[l+4>>>2>>>0]=n.getUTCMinutes(),a()[l+8>>>2>>>0]=n.getUTCHours(),a()[l+12>>>2>>>0]=n.getUTCDate(),a()[l+16>>>2>>>0]=n.getUTCMonth(),a()[l+20>>>2>>>0]=n.getUTCFullYear()-1900,a()[l+24>>>2>>>0]=n.getUTCDay(),n=(n.getTime()-Date.UTC(n.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,a()[l+28>>>2>>>0]=n},da:function(n,l){n=-9007199254740992>n||9007199254740992<n?NaN:Number(n),l>>>=0,n=new Date(1e3*n),a()[l>>>2>>>0]=n.getSeconds(),a()[l+4>>>2>>>0]=n.getMinutes(),a()[l+8>>>2>>>0]=n.getHours(),a()[l+12>>>2>>>0]=n.getDate(),a()[l+16>>>2>>>0]=n.getMonth(),a()[l+20>>>2>>>0]=n.getFullYear()-1900,a()[l+24>>>2>>>0]=n.getDay();var f=(Mt(n.getFullYear())?Qo:Jo)[n.getMonth()]+n.getDate()-1|0;a()[l+28>>>2>>>0]=f,a()[l+36>>>2>>>0]=-(60*n.getTimezoneOffset()),f=new Date(n.getFullYear(),6,1).getTimezoneOffset();var b=new Date(n.getFullYear(),0,1).getTimezoneOffset();n=(f!=b&&n.getTimezoneOffset()==Math.min(b,f))|0,a()[l+32>>>2>>>0]=n},ea:function(n){n>>>=0;var l=new Date(a()[n+20>>>2>>>0]+1900,a()[n+16>>>2>>>0],a()[n+12>>>2>>>0],a()[n+8>>>2>>>0],a()[n+4>>>2>>>0],a()[n>>>2>>>0],0),f=a()[n+32>>>2>>>0],b=l.getTimezoneOffset(),C=new Date(l.getFullYear(),6,1).getTimezoneOffset(),k=new Date(l.getFullYear(),0,1).getTimezoneOffset(),V=Math.min(k,C);return 0>f?a()[n+32>>>2>>>0]=+(C!=k&&V==b):0<f!=(V==b)&&(C=Math.max(k,C),l.setTime(l.getTime()+6e4*((0<f?V:C)-b))),a()[n+24>>>2>>>0]=l.getDay(),f=(Mt(l.getFullYear())?Qo:Jo)[l.getMonth()]+l.getDate()-1|0,a()[n+28>>>2>>>0]=f,a()[n>>>2>>>0]=l.getSeconds(),a()[n+4>>>2>>>0]=l.getMinutes(),a()[n+8>>>2>>>0]=l.getHours(),a()[n+12>>>2>>>0]=l.getDate(),a()[n+16>>>2>>>0]=l.getMonth(),a()[n+20>>>2>>>0]=l.getYear(),n=l.getTime(),BigInt(isNaN(n)?-1:n/1e3)},aa:ei,ba:ti,R:function(n,l,f,b){n>>>=0,l>>>=0,f>>>=0,b>>>=0;var C=new Date().getFullYear(),k=new Date(C,0,1),V=new Date(C,6,1);C=k.getTimezoneOffset();var F=V.getTimezoneOffset(),re=Math.max(C,F);c()[n>>>2>>>0]=60*re,a()[l>>>2>>>0]=+(C!=F),n=J=>J.toLocaleTimeString(void 0,{hour12:!1,timeZoneName:"short"}).split(" ")[1],k=n(k),V=n(V),F<C?(Bt(k,f,17),Bt(V,b,17)):(Bt(k,b,17),Bt(V,f,17))},q:()=>{Kt("")},f:function(n,l,f){return n>>>=0,l=ri(l>>>0,f>>>0),Ir[n](...l)},M:function(n,l,f){return n>>>=0,l=ri(l>>>0,f>>>0),Ir[n](...l)},H:()=>{},L:()=>Date.now(),ra:()=>{throw bt+=1,"unwind"},T:function(){return 4294901760},u:()=>performance.timeOrigin+performance.now(),x:()=>P?(Ri(),wr(Oi)).cpus().length:navigator.hardwareConcurrency,Q:function(n){n>>>=0;var l=o().length;if(n<=l||4294901760<n)return!1;for(var f=1;4>=f;f*=2){var b=l*(1+.2/f);b=Math.min(b,n+100663296);var C=Math;b=Math.max(n,b);e:{C=(C.min.call(C,4294901760,b+(65536-b%65536)%65536)-ae.buffer.byteLength+65535)/65536;try{ae.grow(C),Ne();var k=1;break e}catch{}k=void 0}if(k)return!0}return!1},ja:oi,la:ii,Z:Xe,B:ai,J:si,ga:ui,I:di,a:ae||d.wasmMemory,ka:pi,r:function(n,l,f,b){return pi(n>>>0,l>>>0,f>>>0,b>>>0)}},oe=function(){function n(f,b){return oe=f.exports,oe=dn(),oe=rl(),ge.Qb.push(oe.cb),Rt.unshift(oe.va),Ae=b,xr(),oe}var l={a:tl};if(Vt++,d.instantiateWasm)try{return d.instantiateWasm(l,n)}catch(f){ee(`Module.instantiateWasm callback failed with error: ${f}`),w(f)}return Ar(l,function(f){n(f.instance,f.module)}).catch(w),{}}(),fi=n=>(fi=oe.wa)(n);d.__embind_initialize_bindings=()=>(d.__embind_initialize_bindings=oe.xa)(),d._OrtInit=(n,l)=>(d._OrtInit=oe.ya)(n,l),d._OrtGetLastError=(n,l)=>(d._OrtGetLastError=oe.za)(n,l),d._OrtCreateSessionOptions=(n,l,f,b,C,k,V,F,re,J)=>(d._OrtCreateSessionOptions=oe.Aa)(n,l,f,b,C,k,V,F,re,J),d._OrtAppendExecutionProvider=(n,l)=>(d._OrtAppendExecutionProvider=oe.Ba)(n,l),d._OrtAddFreeDimensionOverride=(n,l,f)=>(d._OrtAddFreeDimensionOverride=oe.Ca)(n,l,f),d._OrtAddSessionConfigEntry=(n,l,f)=>(d._OrtAddSessionConfigEntry=oe.Da)(n,l,f),d._OrtReleaseSessionOptions=n=>(d._OrtReleaseSessionOptions=oe.Ea)(n),d._OrtCreateSession=(n,l,f)=>(d._OrtCreateSession=oe.Fa)(n,l,f),d._OrtReleaseSession=n=>(d._OrtReleaseSession=oe.Ga)(n),d._OrtGetInputOutputCount=(n,l,f)=>(d._OrtGetInputOutputCount=oe.Ha)(n,l,f),d._OrtGetInputName=(n,l)=>(d._OrtGetInputName=oe.Ia)(n,l),d._OrtGetOutputName=(n,l)=>(d._OrtGetOutputName=oe.Ja)(n,l),d._OrtFree=n=>(d._OrtFree=oe.Ka)(n),d._OrtCreateTensor=(n,l,f,b,C,k)=>(d._OrtCreateTensor=oe.La)(n,l,f,b,C,k),d._OrtGetTensorData=(n,l,f,b,C)=>(d._OrtGetTensorData=oe.Ma)(n,l,f,b,C),d._OrtReleaseTensor=n=>(d._OrtReleaseTensor=oe.Na)(n),d._OrtCreateRunOptions=(n,l,f,b)=>(d._OrtCreateRunOptions=oe.Oa)(n,l,f,b),d._OrtAddRunConfigEntry=(n,l,f)=>(d._OrtAddRunConfigEntry=oe.Pa)(n,l,f),d._OrtReleaseRunOptions=n=>(d._OrtReleaseRunOptions=oe.Qa)(n),d._OrtCreateBinding=n=>(d._OrtCreateBinding=oe.Ra)(n),d._OrtBindInput=(n,l,f)=>(d._OrtBindInput=oe.Sa)(n,l,f),d._OrtBindOutput=(n,l,f,b)=>(d._OrtBindOutput=oe.Ta)(n,l,f,b),d._OrtClearBoundOutputs=n=>(d._OrtClearBoundOutputs=oe.Ua)(n),d._OrtReleaseBinding=n=>(d._OrtReleaseBinding=oe.Va)(n),d._OrtRunWithBinding=(n,l,f,b,C)=>(d._OrtRunWithBinding=oe.Wa)(n,l,f,b,C),d._OrtRun=(n,l,f,b,C,k,V,F)=>(d._OrtRun=oe.Xa)(n,l,f,b,C,k,V,F),d._OrtEndProfiling=n=>(d._OrtEndProfiling=oe.Ya)(n),d._JsepOutput=(n,l,f)=>(d._JsepOutput=oe.Za)(n,l,f),d._JsepGetNodeName=n=>(d._JsepGetNodeName=oe._a)(n);var pn=d._pthread_self=()=>(pn=d._pthread_self=oe.$a)(),mn=d._malloc=n=>(mn=d._malloc=oe.ab)(n),At=d._free=n=>(At=d._free=oe.bb)(n);d.__emscripten_tls_init=()=>(d.__emscripten_tls_init=oe.cb)();var hi=d.__emscripten_thread_init=(n,l,f,b,C,k)=>(hi=d.__emscripten_thread_init=oe.eb)(n,l,f,b,C,k);d.__emscripten_thread_crashed=()=>(d.__emscripten_thread_crashed=oe.fb)();var gi=(n,l,f,b,C)=>(gi=oe.gb)(n,l,f,b,C),ro=n=>(ro=oe.hb)(n),no=d.__emscripten_thread_exit=n=>(no=d.__emscripten_thread_exit=oe.ib)(n),yi=()=>(yi=oe.jb)(),bi=(n,l)=>(bi=oe.kb)(n,l),fn=n=>(fn=oe.lb)(n),oo=n=>(oo=oe.mb)(n),io=()=>(io=oe.nb)(),wi=d.dynCall_ii=(n,l)=>(wi=d.dynCall_ii=oe.pb)(n,l),vi=n=>(vi=oe.qb)(n),$i=()=>($i=oe.rb)(),_i=n=>(_i=oe.sb)(n),Si=()=>(Si=oe.tb)();d.___start_em_js=839500,d.___stop_em_js=839661;function rl(){var n=oe;n=Object.assign({},n);var l=b=>C=>b(C)>>>0,f=b=>()=>b()>>>0;return n.wa=l(n.wa),n.$a=f(n.$a),n.ab=l(n.ab),n.emscripten_main_runtime_thread_id=f(n.emscripten_main_runtime_thread_id),n.mb=l(n.mb),n.nb=f(n.nb),n}d.wasmMemory=ae,d.stackSave=()=>io(),d.stackRestore=n=>fn(n),d.stackAlloc=n=>oo(n),d.keepRuntimeAlive=()=>0<bt,d.UTF8ToString=He,d.stringToUTF8=Bt,d.lengthBytesUTF8=gt,d.ExitStatus=st,d.PThread=ge;var hn;jt=function n(){hn||xi(),hn||(jt=n)};function xi(){if(!(0<Vt))if(O)y(d),O||Yt(Rt),startWorker(d);else{if(d.preRun)for(typeof d.preRun=="function"&&(d.preRun=[d.preRun]);d.preRun.length;)mt.unshift(d.preRun.shift());Yt(mt),0<Vt||hn||(hn=!0,d.calledRun=!0,me||(O||Yt(Rt),y(d),O||Yt(qt)))}}return xi(),_}})();typeof Di=="object"&&typeof mo=="object"?mo.exports=Bi:typeof define=="function"&&define.amd&&define([],()=>Bi)});var Mi=Wr((gm,cl)=>{cl.exports=\'"use strict";var Module={},ENVIRONMENT_IS_NODE=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string";if(ENVIRONMENT_IS_NODE){var nodeWorkerThreads=require("worker_threads"),parentPort=nodeWorkerThreads.parentPort;parentPort.on("message",e=>onmessage({data:e}));var fs=require("fs"),vm=require("vm");Object.assign(global,{self:global,require,Module,location:{href:__filename},Worker:nodeWorkerThreads.Worker,importScripts:e=>vm.runInThisContext(fs.readFileSync(e,"utf8"),{filename:e}),postMessage:e=>parentPort.postMessage(e),performance:global.performance||{now:Date.now}})}var initializedJS=!1;function threadPrintErr(...e){var a=e.join(" ");if(ENVIRONMENT_IS_NODE){fs.writeSync(2,a+`\\n`);return}console.error(a)}function threadAlert(...e){var a=e.join(" ");postMessage({cmd:"alert",text:a,threadId:Module._pthread_self()})}var err=threadPrintErr;self.alert=threadAlert,Module.instantiateWasm=(e,a)=>{var r=Module.wasmModule;Module.wasmModule=null;var t=new WebAssembly.Instance(r,e);return a(t)},self.onunhandledrejection=e=>{throw e.reason||e};function handleMessage(e){try{if(e.data.cmd==="load"){let r=[];self.onmessage=t=>r.push(t),self.startWorker=t=>{Module=t,postMessage({cmd:"loaded"});for(let s of r)handleMessage(s);self.onmessage=handleMessage},Module.wasmModule=e.data.wasmModule;for(const t of e.data.handlers)Module[t]=(...s)=>{postMessage({cmd:"callHandler",handler:t,args:s})};if(Module.wasmMemory=e.data.wasmMemory,Module.buffer=Module.wasmMemory.buffer,Module.ENVIRONMENT_IS_PTHREAD=!0,typeof e.data.urlOrBlob=="string")importScripts(e.data.urlOrBlob);else{var a=URL.createObjectURL(e.data.urlOrBlob);importScripts(a),URL.revokeObjectURL(a)}ortWasmThreaded(Module)}else if(e.data.cmd==="run"){Module.__emscripten_thread_init(e.data.pthread_ptr,0,0,1),Module.__emscripten_thread_mailbox_await(e.data.pthread_ptr),Module.establishStackSpace(),Module.PThread.receiveObjectTransfer(e.data),Module.PThread.threadInitTLS(),initializedJS||(initializedJS=!0);try{Module.invokeEntryPoint(e.data.start_routine,e.data.arg)}catch(r){if(r!="unwind")throw r}}else e.data.cmd==="cancel"?Module._pthread_self()&&Module.__emscripten_thread_exit(-1):e.data.target==="setimmediate"||(e.data.cmd==="checkMailbox"?initializedJS&&Module.checkMailbox():e.data.cmd&&(err(`worker.js received unknown command ${e.data.cmd}`),err(e.data)))}catch(r){throw Module.__emscripten_thread_crashed?.(),r}}self.onmessage=handleMessage;\\n\'});var go,Gt,tr,bn,Gr,Hi,yo,ye=Y(()=>{"use strict";go=e=>{switch(e){case"int8":return 3;case"uint8":return 2;case"bool":return 9;case"int16":return 5;case"uint16":return 4;case"int32":return 6;case"uint32":return 12;case"float16":return 10;case"float32":return 1;case"float64":return 11;case"string":return 8;case"int64":return 7;case"uint64":return 13;default:throw new Error(`unsupported data type: ${e}`)}},Gt=e=>{switch(e){case 3:return"int8";case 2:return"uint8";case 9:return"bool";case 5:return"int16";case 4:return"uint16";case 6:return"int32";case 12:return"uint32";case 10:return"float16";case 1:return"float32";case 11:return"float64";case 8:return"string";case 7:return"int64";case 13:return"uint64";default:throw new Error(`unsupported data type: ${e}`)}},tr=e=>[void 0,4,1,1,2,2,4,8,void 0,1,2,8,4,8,void 0,void 0,void 0][e],bn=e=>{switch(e){case"float16":return typeof Float16Array<"u"&&Float16Array.from?Float16Array:Uint16Array;case"float32":return Float32Array;case"uint8":return Uint8Array;case"int8":return Int8Array;case"uint16":return Uint16Array;case"int16":return Int16Array;case"int32":return Int32Array;case"bool":return Uint8Array;case"float64":return Float64Array;case"uint32":return Uint32Array;case"int64":return BigInt64Array;case"uint64":return BigUint64Array;default:throw new Error(`unsupported type: ${e}`)}},Gr=e=>{switch(e){case"verbose":return 0;case"info":return 1;case"warning":return 2;case"error":return 3;case"fatal":return 4;default:throw new Error(`unsupported logging level: ${e}`)}},Hi=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint8"||e==="bool",yo=e=>{switch(e){case"none":return 0;case"cpu":return 1;case"cpu-pinned":return 2;case"texture":return 3;case"gpu-buffer":return 4;default:throw new Error(`unsupported data location: ${e}`)}}});var wn=Y(()=>{"use strict"});var Fi=Y(()=>{"use strict";wn()});var qi,ji=Y(()=>{"use strict";qi="1.18.0"});var Ki,pt,bo=Y(()=>{"use strict";ji();Ki="warning",pt={wasm:{},webgl:{},webgpu:{},versions:{common:qi},set logLevel(e){if(e!==void 0){if(typeof e!="string"||["verbose","info","warning","error","fatal"].indexOf(e)===-1)throw new Error(`Unsupported logging level: ${e}`);Ki=e}},get logLevel(){return Ki}};Object.defineProperty(pt,"logLevel",{enumerable:!0})});var vr,Yi=Y(()=>{"use strict";bo();vr=pt});var Zi=Y(()=>{"use strict"});var Xi=Y(()=>{"use strict";vn()});var Ji=Y(()=>{"use strict"});var ea=Y(()=>{"use strict";vn()});var vn=Y(()=>{"use strict";Zi();Xi();Ji();ea()});var $n=Y(()=>{"use strict";vn()});var wo,ta,rr,Ht,vo=Y(()=>{"use strict";bo();wo=(e,t)=>{(typeof pt.trace>"u"?!pt.wasm.trace:!pt.trace)||console.timeStamp(`${e}::ORT::${t}`)},ta=(e,t)=>{let r=new Error().stack?.split(/\\r\\n|\\r|\\n/g)||[],o=!1;for(let i=0;i<r.length;i++){if(o&&!r[i].includes("TRACE_FUNC")){let u=`FUNC_${e}::${r[i].trim().split(" ")[1]}`;t&&(u+=`::${t}`),wo("CPU",u);return}r[i].includes("TRACE_FUNC")&&(o=!0)}},rr=e=>{(typeof pt.trace>"u"?!pt.wasm.trace:!pt.trace)||ta("BEGIN",e)},Ht=e=>{(typeof pt.trace>"u"?!pt.wasm.trace:!pt.trace)||ta("END",e)}});var ra=Y(()=>{"use strict";wn();$n();vo()});var na=Y(()=>{"use strict";ra()});var oa=Y(()=>{"use strict"});var ia=Y(()=>{"use strict"});var aa=Y(()=>{"use strict"});var sa=Y(()=>{"use strict"});var ua=Y(()=>{"use strict";wn();$n()});var da=Y(()=>{"use strict";ua()});var $r=Y(()=>{"use strict";Fi();Yi();na();$n();oa();ia();vo();aa();sa();da()});var _l,Sl,la,ca,pa,xl,Ve,Lt=Y(()=>{"use strict";ye();_l=["V","I","W","E","F"],Sl=(e,t)=>{console.log(`[${_l[e]},${new Date().toISOString()}]${t}`)},pa=(e,t)=>{la=e,ca=t},xl=(e,t)=>{let r=Gr(e),o=Gr(la);r>=o&&Sl(r,typeof t=="function"?t():t)},Ve=(...e)=>{ca&&xl(...e)}});var ma,fa=Y(()=>{"use strict";ye();ma=(e,t)=>new(bn(t))(e)});var _n=Y(()=>{"use strict"});var ha,$o,_o,Cl,Al,ga,xo,So,ba,wa=Y(()=>{"use strict";Lt();_n();ha=new Map([[64,250],[128,200],[256,200],[512,200],[2048,230],[4096,200],[8192,50],[16384,50],[32768,50],[65536,50],[131072,50],[262144,50],[524288,50],[1048576,50],[2097152,30],[4194304,20],[8388608,10],[12582912,10],[16777216,10],[26214400,15],[33554432,22],[44236800,2],[58982400,6],[67108864,6],[134217728,6],[167772160,6]]),$o=[],_o=e=>Math.ceil(e/16)*16,Cl=e=>{for(let t=0;t<$o.length;t++){let r=$o[t];if(e<=r)return r}return Math.ceil(e/16)*16},Al=1,ga=()=>Al++,xo=async(e,t,r,o)=>{let i=_o(r),u=e.device.createBuffer({size:i,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let a=e.getCommandEncoder();e.endComputePass(),a.copyBufferToBuffer(t,0,u,0,i),e.flush(),await u.mapAsync(GPUMapMode.READ);let c=u.getMappedRange();if(o){let p=o();return p.set(new Uint8Array(c,0,r)),p}else return new Uint8Array(c.slice(0,r))}finally{u.destroy()}},So=class{constructor(t){this.backend=t;this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.buffersForUploadingPending=[],this.buffersPending=[],this.externalBuffers=new Map,this.capturedPendingBuffers=new Map;for(let[r]of ha)$o.push(r),this.freeBuffers.set(r,[]),this.freeUniformBuffers.set(r,[])}upload(t,r){let o=r.buffer,i=r.byteOffset,u=r.byteLength,a=_o(u),c=this.storageCache.get(t);if(!c)throw new Error("gpu data for uploading does not exist");if(c.originalSize!==u)throw new Error(`inconsistent data size. gpu data size=${c.originalSize}, data size=${u}`);let p=this.backend.device.createBuffer({mappedAtCreation:!0,size:a,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),h=p.getMappedRange();new Uint8Array(h).set(new Uint8Array(o,i,u)),p.unmap();let d=this.backend.getCommandEncoder();this.backend.endComputePass(),d.copyBufferToBuffer(p,0,c.gpuData.buffer,0,a),Ve("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${t})`),this.buffersForUploadingPending.push(p)}memcpy(t,r){let o=this.storageCache.get(t);if(!o)throw new Error("source gpu data for memcpy does not exist");let i=this.storageCache.get(r);if(!i)throw new Error("destination gpu data for memcpy does not exist");if(o.originalSize!==i.originalSize)throw new Error("inconsistent source and destination gpu data size");let u=_o(o.originalSize),a=this.backend.getCommandEncoder();this.backend.endComputePass(),a.copyBufferToBuffer(o.gpuData.buffer,0,i.gpuData.buffer,0,u)}registerExternalBuffer(t,r,o){let i;if(o){if(i=this.externalBuffers.get(o),i===void 0)throw new Error("previous buffer is not registered");if(t===o)return Ve("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${r}) => id=${i}, buffer is the same, skip.`),i;if(this.backend.capturedCommandList.has(this.backend.currentSessionId))throw new Error(`Registering a different external buffer under graph capture mode is not supported yet.\n             Please use the previous external buffer!`);this.externalBuffers.delete(o)}else i=ga();return this.storageCache.set(i,{gpuData:{id:i,type:0,buffer:t},originalSize:r}),this.externalBuffers.set(t,i),Ve("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${r}) => id=${i}, registered.`),i}unregisterExternalBuffer(t){let r=this.externalBuffers.get(t);r!==void 0&&(this.storageCache.delete(r),this.externalBuffers.delete(t),Ve("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${r}`))}create(t,r=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let o=Cl(t),i,u=(r&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,a=(r&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM;if(u||a){let h=(u?this.freeBuffers:this.freeUniformBuffers).get(o);h?h.length>0?i=h.pop():i=this.backend.device.createBuffer({size:o,usage:r}):i=this.backend.device.createBuffer({size:o,usage:r})}else i=this.backend.device.createBuffer({size:o,usage:r});let c={id:ga(),type:0,buffer:i};return this.storageCache.set(c.id,{gpuData:c,originalSize:t}),Ve("verbose",()=>`[WebGPU] GpuDataManager.create(size=${t}) => id=${c.id}`),c}get(t){return this.storageCache.get(t)?.gpuData}release(t){let r=this.storageCache.get(t);if(!r)throw new Error("releasing data does not exist");return Ve("verbose",()=>`[WebGPU] GpuDataManager.release(id=${t}), gpuDataId=${r.gpuData.id}`),this.storageCache.delete(t),this.buffersPending.push(r.gpuData.buffer),r.originalSize}async download(t,r){let o=this.storageCache.get(t);if(!o)throw new Error("data does not exist");await xo(this.backend,o.gpuData.buffer,o.originalSize,r)}refreshPendingBuffers(){for(let t of this.buffersForUploadingPending)t.destroy();if(this.buffersForUploadingPending=[],this.buffersPending.length!==0)if(this.backend.sessionStatus==="default"){for(let t of this.buffersPending){let r=ha.get(t.size);if((t.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let o=this.freeBuffers.get(t.size)||[];r===void 0||o.length>=r?t.destroy():o.push(t)}else if((t.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let o=this.freeUniformBuffers.get(t.size)||[];r===void 0||o.length>=r?t.destroy():o.push(t)}else t.destroy()}this.buffersPending=[]}else{let t=this.capturedPendingBuffers.get(this.backend.currentSessionId);t||(t=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,t));for(let r of this.buffersPending)t.push(r);this.buffersPending=[]}}dispose(){this.freeBuffers.forEach(t=>{t.forEach(r=>{r.destroy()})}),this.freeUniformBuffers.forEach(t=>{t.forEach(r=>{r.destroy()})}),this.storageCache.forEach(t=>{t.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(t=>{t.forEach(r=>{r.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onReleaseSession(t){let r=this.capturedPendingBuffers.get(t);r&&(r.forEach(o=>{o.destroy()}),this.capturedPendingBuffers.delete(t))}},ba=(...e)=>new So(...e)});var Co,ve,Ze=Y(()=>{"use strict";Co=class{constructor(t){Object.assign(this,t)}get cacheKey(){return this.key||(this.key=Object.getOwnPropertyNames(this).sort().map(t=>`${this[t]}`).join(";")),this.key}},ve=e=>new Co(e)});var Ao,It,M,nr,Sn,xn,Cn,Se=Y(()=>{"use strict";Ao=class{static calcMatMulShape(t,r){return t[1]!==r[0]?void 0:[t[0],r[1]]}},It=class{static calcShape(t,r,o=!1){let i=t.length,u=r.length;if(i===0)return r;if(u===0)return t;let a=Math.max(t.length,r.length),c=new Array(a);if(o){if(i<2||u<2)return;let p=Ao.calcMatMulShape([t[i-2],t[i-1]],[r[u-2],r[u-1]]);if(p===void 0)return;[c[a-2],c[a-1]]=p}for(let p=o?3:1;p<=a;p++){let h=i-p<0?1:t[i-p],d=u-p<0?1:r[u-p];if(h!==d&&h>1&&d>1)return;let y=Math.max(h,d);if(h&&d)c[a-p]=Math.max(h,d);else{if(y>1)return;c[a-p]=0}}return c}static isValidBroadcast(t,r){let o=t.length,i=r.length;if(o>i)return!1;for(let u=1;u<=o;u++)if(t[o-u]!==1&&t[o-u]!==r[i-u])return!1;return!0}},M=class e{static size(t){return e.getSizeFromDimensionRange(t,0,t.length)}static convertShape(t,r=4){let o=t.length;if(o===0)return[];let i=new Array(o),u=o-1;for(;u>=0;){if(t[u]%r===0){i[u]=t[u]/r;break}if(r%t[u]!==0)throw new Error("cannot convert shape");i[u]=1,r/=t[u],u--}for(u--;u>=0;u--)i[u]=t[u];return i}static sizeFromDimension(t,r){if(r<0||r>t.length)throw new Error(`invalid dimension of ${r} for sizeFromDimension as Tensor has ${t.length} dimensions.`);return e.getSizeFromDimensionRange(t,r,t.length)}static sizeToDimension(t,r){if(r<0||r>t.length)throw new Error(`invalid dimension of ${r} for sizeToDimension as Tensor has ${t.length} dimensions.`);return e.getSizeFromDimensionRange(t,0,r)}static getSizeFromDimensionRange(t,r,o){let i=1;for(let u=r;u<o;u++){if(t[u]<0)throw new Error("cannot get valid size from specified dimension range. Most likely the range contains negative values in them.");i*=t[u]}return i}static computeStrides(t){let r=t.length;if(r===0)return[];if(r===1)return[1];let o=new Array(r);o[r-1]=1,o[r-2]=t[r-1];for(let i=r-3;i>=0;--i)o[i]=o[i+1]*t[i+1];return o}static normalizeAxis(t,r){if(t<-r&&t>=r)throw new Error("unsupported axis for this operation.");return t<0?t+r:t}static normalizeAxes(t,r){return t.map(o=>this.normalizeAxis(o,r??t.length))}static sortBasedOnPerm(t,r){return r?r.map(o=>t[o]):t.slice().reverse()}static padShape(t,r){let o=t.length;return t.map((i,u)=>i+r[u]+r[u+o])}static areEqual(t,r){return t.length!==r.length?!1:t.every((o,i)=>o===r[i])}},nr=class e{static adjustPoolAttributes(t,r,o,i,u,a){if(!t&&o.length!==r.length-2)throw new Error("length of specified kernel shapes should be 2 less than length of input dimensions");if(t)for(let c=0;c<r.length-2;c++)c>=o.length?o.push(r[c+2]):o[c]=r[c+2];for(let c=0;c<o.length;c++)if(c<i.length){if(i[c]<0)throw new Error("strides should be greater than or equal to 1")}else i.push(1);for(let c=0;c<o.length;c++)if(c<u.length){if(u[c]<0)throw new Error("dilations should be greater than or equal to 1")}else u.push(1);for(let c=0;c<o.length*2;c++)if(c<a.length){if(a[c]<0)throw new Error("pad should be greater than or equal to 1")}else a.push(0);for(let c=0;c<o.length;c++){if(o[c]<=0)throw new Error("kernel shapes need to be greater than 0");if(a[c]>=o[c]||a[c+o.length]>=o[c])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(t,r,o,i,u,a,c){if(c){if(u.length!==2*(t.length-2))throw new Error("length of pads should be twice the length of data dimensions");if(r.length!==t.length-2)throw new Error("length of strides should be the length of data dimensions");if(i.length!==t.length-2)throw new Error("length of kernel shapes should be the length of data dimensions");for(let p=0;p<t.length-2;p++)e.adjustPadAndReturnShape(t[p+(a?1:2)],r[p],o[p],i[p],u,p,p+t.length-2,c)}}static computePoolOutputShape(t,r,o,i,u,a,c){if(r.length<=0)throw new Error("input shape must be of size greater than 0");let p=[r[0],r[1]];return e.computeShapeHelper(t,r,p,o,i,u,a,c),p}static computeConvOutputShape(t,r,o,i,u,a,c){if(t.length<=0||r.length<=0)throw new Error("invalid input tensor dims or invalid filter tensor dims");let p=[t[0],r[0]];return e.computeShapeHelper(!1,t,p,o,i,u,a,c),p}static computeShapeHelper(t,r,o,i,u,a,c,p){if(t)for(let h=0;h<r.length-2;h++)o.push(1);else for(let h=0;h<r.length-2;h++)o.push(e.adjustPadAndReturnShape(r[h+2],i[h],u[h],a[h],c,h,h+r.length-2,p))}static adjustPadAndReturnShape(t,r,o,i,u,a,c,p){let h=o*(i-1)+1;if(p&&p!=="NOTSET")switch(p){case"VALID":return u[a]=0,u[c]=0,Math.floor((t-h)/r+1);case"SAME_LOWER":case"SAME_UPPER":if(o!==1)throw new Error("Dilation not supported for SAME_UPPER or SAME_LOWER");{let y=((t+r-1)/r-1)*r+i-t;return u[a]=Math.floor(p==="SAME_LOWER"?(y+1)/2:y/2),u[c]=y-u[a],Math.floor((t+y-i)/r+1)}default:throw new Error("Unsupported AutoPad type")}else return Math.floor((t+u[a]+u[c]-h)/r+1)}},Sn=class{static getShapeOfGemmResult(t,r,o,i,u){if(t.length!==2||o.length!==2)throw new Error("shape need to be of size 2");let a,c,p;r?(a=t[1],c=t[0]):(a=t[0],c=t[1]);let h=-1;if(i?(p=o[0],h=1):(p=o[1],h=0),o[h]!==c)throw new Error("dimension mismatch");if(a<=0||p<=0||c<=0)throw new Error("invalid shape specified");if(u&&!It.isValidBroadcast(u,[a,p]))throw new Error("gemm: invalid bias shape for broadcast");return[a,p,c]}},xn=-34028234663852886e22,Cn=34028234663852886e22});var or,To,De,et,Z,Me,$t,ir,_t,fe,Eo,U,j,An,Io,va,_r,_e=Y(()=>{"use strict";ye();Se();or=64,To=(e,t)=>{if(t===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(e){case 10:return t>1?`vec${t}<f16>`:"f16";case 1:return t>1?`vec${t}<f32>`:"f32";case 6:return t>1?`vec${t}<i32>`:"i32";case 12:return t>1?`vec${t}<u32>`:"u32";case 7:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","i32"];case 13:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","u32"];case 9:if(t!==4)throw new Error("bool must be vec4");return["u32","vec4<bool>"];default:throw new Error(`Unknown data type: ${e}`)}},De=(e,t=1)=>{let r=To(e,t);return typeof r=="string"?r:r[0]},et=(e,t=1)=>{let r=To(e,t);return typeof r=="string"?r:r[1]},Z=(...e)=>{let t=[];return e.forEach(r=>{r.length!==0&&t.push({type:12,data:r},{type:12,data:M.computeStrides(r)})}),t},Me=e=>e%4===0?4:e%2===0?2:1,$t=(e="f32",t,r="0")=>!t||t===1?`${e}(${r})`:`vec${t}<${e}>(${r})`,ir=(e,t,r)=>e==="f32"?r:t===1?`f32(${r})`:`vec${t}<f32>(${r})`,_t=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,fe=(e,t,r,o)=>e.startsWith("uniforms.")&&r>4?typeof t=="string"?o==="f16"?`${e}[(${t}) / 8][(${t}) % 8 / 4][(${t}) % 8 % 4]`:`${e}[(${t}) / 4][(${t}) % 4]`:o==="f16"?`${e}[${Math.floor(t/8)}][${Math.floor(t%8/4)}][${t%8%4}]`:`${e}[${Math.floor(t/4)}][${t%4}]`:r>1?`${e}[${t}]`:e,Eo=(e,t,r,o,i)=>{let u=typeof r=="number",a=u?r:r.length,c=[...new Array(a).keys()],p=a<2?"u32":a<=4?`vec${a}<u32>`:`array<u32, ${a}>`,h=To(t,i),d=typeof h=="string"?h:h[1],y=typeof h=="string"?h:h[0],w={indices:p,value:d,storage:y,tensor:t},_=G=>typeof G=="string"?G:`${G}u`,v={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},S=u?"uniforms.":"",A=`${S}${e}_shape`,I=`${S}${e}_strides`,x="";for(let G=0;G<a-1;G++)x+=`\n    let dim${G} = current / ${fe(I,G,a)};\n    let rest${G} = current % ${fe(I,G,a)};\n    indices[${G}] = dim${G};\n    current = rest${G};\n    `;x+=`indices[${a-1}] = current;`;let E=a<2?"":`\n  fn o2i_${e}(offset: u32) -> ${w.indices} {\n    var indices: ${w.indices};\n    var current = offset;\n    ${x}\n    return indices;\n  }`,P=G=>(v.offsetToIndices=!0,a<2?G:`o2i_${e}(${G})`),O=[];if(a>=2)for(let G=a-1;G>=0;G--)O.push(`${fe(I,G,a)} * (indices[${G}])`);let R=a<2?"":`\n  fn i2o_${e}(indices: ${w.indices}) -> u32 {\n    return ${O.join("+")};\n  }`,L=G=>(v.indicesToOffset=!0,a<2?G:`i2o_${e}(${G})`),N=(...G)=>a===0?"0u":`${w.indices}(${G.map(_).join(",")})`,K=(G,ne)=>a<2?`${G}`:`${fe(G,ne,a)}`,Q=(G,ne,xe)=>a<2?`${G}=${xe};`:`${fe(G,ne,a)}=${xe};`,he={},W=(G,ne)=>{v.broadcastedIndicesToOffset=!0;let xe=`${ne.name}broadcastedIndicesTo${e}Offset`;if(xe in he)return`${xe}(${G})`;let Ke=[];for(let Be=a-1;Be>=0;Be--){let Ge=ne.indicesGet("outputIndices",Be+ne.rank-a);Ke.push(`${K(I,Be)} * (${Ge} % ${K(A,Be)})`)}return he[xe]=`fn ${xe}(outputIndices: ${ne.type.indices}) -> u32 {\n             return ${Ke.length>0?Ke.join("+"):"0u"};\n           }`,`${xe}(${G})`},se=(G,ne)=>(()=>{if(w.storage===w.value)return`${e}[${G}]=${ne};`;if(w.storage==="vec2<u32>"&&w.value==="i32")return`${e}[${G}]=vec2<u32>(u32(${ne}), select(0u, 0xFFFFFFFFu, ${ne} < 0));`;if(w.storage==="vec2<u32>"&&w.value==="u32")return`${e}[${G}]=vec2<u32>(u32(${ne}), 0u);`;if(w.storage==="u32"&&w.value==="vec4<bool>")return`${e}[${G}]=dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(${ne}));`;throw new Error(`not supported combination of storage type ${w.storage} and value type ${w.value} yet`)})(),Ce=G=>(()=>{if(w.storage===w.value)return`${e}[${G}]`;if(w.storage==="vec2<u32>"&&w.value==="i32")return`i32(${e}[${G}].x)`;if(w.storage==="vec2<u32>"&&w.value==="u32")return`u32(${e}[${G}].x)`;if(w.storage==="u32"&&w.value==="vec4<bool>")return`vec4<bool>(bool(${e}[${G}] & 0xFFu), bool(${e}[${G}] & 0xFF00u), bool(${e}[${G}] & 0xFF0000u), bool(${e}[${G}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${w.storage} and value type ${w.value} yet`)})(),We=a<2?"":`\n  fn get_${e}ByIndices(indices: ${w.indices}) -> ${d} {\n    return ${Ce(`i2o_${e}(indices)`)};\n  }`,ee=a<2?"":(()=>{let G=c.map(xe=>`d${xe}: u32`).join(", "),ne=c.map(xe=>`d${xe}`).join(", ");return`\n  fn get_${e}(${G}) -> ${d} {\n    return get_${e}ByIndices(${N(ne)});\n  }`})(),ae=(...G)=>{if(G.length!==a)throw new Error(`indices length must be ${a}`);let ne=G.map(_).join(",");return a===0?Ce("0u"):a===1?Ce(ne[0]):(v.get=!0,v.getByIndices=!0,v.indicesToOffset=!0,`get_${e}(${ne})`)},Ae=G=>a<2?Ce(G):(v.getByIndices=!0,v.indicesToOffset=!0,`get_${e}ByIndices(${G})`),me=a<2?"":`\n  fn set_${e}ByIndices(indices: ${w.indices}, value: ${d}) {\n    ${se(`i2o_${e}(indices)`,"value")}\n  }`,ie=a<2?"":(()=>{let G=c.map(xe=>`d${xe}: u32`).join(", "),ne=c.map(xe=>`d${xe}`).join(", ");return`\n  fn set_${e}(${G}, value: ${d}) {\n    set_${e}ByIndices(${N(ne)}, value);\n  }`})();return{impl:()=>{let G=[],ne=!1;return v.offsetToIndices&&(G.push(E),ne=!0),v.indicesToOffset&&(G.push(R),ne=!0),v.broadcastedIndicesToOffset&&(Object.values(he).forEach(xe=>G.push(xe)),ne=!0),v.set&&(G.push(ie),ne=!0),v.setByIndices&&(G.push(me),ne=!0),v.get&&(G.push(ee),ne=!0),v.getByIndices&&(G.push(We),ne=!0),!u&&ne&&G.unshift(`const ${A} = ${w.indices}(${r.join(",")});`,`const ${I} = ${w.indices}(${M.computeStrides(r).join(",")});`),G.join(`\n`)},type:w,offsetToIndices:P,indicesToOffset:L,broadcastedIndicesToOffset:W,indices:N,indicesGet:K,indicesSet:Q,set:(...G)=>{if(G.length!==a+1)throw new Error(`indices length must be ${a}`);let ne=G[a];if(typeof ne!="string")throw new Error("value must be string");let xe=G.slice(0,a).map(_).join(",");return a===0?se("0u",ne):a===1?se(xe[0],ne):(v.set=!0,v.setByIndices=!0,v.indicesToOffset=!0,`set_${e}(${xe}, ${ne})`)},setByOffset:se,setByIndices:(G,ne)=>a<2?se(G,ne):(v.setByIndices=!0,v.indicesToOffset=!0,`set_${e}ByIndices(${G}, ${ne});`),get:ae,getByOffset:Ce,getByIndices:Ae,usage:o,name:e,strides:I,shape:A,rank:a}},U=(e,t,r,o=1)=>Eo(e,t,r,"input",o),j=(e,t,r,o=1)=>Eo(e,t,r,"output",o),An=(e,t,r,o=1)=>Eo(e,t,r,"internal",o),Io=class{constructor(t,r){this.normalizedDispatchGroup=t;this.limits=r;this.internalVariables=[];this.variables=[];this.uniforms=[];this.variableIndex=0}guardAgainstOutOfBoundsWorkgroupSizes(t){return`if (global_idx >= ${typeof t=="number"?`${t}u`:t}) { return; }`}mainStart(t=or){let r=typeof t=="number"?t:t[0],o=typeof t=="number"?1:t[1],i=typeof t=="number"?1:t[2];if(r>this.limits.maxComputeWorkgroupSizeX||o>this.limits.maxComputeWorkgroupSizeY||i>this.limits.maxComputeWorkgroupSizeZ)throw new Error(`workgroup size [${r}, ${o}, ${i}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(r*o*i>this.limits.maxComputeInvocationsPerWorkgroup)throw new Error(`workgroup size [${r}, ${o}, ${i}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let u=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,a=u?`@builtin(global_invocation_id) global_id : vec3<u32>,\n    @builtin(workgroup_id) workgroup_id : vec3<u32>,\n    @builtin(local_invocation_id) local_id : vec3<u32>`:`@builtin(global_invocation_id) global_id : vec3<u32>,\n                                             @builtin(local_invocation_id) local_id : vec3<u32>,\n    @builtin(local_invocation_index) local_idx : u32,\n    @builtin(workgroup_id) workgroup_id : vec3<u32>,\n    @builtin(num_workgroups) num_workgroups : vec3<u32>`,c=u?"let global_idx = global_id.x; let local_idx = local_id.x;":`let global_idx = (workgroup_id.z * num_workgroups[0] * num_workgroups[1] +\n          workgroup_id.y * num_workgroups[0] + workgroup_id.x) * ${r*o*i}u + local_idx;`;return`@compute @workgroup_size(${r}, ${o}, ${i})\n  fn main(${a}) {\n    ${c}\n  `}appendVariableUniforms(t){t.rank!==0&&(t.shape.startsWith("uniforms.")&&this.uniforms.push({name:t.shape.replace("uniforms.",""),type:"u32",length:t.rank}),t.strides.startsWith("uniforms.")&&this.uniforms.push({name:t.strides.replace("uniforms.",""),type:"u32",length:t.rank}))}declareVariable(t,r){if(t.usage==="internal")throw new Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(t),this.appendVariableUniforms(t);let o=t.usage==="input"?"read":"read_write",i=t.type.storage;return`@group(0) @binding(${r}) var<storage, ${o}> ${t.name}: array<${i}>;`}declareVariables(...t){return t.map(r=>this.declareVariable(r,this.variableIndex++)).join(`\n`)}registerInternalVariable(t){if(t.usage!=="internal")throw new Error("cannot use input or output variable with registerInternalVariable(). use declareVariables() instead.");this.internalVariables.push(t),this.appendVariableUniforms(t)}registerInternalVariables(...t){return t.forEach(r=>this.registerInternalVariable(r)),this}registerUniform(t,r,o=1){return this.uniforms.push({name:t,type:r,length:o}),this}registerUniforms(t){return this.uniforms=this.uniforms.concat(t),this}uniformDeclaration(){if(this.uniforms.length===0)return"";let t=[];for(let{name:r,type:o,length:i}of this.uniforms)if(i&&i>4)o==="f16"?t.push(`@align(16) ${r}:array<mat2x4<${o}>, ${Math.ceil(i/8)}>`):t.push(`${r}:array<vec4<${o}>, ${Math.ceil(i/4)}>`);else{let u=i==null||i===1?o:`vec${i}<${o}>`;t.push(`${r}:${u}`)}return`\n      struct Uniforms { ${t.join(", ")} };\n      @group(0) @binding(${this.variableIndex}) var<uniform> uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(t=>t.impl()).join(`\n`)+this.internalVariables.map(t=>t.impl()).join(`\n`)}get variablesInfo(){if(this.uniforms.length===0)return;let t=r=>[12,10,1,6][["u32","f16","f32","i32"].indexOf(r)];return this.uniforms.map(r=>[t(r.type),r.length??1])}},va=(e,t)=>new Io(e,t),_r=(e,t)=>{let r=e.length,o=[];for(let i=0;i<r;i++){let u=r-1-i,a=e[u]||1;(t[t.length-1-i]||1)>1&&a===1&&o.unshift(u)}return o}});var Il,$a,Tl,El,yt,_a,Sa,Sr=Y(()=>{"use strict";ye();Se();Ze();_e();Il=e=>{if(!e||e.length!==1)throw new Error("Transpose requires 1 input.")},$a=(e,t)=>t&&t.length!==e?[...new Array(e).keys()].reverse():t,Tl=(e,t)=>M.sortBasedOnPerm(e,$a(e.length,t)),El=(e,t,r,o)=>{let i=[];i.push(`fn perm(i: ${o.type.indices}) -> ${r.type.indices} {\n    var a: ${r.type.indices};`);for(let u=0;u<t;++u)i.push(r.indicesSet("a",e[u],`i[${u}]`));return i.push("return a;}"),i.join(`\n`)},yt=(e,t)=>{let r=e.dataType,o=e.dims.length,i=$a(o,t),u=Tl(e.dims,i),a=j("output",r,u.length),c=U("a",r,o),p=h=>`\n  ${h.registerUniform("output_size","u32").declareVariables(c,a)}\n\n  ${El(i,o,c,a)}\n\n  ${h.mainStart()}\n    ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n    let indices = ${a.offsetToIndices("global_idx")};\n    let aIndices = perm(indices);\n\n    ${a.setByOffset("global_idx",c.getByIndices("aIndices"))}\n  }`;return{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:["rank"]},getRunData:h=>{let d=M.size(u);return{outputs:[{dims:u,dataType:h[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:[{type:12,data:d},...Z(h[0].dims,u)]}},getShaderSource:p}},_a=(e,t)=>{Il(e.inputs),e.compute(yt(e.inputs[0],t.perm))},Sa=e=>ve({perm:e.perm})});var Pl,kl,Ol,Rl,Bl,Dl,zl,Ml,Ul,Vl,Tt,xa,Ca,Aa,Ia,Ta,Ea,Pa,ka,Oa,Ra,Ba=Y(()=>{"use strict";ye();Se();_e();In();Sr();Pl={max:"select(bestValue, candidate, candidate > bestValue)",min:"select(bestValue, candidate, candidate < bestValue)",mean:"bestValue + candidate",sum:"bestValue + candidate",prod:"bestValue * candidate",sumSquare:"bestValue + candidate * candidate",logSumExp:"bestValue + exp(candidate)",l1:"bestValue + abs(candidate)",l2:"bestValue + candidate * candidate",logSum:"bestValue + candidate"},kl={max:"select(bestValue, candidate, candidate > bestValue)",min:"select(bestValue, candidate, candidate < bestValue)",mean:"bestValue + candidate",sum:"bestValue + candidate",prod:"bestValue * candidate",sumSquare:"bestValue + candidate",logSumExp:"bestValue + candidate",l1:"bestValue + candidate",l2:"bestValue + candidate",logSum:"bestValue + candidate"},Ol={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},Rl={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},Bl=(e,t)=>{let r=[];for(let o=t-e;o<t;++o)r.push(o);return r},Dl=(e,t)=>{let r=[],o=e.length;for(let u=0;u<o;u++)t.indexOf(u)===-1&&r.push(e[u]);let i=t.map(u=>e[u]);return[r,i]},zl=(e,t)=>{let r=e.length+t.length,o=[],i=0;for(let u=0;u<r;u++)t.indexOf(u)===-1?o.push(e[i++]):o.push(1);return o},Ml=(e,t)=>{for(let r=0;r<e.length;++r)if(e[e.length-r-1]!==t-1-r)return!1;return!0},Ul=(e,t)=>{let r=[];if(!Ml(e,t)){for(let o=0;o<t;++o)e.indexOf(o)===-1&&r.push(o);e.forEach(o=>r.push(o))}return r},Vl=(e,t,r,o,i,u,a)=>{let c=r[0].dims,p=M.size(u),h=M.size(a),d=U("_A",r[0].dataType,c),y=j("output",i,u),w=32,_=`\n          var<workgroup> aBestValues : array<f32, ${w}>;\n       `;return{name:e,shaderCache:t,getShaderSource:S=>`\n        ${S.registerUniform("reduceSize","u32").declareVariables(d,y)}\n        ${_}\n        fn DIV_CEIL(a : u32, b : u32) -> u32 {\n          return ((a - 1u) / b + 1u);\n         }\n         ${S.mainStart(w)}\n\n          let outputIndex = global_idx / ${w};\n          let offset = outputIndex * uniforms.reduceSize;\n\n          var bestValue = f32(${Ol[o]});\n          let Length = uniforms.reduceSize;\n          for (var k = local_idx; k < Length; k = k + ${w}) {\n           let candidate = f32(${d.getByOffset("offset + k")});\n           bestValue = ${Pl[o]};\n          }\n          aBestValues[local_idx] = bestValue;\n          workgroupBarrier();\n\n         var reduceSize = min(Length, ${w}u);\n         for (var currentSize = reduceSize / 2u; reduceSize > 1u;\n             currentSize = reduceSize / 2u) {\n           let interval = DIV_CEIL(reduceSize, 2u);\n           if (local_idx < currentSize) {\n            let candidate = aBestValues[local_idx + interval];\n            bestValue = ${kl[o]};\n            aBestValues[local_idx] = bestValue;\n           }\n           reduceSize = interval;\n           workgroupBarrier();\n         }\n\n         if (local_idx == 0u) {\n          ${y.setByOffset("outputIndex",`${o==="mean"?`${y.type.storage}(bestValue / f32(uniforms.reduceSize))`:`${y.type.storage}(${Rl[o]})`}`)};\n         }\n        }`,getRunData:()=>({outputs:[{dims:u,dataType:i}],dispatchGroup:{x:p},programUniforms:[{type:12,data:h}]})}},Tt=(e,t,r,o)=>{let i=e.inputs.length===1?r:Po(e.inputs,r),u=i.axes;u.length===0&&!i.noopWithEmptyAxes&&(u=e.inputs[0].dims.map((_,v)=>v));let a=M.normalizeAxes(u,e.inputs[0].dims.length),c=a,p=e.inputs[0],h=Ul(c,e.inputs[0].dims.length);h.length>0&&(p=e.compute(yt(e.inputs[0],h),{inputs:[0],outputs:[-1]})[0],c=Bl(c.length,p.dims.length));let[d,y]=Dl(p.dims,c),w=d;i.keepDims&&(w=zl(d,a)),e.compute(Vl(t,{hint:i.cacheKey,inputDependencies:["type"]},[p],o,e.inputs[0].dataType,w,y),{inputs:[p]})},xa=(e,t)=>{Tt(e,"ReduceMeanShared",t,"mean")},Ca=(e,t)=>{Tt(e,"ReduceL1Shared",t,"l1")},Aa=(e,t)=>{Tt(e,"ReduceL2Shared",t,"l2")},Ia=(e,t)=>{Tt(e,"ReduceLogSumExpShared",t,"logSumExp")},Ta=(e,t)=>{Tt(e,"ReduceMaxShared",t,"max")},Ea=(e,t)=>{Tt(e,"ReduceMinShared",t,"min")},Pa=(e,t)=>{Tt(e,"ReduceProdShared",t,"prod")},ka=(e,t)=>{Tt(e,"ReduceSumShared",t,"sum")},Oa=(e,t)=>{Tt(e,"ReduceSumSquareShared",t,"sumSquare")},Ra=(e,t)=>{Tt(e,"ReduceLogSumShared",t,"logSum")}});var Et,Wl,Tn,Po,Pt,Nl,Gl,Hl,Ll,Fl,ql,jl,Kl,Yl,Zl,kt,Da,za,Ma,Ua,Va,Wa,Na,Ga,Ha,La,In=Y(()=>{"use strict";ye();Se();Ze();_e();Ba();Et=e=>{if(!e||e.length===0||e.length>2)throw new Error("Reduce op requires 1 or 2 inputs.");if(e.length===2&&e[1].dims.length!==1)throw new Error("Invalid axes input dims.")},Wl=e=>["","",`var value = ${e.getByIndices("input_indices")};`,""],Tn=(e,t,r,o,i,u,a=!1,c=!1)=>{let p=[],h=r[0].dims,d=h.length,y=M.normalizeAxes(i,d),w=!c&&y.length===0;h.forEach((A,I)=>{w||y.indexOf(I)>=0?a&&p.push(1):p.push(A)});let _=p.length,v=M.size(p);return{name:e,shaderCache:t,getShaderSource:A=>{let I=[],x=U("_A",r[0].dataType,d),E=j("output",u,_),P=o(x,E,y),O=P[2];for(let R=0,L=0;R<d;R++)w||y.indexOf(R)>=0?(a&&L++,O=`for(var j${R}: u32 = 0; j${R} < ${h[R]}; j${R}++) {\n                  ${P[2].includes("last_index")?`let last_index = j${R};`:""}\n                  ${x.indicesSet("input_indices",R,`j${R}`)}\n                  ${O}\n                }`):(I.push(`${x.indicesSet("input_indices",R,E.indicesGet("output_indices",L))};`),L++);return`\n\n        ${A.registerUniform("output_size","u32").declareVariables(x,E)}\n\n        ${A.mainStart()}\n          ${A.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n          var input_indices: ${x.type.indices};\n          let output_indices = ${E.offsetToIndices("global_idx")};\n\n          ${I.join(`\n`)}\n          ${P[0]}       // init ops for reduce max/min\n          ${P[1]}\n          ${O}\n          ${P[3]}\n          ${P.length===4?E.setByOffset("global_idx","value"):P.slice(4).join(`\n`)}\n        }`},getRunData:()=>({outputs:[{dims:p,dataType:u}],dispatchGroup:{x:Math.ceil(v/64)},programUniforms:[{type:12,data:v},...Z(h,p)]})}},Po=(e,t)=>{let r=[];return e[1].dims[0]>0&&e[1].getBigInt64Array().forEach(o=>r.push(Number(o))),ve({axes:r,keepDims:t.keepDims,noopWithEmptyAxes:t.noopWithEmptyAxes})},Pt=(e,t,r,o)=>{let i=e.inputs,u=i.length===1?r:Po(i,r);e.compute(Tn(t,{hint:u.cacheKey,inputDependencies:["rank"]},[i[0]],u.noopWithEmptyAxes&&u.axes.length===0?Wl:o,u.axes,i[0].dataType,u.keepDims,u.noopWithEmptyAxes),{inputs:[0]})},Nl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceLogSum",t,(o,i)=>[`var value = ${i.type.storage}(0);`,"",`value += ${o.getByIndices("input_indices")};`,"value = log(value);"])},Gl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceL1",t,(o,i)=>[`var value = ${i.type.storage}(0);`,"",`value += abs(${o.getByIndices("input_indices")});`,""])},Hl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceL2",t,(o,i)=>[`var t = ${i.type.value}(0); var value = ${i.type.value}(0);`,"",`t = ${o.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},Ll=(e,t)=>{Et(e.inputs),Pt(e,"ReduceLogSumExp",t,(o,i)=>[`var value = ${i.type.storage}(0);`,"",`value += exp(${o.getByIndices("input_indices")});`,"value = log(value);"])},Fl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceMax",t,(o,i,u)=>{let a=[];for(let c=0;c<o.rank;c++)(u.indexOf(c)>=0||u.length===0)&&a.push(o.indicesSet("input_indices",c,0));return[`${a.join(`\n`)}`,`var value = ${o.getByIndices("input_indices")};`,`value = max(value, ${o.getByIndices("input_indices")});`,""]})},ql=(e,t)=>{Et(e.inputs),Pt(e,"ReduceMean",t,(o,i,u)=>{let a=1;for(let c=0;c<o.rank;c++)(u.indexOf(c)>=0||u.length===0)&&(a*=e.inputs[0].dims[c]);return["var sum = f32(0);","",`sum += f32(${o.getByIndices("input_indices")});`,`let value = ${i.type.value}(sum / ${a});`]})},jl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceMin",t,(o,i,u)=>{let a=[];for(let c=0;c<o.rank;c++)(u.indexOf(c)>=0||u.length===0)&&a.push(`input_indices[${c}] = 0;`);return[`${a.join(`\n`)}`,`var value = ${o.getByIndices("input_indices")};`,`value = min(value, ${o.getByIndices("input_indices")});`,""]})},Kl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceProd",t,(o,i)=>[`var value = ${i.type.storage}(1);`,"",`value *= ${o.getByIndices("input_indices")};`,""])},Yl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceSum",t,(o,i)=>[`var value = ${i.type.storage}(0);`,"",`value += ${o.getByIndices("input_indices")};`,""])},Zl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceSumSquare",t,(o,i)=>[`var t = ${i.type.value}(0); var value = ${i.type.value}(0);`,"",`t = ${o.getByIndices("input_indices")}; value += t * t;`,""])},kt=(e,t,r)=>{if(t.length===0)return r;let o=1,i=1;for(let u=0;u<t.length;u++)t.indexOf(u)===-1?o*=e[u]:i*=e[u];return i<32&&o>1024},Da=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?ql(e,t):xa(e,t)},za=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Gl(e,t):Ca(e,t)},Ma=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Hl(e,t):Aa(e,t)},Ua=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Ll(e,t):Ia(e,t)},Va=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Fl(e,t):Ta(e,t)},Wa=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?jl(e,t):Ea(e,t)},Na=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Kl(e,t):Pa(e,t)},Ga=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Yl(e,t):ka(e,t)},Ha=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Zl(e,t):Oa(e,t)},La=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Nl(e,t):Ra(e,t)}});var Fa,qa,ja,ko,Ka=Y(()=>{"use strict";ye();Ze();In();Fa=e=>{if(!e||e.length===0||e.length>2)throw new Error("ArgMinMaxOp op requires 1 or 2 inputs.");if(e[0].dataType!==1)throw new Error("Invalid input type.")},qa=(e,t)=>{Fa(e.inputs);let r=(o,i,u)=>{let a=[];for(let c=0;c<o.rank;c++)(u.indexOf(c)>=0||u.length===0)&&a.push(`input_indices[${c}] = 0;`);return[`${a.join(`\n`)}`,`var value = ${o.getByIndices("input_indices")};\nvar best_index : i32 = 0;`,`if (${o.getByIndices("input_indices")} ${t.selectLastIndex>0?"<=":"<"} value) {\n         value = ${o.getByIndices("input_indices")};\n         best_index = i32(last_index);\n       }`,"",i.setByOffset("global_idx","best_index")]};e.compute(Tn("ArgMin",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},ja=(e,t)=>{Fa(e.inputs);let r=(o,i,u)=>{let a=[];for(let c=0;c<o.rank;c++)(u.indexOf(c)>=0||u.length===0)&&a.push(`input_indices[${c}] = 0;`);return[`${a.join(`\n`)}`,`var value = ${o.getByIndices("input_indices")};\nvar best_index : i32 = 0;`,`if (${o.getByIndices("input_indices")} ${t.selectLastIndex>0?">=":">"} value) {\n         value = ${o.getByIndices("input_indices")};\n         best_index = i32(last_index);\n       }`,"",i.setByOffset("global_idx","best_index")]};e.compute(Tn("argMax",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},ko=e=>ve(e)});var Xl,Ql,Jl,En,Ya,Za,Oo=Y(()=>{"use strict";ye();Se();Ze();_e();Xl=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");let r=0,o=e[r],i=o.dataType,u=o.dims.length;e.forEach((a,c)=>{if(c!==r){if(a.dataType!==i)throw new Error("input tensors should be one type");if(a.dims.length!==u)throw new Error("input tensors should have the same shape");a.dims.forEach((p,h)=>{if(h!==t&&p!==o.dims[h])throw new Error("non concat dimensions must match")})}})},Ql=(e,t)=>`\n  fn calculateInputIndex(index: u32) -> u32 {\n    let sizeInConcatAxis = array<u32, ${e}u>(${t});\n    for (var i: u32 = 0u; i < ${e}; i += 1u ) {\n      if (index < sizeInConcatAxis[i]) {\n        return i;\n      }\n    }\n    return ${e}u;\n  }`,Jl=(e,t)=>{let r=e.length,o=[];for(let i=0;i<r;++i){let u=t.setByOffset("global_idx",e[i].getByIndices("indices"));r===1?o.push(u):i===0?o.push(`if (inputIndex == ${i}u) { ${u} }`):i===r-1?o.push(`else { ${u} }`):o.push(`else if (inputIndex == ${i}) { ${u} }`)}return o.join(`\n`)},En=(e,t,r,o)=>{let i=M.size(r),u=new Array(e.length),a=new Array(e.length),c=0,p=[],h=[],d=[{type:12,data:i}];for(let S=0;S<e.length;++S)c+=e[S].dims[t],u[S]=c,h.push(e[S].dims.length),a[S]=U(`input${S}`,o,h[S]),p.push("rank"),d.push({type:12,data:u[S]});for(let S=0;S<e.length;++S)d.push(...Z(e[S].dims));d.push(...Z(r));let y=j("output",o,r.length),w=y.indicesGet("indices",t),_=Array.from(Array(u.length).keys()).map(S=>`uniforms.sizeInConcatAxis${S}`).join(","),v=S=>`\n\n  ${(()=>{S.registerUniform("outputSize","u32");for(let A=0;A<e.length;A++)S.registerUniform(`sizeInConcatAxis${A}`,"u32");return S.declareVariables(...a,y)})()}\n\n  ${Ql(u.length,_)}\n\n  ${S.mainStart()}\n    ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n\n    var indices = ${y.offsetToIndices("global_idx")};\n\n    let inputIndex = calculateInputIndex(${w});\n    if (inputIndex != 0u) {\n      let sizeInConcatAxis = array<u32, ${u.length}u>(${_});\n      ${w} -= sizeInConcatAxis[inputIndex - 1u];\n    }\n\n    ${Jl(a,y)}\n  }`;return{name:"Concat",shaderCache:{hint:`${t}`,inputDependencies:p},getRunData:()=>({outputs:[{dims:r,dataType:o}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:d}),getShaderSource:v}},Ya=(e,t)=>{let r=e.inputs,o=r[0].dims,i=M.normalizeAxis(t.axis,o.length);Xl(r,i);let u=o.slice();u[i]=r.reduce((c,p)=>c+(p.dims.length>i?p.dims[i]:0),0);let a=r.filter(c=>M.size(c.dims)>0);e.compute(En(a,i,u,r[0].dataType),{inputs:a})},Za=e=>ve({axis:e.axis})});var ec,tc,rc,nc,Pn,oc,Xa,Ro=Y(()=>{"use strict";ye();_n();_e();Oo();ec=(e,t)=>{let r=e[0],o=e[1],i=e[2],u=e[3],a=e[4],c=e[5];if(a&&c)throw new Error("Attention cannot have both past and relative_position_bias");if(r.dims.length!==3)throw new Error(\'Input "input" must have 3 dimensions\');let p=r.dims[0],h=r.dims[1],d=r.dims[2];if(i.dims.length!==1)throw new Error(\'Input "bias" is expected to have 1 dimensions\');if(o.dims.length!==2)throw new Error(\'Input "weights" is expected to have 2 dimensions\');if(o.dims[0]!==d)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(i.dims[0]!==o.dims[1])throw new Error(\'Input "bias" dimension 0 should have same length as dimension 1 of input "weights"\');let y=i.dims[0]/3,w=y,_=w;if(t.qkvHiddenSizes.length>0){if(t.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let E of t.qkvHiddenSizes)if(E%t.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");y=t.qkvHiddenSizes[0],w=t.qkvHiddenSizes[1],_=t.qkvHiddenSizes[2]}let v=h;if(y!==w)throw new Error("qkv_hidden_sizes first element should be same as the second");if(i.dims[0]!==y+w+_)throw new Error(\'Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes\');let S=0;if(a){if(w!==_)throw new Error(\'Input "past" expect k_hidden_size == v_hidden_size\');if(a.dims.length!==5)throw new Error(\'Input "past" must have 5 dimensions\');if(a.dims[0]!==2)throw new Error(\'Input "past" first dimension must be 2\');if(a.dims[1]!==p)throw new Error(\'Input "past" second dimension must be batch_size\');if(a.dims[2]!==t.numHeads)throw new Error(\'Input "past" third dimension must be num_heads\');if(a.dims[4]!==w/t.numHeads)throw new Error(\'Input "past" fifth dimension must be k_hidden_size / num_heads\');t.pastPresentShareBuffer||(S=a.dims[3])}let A=v+S,I=-1,x=0;if(u)throw new Error("Mask not supported");if(a)throw new Error("past is not supported");return{batchSize:p,sequenceLength:h,pastSequenceLength:S,kvSequenceLength:v,totalSequenceLength:A,maxSequenceLength:I,inputHiddenSize:d,hiddenSize:y,vHiddenSize:_,headSize:Math.floor(y/t.numHeads),vHeadSize:Math.floor(_/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:x,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},tc=(e,t,r,o)=>{let i=Me(o),u=64,a=o/i;a<u?u=1:a/8<64&&(u=Math.ceil(a/8));let c=Math.ceil(o/i/u),p=[{type:t.dataType,data:1/o},{type:12,data:a},{type:12,data:c}],h=De(t.dataType,i),d=et(1,i),y=w=>{let _=j("x",t.dataType,t.dims,i),S=[{name:"d_inv",type:et(t.dataType)},{name:"d_comp",type:"u32"},{name:"elements_per_thread",type:"u32"}];return`\n  var<workgroup> thread_max: array<f32, ${u}>;\n  var<workgroup> thread_sum: array<f32, ${u}>;\n  ${w.registerUniforms(S).declareVariables(_)}\n  ${w.mainStart([u,1,1])}\n    let local_offset = local_idx * uniforms.elements_per_thread;\n    let offset = workgroup_id.x * uniforms.d_comp + local_offset;\n\n    var thread_max_vector = ${d}(-3.402823e+38f);\n    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {\n      thread_max_vector = max(${d}(x[offset + i]), thread_max_vector);\n    }\n    thread_max[local_idx] = ${(()=>{switch(i){case 1:return"thread_max_vector";case 2:return"max(thread_max_vector.x, thread_max_vector.y)";case 4:return"max(max(thread_max_vector.x, thread_max_vector.y), max(thread_max_vector.z, thread_max_vector.w))";default:throw new Error(`Unsupported components: ${i}`)}})()};\n    workgroupBarrier();\n\n    var max_value =  f32(-3.402823e+38f);\n    for (var i = 0u; i < ${u}; i++) {\n      max_value = max(thread_max[i], max_value);\n    }\n\n    var sum_vector = ${d}(0);\n    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {\n      sum_vector += exp(${d}(x[offset + i]) - max_value);\n    }\n    thread_sum[local_idx] = ${(()=>{switch(i){case 1:return"sum_vector";case 2:return"sum_vector.x + sum_vector.y";case 4:return"sum_vector.x + sum_vector.y + sum_vector.z + sum_vector.w";default:throw new Error(`Unsupported components: ${i}`)}})()};\n    workgroupBarrier();\n\n    var sum: f32 = 0;\n    for (var i = 0u; i < ${u}; i++) {\n      sum += thread_sum[i];\n    }\n\n    if (sum == 0) {\n      for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {\n        x[offset + i] = ${_.type.value}(uniforms.d_inv);\n      }\n    } else {\n      for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {\n        var f32input = ${d}(x[offset + i]);\n        x[offset + i] = ${_.type.value}(exp(f32input - max_value) / sum);\n      }\n    }\n  }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${u};${h};${i}`},getShaderSource:y,getRunData:()=>({outputs:[],dispatchGroup:{x:r},programUniforms:p})}},rc=(e,t,r,o,i,u,a)=>{let c=a+i.kvSequenceLength,p=[i.batchSize,i.numHeads,i.sequenceLength,c],h=u.scale===0?1/Math.sqrt(i.headSize):u.scale,d=Me(i.headSize),y=i.headSize/d,w=12,_={x:Math.ceil(c/w),y:Math.ceil(i.sequenceLength/w),z:i.batchSize*i.numHeads},v=[{type:12,data:i.sequenceLength},{type:12,data:y},{type:12,data:c},{type:12,data:i.numHeads},{type:1,data:h}],S=o?["type","type","type"]:["type","type"],A=I=>{let x=U("q",t.dataType,t.dims,d),E=U("key",r.dataType,r.dims,d),P=[x,E];o&&P.push(U("relative_position_bias",o.dataType,o.dims));let O=j("output",t.dataType,p),R=et(1,d),L=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"alpha",type:"f32"}];return`\n  const TILE_SIZE = ${w}u;\n\n  var<workgroup> tileQ: array<${x.type.storage}, ${w*w}>;\n  var<workgroup> tileK: array<${x.type.storage}, ${w*w}>;\n  ${I.registerUniforms(L).declareVariables(...P,O)}\n  ${I.mainStart([w,w,1])}\n    // x holds the N and y holds the M\n    let headIdx = workgroup_id.z;\n    let m = workgroup_id.y * TILE_SIZE;\n    let n = workgroup_id.x * TILE_SIZE;\n    let qOffset = uniforms.M * uniforms.K * headIdx + m * uniforms.K;\n    let kOffset = uniforms.N * uniforms.K * headIdx + n * uniforms.K;\n\n    var value = ${R}(0);\n    for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {\n      if (global_id.y < uniforms.M && w + local_id.x < uniforms.K) {\n        tileQ[TILE_SIZE * local_id.y + local_id.x] = q[qOffset + local_id.y * uniforms.K + w + local_id.x];\n      }\n      if (n + local_id.y < uniforms.N && w + local_id.x < uniforms.K) {\n        tileK[TILE_SIZE * local_id.y + local_id.x] = key[kOffset + local_id.y * uniforms.K + w + local_id.x];\n      }\n      workgroupBarrier();\n\n      for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) {\n        value += ${R}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]);\n      }\n\n      workgroupBarrier();\n    }\n\n    let headOffset = headIdx * uniforms.M * uniforms.N;\n    if (global_id.y < uniforms.M && global_id.x < uniforms.N) {\n      let outputIdx = headOffset + global_id.y * uniforms.N + global_id.x;\n      var sum: f32 = ${(()=>{switch(d){case 1:return"value";case 2:return"value.x + value.y";case 4:return"value.x + value.y + value.z + value.w";default:throw new Error(`Unsupported components: ${d}`)}})()};\n        output[outputIdx] = ${O.type.value} (sum * uniforms.alpha) + ${o?"relative_position_bias[outputIdx]":"0.0"};\n    }\n  }`};return{name:"AttentionProbs",shaderCache:{hint:`${d}`,inputDependencies:S},getRunData:()=>({outputs:[{dims:p,dataType:t.dataType,gpuDataType:0}],dispatchGroup:_,programUniforms:v}),getShaderSource:A}},nc=(e,t,r,o,i)=>{let u=i+o.kvSequenceLength,a=[o.batchSize,o.sequenceLength,o.vHiddenSize],c=12,p={x:Math.ceil(o.vHeadSize/c),y:Math.ceil(o.sequenceLength/c),z:o.batchSize*o.numHeads},h=[{type:12,data:o.sequenceLength},{type:12,data:u},{type:12,data:o.vHeadSize},{type:12,data:o.numHeads},{type:12,data:o.vHiddenSize}];return{name:"AttentionScore",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:a,dataType:t.dataType,gpuDataType:0}],dispatchGroup:p,programUniforms:h}),getShaderSource:w=>{let _=U("probs",t.dataType,t.dims),v=U("v",r.dataType,r.dims),S=j("output",t.dataType,a),A=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"v_hidden_size",type:"u32"}];return`\n  const TILE_SIZE = ${c}u;\n  var<workgroup> tileQ: array<${_.type.value}, ${c*c}>;\n  var<workgroup> tileK: array<${_.type.value}, ${c*c}>;\n  ${w.registerUniforms(A).declareVariables(_,v,S)}\n  ${w.mainStart([c,c,1])}\n   let headIdx = workgroup_id.z;\n   let m = global_id.y;\n   let n = global_id.x;\n\n   let offsetA = headIdx * (uniforms.M * uniforms.K) + m * uniforms.K;\n   let offsetB = headIdx * (uniforms.N * uniforms.K) + n;\n\n   var value = ${_.type.storage}(0);\n   for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {\n     if (m < uniforms.M && w + local_id.x < uniforms.K) {\n       tileQ[TILE_SIZE * local_id.y + local_id.x] = probs[offsetA + w + local_id.x];\n     }\n     if (n < uniforms.N && w + local_id.y < uniforms.K) {\n       tileK[TILE_SIZE * local_id.y + local_id.x] = v[offsetB + (w + local_id.y) * uniforms.N];\n     }\n     workgroupBarrier();\n     for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) {\n       value += tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * k + local_id.x];\n     }\n     workgroupBarrier();\n   }\n\n   // we need to transpose output from BNSH_v to BSND_v\n   let batchIdx = workgroup_id.z / uniforms.num_heads;\n   let currentBatchHeadNumber = workgroup_id.z % uniforms.num_heads;\n   if (m < uniforms.M && n < uniforms.N) {\n     let outputIdx = batchIdx * uniforms.M * uniforms.v_hidden_size + m * uniforms.v_hidden_size\n       + currentBatchHeadNumber * uniforms.N + n;\n     output[outputIdx] = value;\n   }\n  }`}}},Pn=(e,t,r,o,i,u,a,c,p,h,d)=>{let y=e.outputCount>1,w=e.outputCount>2,_=y&&w?h.pastSequenceLength:0,v=_+h.kvSequenceLength,S=[h.batchSize,h.numHeads,v,h.headSize],A=a?[a,r]:[r],I=y?e.compute(En(A,2,S,r.dataType),{inputs:A,outputs:[1]})[0]:r,x=[h.batchSize,h.numHeads,v,h.headSize],E=c?[c,o]:[o],P=w?e.compute(En(E,2,x,o.dataType),{inputs:E,outputs:[2]})[0]:o,O=[t,I];p&&O.push(p);let R=e.compute(rc(e,t,I,p,h,d,_),{inputs:O,outputs:[-1]})[0];e.compute(tc(e,R,h.batchSize*h.numHeads*h.sequenceLength,v),{inputs:[R],outputs:[]});let L=[R,P];e.compute(nc(e,R,P,h,_),{inputs:L,outputs:[0]})},oc=(e,t)=>{let r=[t.batchSize,t.numHeads,t.sequenceLength,t.headSize],o=t.sequenceLength,i=t.inputHiddenSize,u=t.headSize,a=12,c={x:Math.ceil(t.headSize/a),y:Math.ceil(t.sequenceLength/a),z:t.batchSize*t.numHeads},p=[e.inputs[0],e.inputs[1],e.inputs[2]],h=[{type:12,data:o},{type:12,data:i},{type:12,data:u},{type:12,data:t.numHeads},{type:12,data:t.headSize},{type:12,data:t.hiddenSize},{type:12,data:t.hiddenSize+t.hiddenSize+t.vHiddenSize}],d=y=>{let w=j("output_q",p[0].dataType,r),_=j("output_k",p[0].dataType,r),v=j("output_v",p[0].dataType,r),S=U("input",p[0].dataType,p[0].dims),A=U("weight",p[1].dataType,p[1].dims),I=U("bias",p[2].dataType,p[2].dims),x=S.type.storage,E=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"hidden_size",type:"u32"},{name:"ldb",type:"u32"}];return`\n  const TILE_SIZE = ${a}u;\n  var<workgroup> tileInput: array<${x}, ${a*a}>;\n  var<workgroup> tileWeightQ: array<${x}, ${a*a}>;\n  var<workgroup> tileWeightK: array<${x}, ${a*a}>;\n  var<workgroup> tileWeightV: array<${x}, ${a*a}>;\n  ${y.registerUniforms(E).declareVariables(S,A,I,w,_,v)}\n  ${y.mainStart([a,a,1])}\n    let batchIndex = workgroup_id.z / uniforms.num_heads;\n    let headNumber = workgroup_id.z % uniforms.num_heads;\n    let m = global_id.y;\n    let n = global_id.x;\n\n    let inputOffset = batchIndex * (uniforms.M * uniforms.K) + m * uniforms.K;\n    let biasOffsetQ = headNumber * uniforms.head_size;\n    let biasOffsetK = uniforms.hidden_size + biasOffsetQ;\n    let biasOffsetV = uniforms.hidden_size + biasOffsetK;\n\n    var valueQ = ${x}(0);\n    var valueK = ${x}(0);\n    var valueV = ${x}(0);\n    for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {\n      if (m < uniforms.M && w + local_id.x < uniforms.K) {\n        tileInput[TILE_SIZE * local_id.y + local_id.x] = input[inputOffset + w + local_id.x];\n      }\n      if (n < uniforms.N && w + local_id.y < uniforms.K) {\n        let offset = n + (w + local_id.y) * uniforms.ldb;\n        tileWeightQ[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetQ + offset];\n        tileWeightK[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetK + offset];\n        tileWeightV[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetV + offset];\n      }\n      workgroupBarrier();\n      for (var k: u32 = 0u; k<TILE_SIZE && w+k < uniforms.K; k++) {\n        let inputTileOffset = TILE_SIZE * local_id.y + k;\n        let weightTileOffset = TILE_SIZE * k + local_id.x;\n        valueQ += tileInput[inputTileOffset] * tileWeightQ[weightTileOffset];\n        valueK += tileInput[inputTileOffset] * tileWeightK[weightTileOffset];\n        valueV += tileInput[inputTileOffset] * tileWeightV[weightTileOffset];\n      }\n\n      workgroupBarrier();\n    }\n\n    let headOffset = (m * uniforms.N + n) % uniforms.head_size;\n    valueQ += bias[headOffset + biasOffsetQ];\n    valueK += bias[headOffset + biasOffsetK];\n    valueV += bias[headOffset + biasOffsetV];\n\n    let offset = workgroup_id.z * uniforms.M * uniforms.N;\n    if (m < uniforms.M && n < uniforms.N) {\n      let outputIdx = offset + m * uniforms.N + n;\n      output_q[outputIdx] = valueQ;\n      output_k[outputIdx] = valueK;\n      output_v[outputIdx] = valueV;\n    }\n  }`};return e.compute({name:"AttentionPrepare",shaderCache:{inputDependencies:["type","type","type"]},getRunData:()=>({outputs:[{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0}],dispatchGroup:c,programUniforms:h}),getShaderSource:d},{inputs:p,outputs:[-1,-1,-1]})},Xa=(e,t)=>{let r=ec(e.inputs,t),[o,i,u]=oc(e,r);return Pn(e,o,i,u,e.inputs[4],void 0,void 0,void 0,e.inputs[5],r,t)}});var ic,ac,sc,Qa,Ja=Y(()=>{"use strict";$r();ye();Se();Ze();_e();ic=(e,t)=>{if(!e||e.length!==5)throw new Error("BatchNormalization requires 5 inputs");let r=(o,i,u)=>{let a=i.length;if(a!==o.length)throw new Error(`${u}: num dimensions != ${a}`);i.forEach((c,p)=>{if(c!==o[p])throw new Error(`${u}: dim[${p}] do not match`)})};if(e[0].dims.length>1){let o=t.format==="NHWC"?t.spatial?e[0].dims.slice(-1):e[0].dims.slice(-1).concat(e[0].dims.slice(1,e[0].dims.length-1)):e[0].dims.slice(1,t.spatial?2:void 0);r(e[1].dims,o,"Invalid input scale"),r(e[2].dims,o,"Invalid input B"),r(e[3].dims,o,"Invalid input mean"),r(e[4].dims,o,"Invalid input var")}else r(e[1].dims,[1],"Invalid input scale"),r(e[2].dims,[1],"Invalid input B"),r(e[3].dims,[1],"Invalid input mean"),r(e[4].dims,[1],"Invalid input var")},ac=(e,t)=>{let{epsilon:r,spatial:o,format:i}=t,u=e[0].dims,a=o?Me(u[u.length-1]):1,c=i==="NHWC"&&u.length>1?a:1,p=M.size(u)/a,h=o,d=h?u.length:u,y=U("x",e[0].dataType,e[0].dims,a),w=U("scale",e[1].dataType,e[1].dims,c),_=U("bias",e[2].dataType,e[2].dims,c),v=U("inputMean",e[3].dataType,e[3].dims,c),S=U("inputVar",e[4].dataType,e[4].dims,c),A=j("y",e[0].dataType,d,a),I=()=>{let E="";if(o)E=`let cOffset = ${u.length===1?"0u":i==="NHWC"?`outputIndices[${u.length-1}] / ${a}`:"outputIndices[1]"};`;else if(i==="NCHW")E=`\n            ${A.indicesSet("outputIndices","0","0")}\n            let cOffset = ${A.indicesToOffset("outputIndices")};`;else{E=`var cIndices = ${w.type.indices}(0);\n                       cIndices[0] = outputIndices[${u.length-1}];`;for(let P=1;P<w.rank;P++)E+=`cIndices[${P}] = outputIndices[${P}];`;E+=`let cOffset = ${w.indicesToOffset("cIndices")};`}return E},x=E=>`\n  const epsilon = ${r};\n  ${E.registerUniform("outputSize","u32").declareVariables(y,w,_,v,S,A)}\n  ${E.mainStart()}\n  ${E.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n    var outputIndices = ${A.offsetToIndices(`global_idx * ${a}`)};\n    ${I()}\n    let scale = ${w.getByOffset("cOffset")};\n    let bias = ${_.getByOffset("cOffset")};\n    let inputMean = ${v.getByOffset("cOffset")};\n    let inputVar = ${S.getByOffset("cOffset")};\n    let x = ${y.getByOffset("global_idx")};\n    let value = (x - inputMean) * inverseSqrt(inputVar + epsilon) * scale + bias;\n    ${A.setByOffset("global_idx","value")}\n  }`;return{name:"BatchNormalization",shaderCache:{hint:`${t.epsilon}_${t.format}_${o}_${a}`,inputDependencies:h?["rank","type","type","type","type"]:void 0},getShaderSource:x,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:h?[{type:12,data:p},...Z(u)]:[{type:12,data:p}]})}},sc=e=>ve(e),Qa=(e,t)=>{let{inputs:r,outputCount:o}=e,i=sc({...t,outputCount:o});if(vr.webgpu.validateInputContent&&ic(r,i),t.trainingMode)throw new Error("BatchNormalization trainingMode is not supported yet.");e.compute(ac(r,i))}});var uc,dc,es,ts=Y(()=>{"use strict";Se();_e();uc=e=>{if(e[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![320,640,1280].includes(e[0].dims[2]))throw new Error("number of channels should be 320, 640 or 1280");if(e[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(e[0].dims[2]!==e[1].dims[0])throw new Error("last dimension of input and bias are not the same")},dc=e=>{let t=e[0].dims,r=e[0].dims[2],o=M.size(t)/4,i=e[0].dataType,u=U("input",i,t,4),a=U("bias",i,[r],4),c=U("residual",i,t,4),p=j("output",i,t,4);return{name:"BiasAdd",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(o/64)}}),getShaderSource:d=>`\n  const channels = ${r}u / 4;\n  ${d.declareVariables(u,a,c,p)}\n\n  ${d.mainStart()}\n    ${d.guardAgainstOutOfBoundsWorkgroupSizes(o)}\n    let value = ${u.getByOffset("global_idx")}\n      + ${a.getByOffset("global_idx % channels")} + ${c.getByOffset("global_idx")};\n    ${p.setByOffset("global_idx","value")}\n  }`}},es=e=>{uc(e.inputs),e.compute(dc(e.inputs))}});var lc,ke,rs,ns,os,is,as,ss,us,ds,ls,cc,cs,ps,ms,fs,kn,hs,On,gs,ys,bs,ws,vs,$s,_s,Ss,xs,Cs,As,Is,Ts,Es,Ps,ks,Os,Rs,Bo,Do,Bs,Ds,zs,Rn=Y(()=>{"use strict";ye();Se();Ze();_e();lc=(e,t,r,o,i,u)=>{let a=Math.ceil(t/4),c="";typeof i=="string"?c=`${i}(a)`:c=i("a");let p=U("inputData",r,[a],4),h=j("outputData",o,[a],4);return`\n      ${e.registerUniform("vec_size","u32").declareVariables(p,h)}\n\n  ${u??""}\n\n  ${e.mainStart()}\n    ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}\n\n    let a = ${p.getByOffset("global_idx")};\n    ${h.setByOffset("global_idx",c)}\n  }`},ke=(e,t,r,o,i,u=e.dataType)=>({name:t,shaderCache:{hint:i,inputDependencies:["type"]},getShaderSource:a=>lc(a,M.size(e.dims),e.dataType,u,r,o),getRunData:a=>({outputs:[{dims:e.dims,dataType:u}],dispatchGroup:{x:Math.ceil(M.size(a[0].dims)/64/4)},programUniforms:[{type:12,data:Math.ceil(M.size(e.dims)/4)}]})}),rs=e=>{e.compute(ke(e.inputs[0],"Abs","abs"))},ns=e=>{e.compute(ke(e.inputs[0],"Acos","acos"))},os=e=>{e.compute(ke(e.inputs[0],"Acosh","acosh"))},is=e=>{e.compute(ke(e.inputs[0],"Asin","asin"))},as=e=>{e.compute(ke(e.inputs[0],"Asinh","asinh"))},ss=e=>{e.compute(ke(e.inputs[0],"Atan","atan"))},us=e=>{e.compute(ke(e.inputs[0],"Atanh","atanh"))},ds=e=>ve(e),ls=(e,t)=>{let r;switch(t.to){case 10:r="vec4<f16>";break;case 1:r="vec4<f32>";break;case 12:r="vec4<u32>";break;case 6:r="vec4<i32>";break;case 9:r="vec4<bool>";break;default:throw new RangeError(`not supported type (specified in attribute \'to\' from \'Cast\' operator): ${t.to}`)}e.compute(ke(e.inputs[0],"Cast",r,void 0,t.cacheKey,t.to))},cc=e=>{let t=e.length>=2&&e[1].data!==0?e[1].getFloat32Array()[0]:xn,r=e.length>=3&&e[2].data!==0?e[2].getFloat32Array()[0]:Cn;return ve({min:t,max:r})},cs=(e,t)=>{let r=e.inputs.length===1?t:cc(e.inputs),o=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"Clip",i=>`clamp(${i}, clip_min_, clip_max_)`,`\n    const clip_min_: vec4<${o}> = vec4(${o}(${r.min}));\n    const clip_max_: vec4<${o}> = vec4(${o}(${r.max}));\n`,r.cacheKey),{inputs:[0]})},ps=e=>{e.compute(ke(e.inputs[0],"Ceil","ceil"))},ms=e=>{e.compute(ke(e.inputs[0],"Cos","cos"))},fs=e=>{e.compute(ke(e.inputs[0],"Cosh","cosh"))},kn=e=>ve(e),hs=(e,t)=>{let r=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"Elu",o=>`elu_vf32(${o})`,`\n  const elu_alpha_ = ${r}(${t.alpha});\n\n  fn elu_f32(a: ${r}) -> ${r} {\n  return select((exp(a) - 1.0) * elu_alpha_, a, a >= 0.0);\n  }\n\n  fn elu_vf32(v: vec4<${r}>) -> vec4<${r}> {\n  return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w));\n  }`,t.cacheKey))},On=(e="f32")=>`\nconst r0: ${e} = 0.3275911;\nconst r1: ${e} = 0.254829592;\nconst r2: ${e} = -0.284496736;\nconst r3: ${e} = 1.421413741;\nconst r4: ${e} = -1.453152027;\nconst r5: ${e} = 1.061405429;\n\nfn erf_vf32(v: vec4<${e}>) -> vec4<${e}> {\n  let absv = abs(v);\n  let x = 1.0 / (1.0 + r0 * absv);\n  return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv));\n}`,gs=e=>{let t=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"Erf",r=>`erf_vf32(${r})`,On(t)))},ys=e=>{e.compute(ke(e.inputs[0],"Exp","exp"))},bs=e=>{e.compute(ke(e.inputs[0],"Floor","floor"))},ws=e=>{let t=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,On(t)))},vs=(e,t)=>{let r=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"LeakyRelu",o=>`select(leaky_relu_alpha_ * ${o}, ${o}, ${o} >= vec4<${r}>(0.0))`,`const leaky_relu_alpha_ = ${r}(${t.alpha});`,t.cacheKey))},$s=e=>{e.compute(ke(e.inputs[0],"Not",t=>`!${t}`))},_s=e=>{e.compute(ke(e.inputs[0],"Neg",t=>`-${t}`))},Ss=e=>{e.compute(ke(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},xs=e=>{let t=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"Relu",r=>`select(vec4<${t}>(0.0), ${r}, ${r} > vec4<${t}>(0.0))`))},Cs=e=>{e.compute(ke(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},As=e=>ve(e),Is=(e,t)=>{let r=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"HardSigmoid",o=>`max(vec4<${r}>(0.0), min(vec4<${r}>(1.0), ${t.alpha} * ${o} + vec4<${r}>(${t.beta})))`,void 0,t.cacheKey))},Ts=e=>{e.compute(ke(e.inputs[0],"Sin","sin"))},Es=e=>{e.compute(ke(e.inputs[0],"Sinh","sinh"))},Ps=e=>{e.compute(ke(e.inputs[0],"Sqrt","sqrt"))},ks=e=>{e.compute(ke(e.inputs[0],"Tan","tan"))},Os=e=>`sign(${e}) * (1 - exp(-2 * abs(${e}))) / (1 + exp(-2 * abs(${e})))`,Rs=e=>{e.compute(ke(e.inputs[0],"Tanh",Os))},Bo=(e="f32")=>`\nconst fast_gelu_a: ${e} = 0.5;\nconst fast_gelu_b: ${e} = 0.7978845608028654;\nconst fast_gelu_c: ${e} = 0.035677408136300125;\n\nfn tanh_v(v: vec4<${e}>) -> vec4<${e}> {\n  return ${Os("v")};\n}\n`,Do=e=>`(fast_gelu_a + fast_gelu_a * tanh_v(${e} * (fast_gelu_c * ${e} * ${e} + fast_gelu_b))) * ${e}`,Bs=e=>{let t=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"FastGelu",Do,Bo(t),void 0,e.inputs[0].dataType))},Ds=(e,t)=>{let r=et(e.inputs[0].dataType);return e.compute(ke(e.inputs[0],"ThresholdedRelu",o=>`select(vec4<${r}>(0.0), ${o}, ${o} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${t.alpha});`,t.cacheKey)),0},zs=e=>{e.compute(ke(e.inputs[0],"Log","log"))}});var pc,mc,Us,Vs=Y(()=>{"use strict";Se();_e();Rn();pc=e=>{if(e[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![2560,5120,10240].includes(e[0].dims[2]))throw new Error("hidden state should be 2560, 5120 or 10240");if(e[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(e[0].dims[2]!==e[1].dims[0])throw new Error("last dimension of input and bias are not the same")},mc=e=>{let t=e[0].dims.slice();t[2]=t[2]/2;let r=U("input",e[0].dataType,e[0].dims,4),o=U("bias",e[0].dataType,[e[0].dims[2]],4),i=j("output",e[0].dataType,t,4),u=M.size(t)/4,a=De(e[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(u/64)}}),getShaderSource:p=>`\n  const M_SQRT2 = sqrt(2.0);\n  const halfChannels = ${e[0].dims[2]/4/2}u;\n\n  ${p.declareVariables(r,o,i)}\n\n  ${On(a)}\n\n  ${p.mainStart()}\n    ${p.guardAgainstOutOfBoundsWorkgroupSizes(u)}\n    let biasIdx = global_idx % halfChannels;\n    let batchIndex = global_idx / halfChannels;\n    let inputOffset = biasIdx + batchIndex * halfChannels * 2;\n    let valueLeft = input[inputOffset] + bias[biasIdx];\n    let valueRight = input[inputOffset + halfChannels] + bias[biasIdx + halfChannels];\n    let geluRight = valueRight * 0.5 * (erf_vf32(valueRight / M_SQRT2) + 1);\n\n    ${i.setByOffset("global_idx","valueLeft * geluRight")}\n  }`}},Us=e=>{pc(e.inputs),e.compute(mc(e.inputs))}});var fc,hc,Ot,Ws,Ns,Gs,Hs,Ls,Fs,qs,js,Ks,Ys,Zs=Y(()=>{"use strict";ye();Se();_e();fc=(e,t,r,o,i,u,a,c,p,h,d,y)=>{let w,_;typeof c=="string"?w=_=(x,E)=>`${c}((${x}),(${E}))`:typeof c=="function"?w=_=c:(w=c.scalar,_=c.vector);let v=j("outputData",d,o.length,4),S=U("aData",p,t.length,4),A=U("bData",h,r.length,4),I;if(i)if(u){let x=M.size(t)===1,E=M.size(r)===1,P=t.length>0&&t[t.length-1]%4===0,O=r.length>0&&r[r.length-1]%4===0;x||E?I=v.setByOffset("global_idx",_(x?`${S.type.value}(${S.getByOffset("0")}.x)`:S.getByOffset("global_idx"),E?`${A.type.value}(${A.getByOffset("0")}.x)`:A.getByOffset("global_idx"))):I=`\n            let outputIndices = ${v.offsetToIndices("global_idx * 4u")};\n            let offsetA = ${S.broadcastedIndicesToOffset("outputIndices",v)};\n            let offsetB = ${A.broadcastedIndicesToOffset("outputIndices",v)};\n            ${v.setByOffset("global_idx",_(a||P?S.getByOffset("offsetA / 4u"):`${S.type.value}(${S.getByOffset("offsetA / 4u")}[offsetA % 4u])`,a||O?A.getByOffset("offsetB / 4u"):`${A.type.value}(${A.getByOffset("offsetB / 4u")}[offsetB % 4u])`))}\n          `}else I=v.setByOffset("global_idx",_(S.getByOffset("global_idx"),A.getByOffset("global_idx")));else{if(!u)throw new Error("no necessary to use scalar implementation for element-wise binary op implementation.");let x=(E,P,O="")=>{let R=`aData[indexA${P}][componentA${P}]`,L=`bData[indexB${P}][componentB${P}]`;return`\n            let outputIndices${P} = ${v.offsetToIndices(`global_idx * 4u + ${P}u`)};\n            let offsetA${P} = ${S.broadcastedIndicesToOffset(`outputIndices${P}`,v)};\n            let offsetB${P} = ${A.broadcastedIndicesToOffset(`outputIndices${P}`,v)};\n            let indexA${P} = offsetA${P} / 4u;\n            let indexB${P} = offsetB${P} / 4u;\n            let componentA${P} = offsetA${P} % 4u;\n            let componentB${P} = offsetB${P} % 4u;\n            ${E}[${P}] = ${O}(${w(R,L)});\n          `};d===9?I=`\n            var data = vec4<u32>(0);\n            ${x("data",0,"u32")}\n            ${x("data",1,"u32")}\n            ${x("data",2,"u32")}\n            ${x("data",3,"u32")}\n            outputData[global_idx] = dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(data));`:I=`\n            ${x("outputData[global_idx]",0)}\n            ${x("outputData[global_idx]",1)}\n            ${x("outputData[global_idx]",2)}\n            ${x("outputData[global_idx]",3)}\n          `}return`\n        ${e.registerUniform("vec_size","u32").declareVariables(S,A,v)}\n\n        ${y??""}\n\n        ${e.mainStart()}\n        ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}\n        ${I}\n      }`},hc=(e,t,r,o,i,u,a=r.dataType)=>{let c=!M.areEqual(r.dims,o.dims),p=r.dims,h=M.size(r.dims),d=!1,y=!1,w=[c];if(c){let _=It.calcShape(r.dims,o.dims,!1);if(!_)throw new Error("Can\'t perform binary op on the given tensors");p=_,h=M.size(p);let v=M.size(r.dims)===1,S=M.size(o.dims)===1,A=r.dims.length>0&&r.dims[r.dims.length-1]%4===0,I=o.dims.length>0&&o.dims[o.dims.length-1]%4===0;w.push(v),w.push(S),w.push(A),w.push(I);let x=1;for(let E=1;E<p.length;E++){let P=r.dims[r.dims.length-E]??1,O=o.dims[o.dims.length-E]??1;if(P===O)x*=P;else break}x%4===0?(y=!0,d=!0):(v||S||A||I)&&(d=!0)}else d=!0;return w.push(d),{name:e,shaderCache:{hint:t+w.map(_=>_.toString()).join("_"),inputDependencies:["rank","rank"]},getShaderSource:_=>fc(_,r.dims,o.dims,p,d,c,y,i,r.dataType,o.dataType,a,u),getRunData:()=>({outputs:[{dims:p,dataType:a}],dispatchGroup:{x:Math.ceil(h/64/4)},programUniforms:[{type:12,data:Math.ceil(M.size(p)/4)},...Z(r.dims,o.dims,p)]})}},Ot=(e,t,r,o,i,u)=>{e.compute(hc(t,i??"",e.inputs[0],e.inputs[1],r,o,u))},Ws=e=>{Ot(e,"Add",(t,r)=>`${t}+${r}`)},Ns=e=>{Ot(e,"Div",(t,r)=>`${t}/${r}`)},Gs=e=>{Ot(e,"Equal",{scalar:(t,r)=>`u32(${t}==${r})`,vector:(t,r)=>`vec4<u32>(${t}==${r})`},void 0,void 0,9)},Hs=e=>{Ot(e,"Mul",(t,r)=>`${t}*${r}`)},Ls=e=>{let t=U("input",e.inputs[0].dataType,e.inputs[0].dims).type.value;Ot(e,"Pow",{scalar:(o,i)=>`pow_custom(${o},${i})`,vector:(o,i)=>`pow_vector_custom(${o},${i})`},`\n    fn pow_custom(a : ${t}, b : ${t}) -> ${t} {\n      if (b == ${t}(0.0)) {\n        return ${t}(1.0);\n      } else if (a < ${t}(0.0) && f32(b) != floor(f32(b))) {\n        return ${t}(pow(f32(a), f32(b))); // NaN\n      }\n      return select(sign(a), ${t}(1.0), round(f32(abs(b) % ${t}(2.0))) != 1.0) * ${t}(${t==="i32"?"round":""}(pow(f32(abs(a)), f32(b))));\n    }\n    fn pow_vector_custom(a : vec4<${t}>, b : vec4<${t}>) -> vec4<${t}> {\n      // TODO: implement vectorized pow\n      return vec4<${t}>(pow_custom(a.x, b.x), pow_custom(a.y, b.y), pow_custom(a.z, b.z), pow_custom(a.w, b.w));\n    }\n      `)},Fs=e=>{Ot(e,"Sub",(t,r)=>`${t}-${r}`)},qs=e=>{Ot(e,"Greater",{scalar:(t,r)=>`u32(${t}>${r})`,vector:(t,r)=>`vec4<u32>(${t}>${r})`},void 0,void 0,9)},js=e=>{Ot(e,"Less",{scalar:(t,r)=>`u32(${t}<${r})`,vector:(t,r)=>`vec4<u32>(${t}<${r})`},void 0,void 0,9)},Ks=e=>{Ot(e,"GreaterOrEqual",{scalar:(t,r)=>`u32(${t}>=${r})`,vector:(t,r)=>`vec4<u32>(${t}>=${r})`},void 0,void 0,9)},Ys=e=>{Ot(e,"LessOrEqual",{scalar:(t,r)=>`u32(${t}<=${r})`,vector:(t,r)=>`vec4<u32>(${t}<=${r})`},void 0,void 0,9)}});var St,xt,Ct,Bn,Ft=Y(()=>{"use strict";ye();Se();St=(e,t,r="f32")=>{switch(e.activation){case"Relu":return`value = max(value, ${t}(0.0));`;case"Sigmoid":return`value = (${t}(1.0) / (${t}(1.0) + exp(-value)));`;case"Clip":return`value = clamp(value, ${t}(${r}(uniforms.clip_min)), ${t}(${r}(uniforms.clip_max)));`;case"HardSigmoid":return`value = max(${t}(0.0), min(${t}(1.0), ${r}(uniforms.alpha) * value + ${r}(uniforms.beta)));`;case"LeakyRelu":return`value = select(${r}(uniforms.alpha) * value, value, value >= ${t}(0.0));`;case"":return"";default:throw new Error(`Unsupported activation ${e.activation}`)}},xt=(e,t)=>{e.activation==="Clip"?t.push({type:1,data:e.clipMax},{type:1,data:e.clipMin}):e.activation==="HardSigmoid"?t.push({type:1,data:e.alpha},{type:1,data:e.beta}):e.activation==="LeakyRelu"&&t.push({type:1,data:e.alpha})},Ct=(e,t)=>{e.activation==="Clip"?t.push({name:"clip_max",type:"f32"},{name:"clip_min",type:"f32"}):e.activation==="HardSigmoid"?t.push({name:"alpha",type:"f32"},{name:"beta",type:"f32"}):e.activation==="LeakyRelu"&&t.push({name:"alpha",type:"f32"})},Bn=e=>{let t=e?.activation||"";if(t==="HardSigmoid"){let[r,o]=e?.activation_params||[.2,.5];return{activation:t,alpha:r,beta:o}}else if(t==="Clip"){let[r,o]=e?.activation_params||[xn,Cn];return{activation:t,clipMax:o,clipMin:r}}else if(t==="LeakyRelu"){let[r]=e?.activation_params||[.01];return{activation:t,alpha:r}}return{activation:t}}});var tt,Dn,zn=Y(()=>{"use strict";tt=(e,t)=>{switch(e){case 1:return t;case 2:return`vec2<${t}>`;case 3:return`vec3<${t}>`;case 4:return`vec4<${t}>`;default:throw new Error(`${e}-component is not supported.`)}},Dn=e=>`\n      ${e?"value = value + getBiasByOutputCoords(coords);":""}\n      `});var Mn,zo=Y(()=>{"use strict";Mn=e=>`\nfn getIndexFromCoords4D(coords : vec4<i32>, shape : vec4<i32>) -> i32 {\n  return dot(coords, vec4<i32>(\n      shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1));\n}\nfn getOutputIndexFromCoords(coords : vec4<i32>) -> i32 {\n  return dot(coords, vec4<i32>(\n    i32(${e}.x), i32(${e}.y), i32(${e}.z), 1));\n}\n`});var yc,bc,Hr,Xs,wc,Lr,vc,Un,Fr=Y(()=>{"use strict";ye();Se();_e();Ft();zn();yc=(e,t)=>e?`\n        mm_Asub[inputRow][inputCol] = mm_readA(batch,\n          kStart + inputRow,\n          globalRowStart / innerElementSize + inputCol${t?", batchIndices":""});\n        `:`\n        mm_Asub[inputRow][inputCol] = mm_readA(batch,\n          globalRow + innerRow,\n          kStart / innerElementSize + inputCol${t?", batchIndices":""});\n        `,bc=(e,t)=>e?`\n        let ACached0 = mm_Asub[k * innerElementSize][localRow];\n        let ACached1 = mm_Asub[k * innerElementSize + 1][localRow];\n        let ACached2 = mm_Asub[k * innerElementSize + 2][localRow];\n        ${t===3?"":"let ACached3 = mm_Asub[k * innerElementSize + 3][localRow];"}\n        for (var i = 0; i < rowPerThread; i = i + 1) {\n          acc[i] = BCached0 * ACached0[i] + acc[i];\n          acc[i] = BCached1 * ACached1[i] + acc[i];\n          acc[i] = BCached2 * ACached2[i] + acc[i];\n          ${t===3?"":"acc[i] = BCached3 * ACached3[i] + acc[i];"}\n        }`:`\n        for (var i = 0; i < rowPerThread; i = i + 1) {\n          let ACached = mm_Asub[tileRow + i][k];\n          acc[i] = BCached0 * ACached.x + acc[i];\n          acc[i] = BCached1 * ACached.y + acc[i];\n          acc[i] = BCached2 * ACached.z + acc[i];\n          ${t===3?"":"acc[i] = BCached3 * ACached.w + acc[i];"}\n        }`,Hr=(e,t,r="f32",o,i=!1,u=32,a=!1,c=32)=>{let p=t[1]*e[1],h=t[0]*e[0],d=i?p:u,y=i?u:p,w=d/t[0],_=u/t[1];if(!((i&&w===4&&e[1]===4||!i&&(w===3||w===4))&&d%t[0]===0&&u%t[1]===0&&e[0]===4))throw new Error(`If transposeA ${i} is true, innerElementSize ${w} and workPerThread[1] ${e[1]} must be 4.\n      Otherwise, innerElementSize ${w} must be 3 or 4.\n  tileAWidth ${d} must be divisible by workgroupSize[0]${t[0]}. tileInner ${u} must be divisible by workgroupSize[1] ${t[1]}. colPerThread ${e[0]} must be 4.`);return`\nvar<workgroup> mm_Asub: array<array<vec${w}<${r}>, ${d/w}>, ${y}>;\nvar<workgroup> mm_Bsub: array<array<vec4<${r}>, ${h/e[0]}>, ${u}>;\n\nconst rowPerThread = ${e[1]};\nconst colPerThread = ${e[0]};\nconst innerElementSize = ${w};\nconst tileInner = ${u};\n\n@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]})\nfn main(@builtin(local_invocation_id) localId : vec3<u32>,\n        @builtin(global_invocation_id) globalId : vec3<u32>,\n        @builtin(workgroup_id) workgroupId : vec3<u32>) {\n  let localRow = i32(localId.y);\n  let tileRow = localRow * rowPerThread;\n  let tileCol = i32(localId.x);\n\n  let globalRow =i32(globalId.y) * rowPerThread;\n  let globalCol = i32(globalId.x);\n  let batch = ${a?"0":"i32(globalId.z)"};\n  ${o?`let batchIndices = ${o.offsetToIndices("u32(batch)")};`:""}\n  let globalRowStart = i32(workgroupId.y) * ${p};\n\n  let num_tiles = ${a?`${Math.ceil(c/u)}`:"(uniforms.dim_inner - 1) / tileInner + 1"};\n  var kStart = ${a?`i32(globalId.z) * ${c}`:"0"};\n\n  var acc: array<vec4<${r}>, rowPerThread>;\n\n  // Loop over shared dimension.\n  let tileRowB = localRow * ${_};\n  for (var t = 0; t < num_tiles; t = t + 1) {\n      // Load one tile of A into local memory.\n      for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n          let inputRow = tileRow + innerRow;\n          let inputCol = tileCol;\n          ${yc(i,o)}\n      }\n\n      // Load one tile of B into local memory.\n      for (var innerRow = 0; innerRow < ${_}; innerRow = innerRow + 1) {\n          let inputRow = tileRowB + innerRow;\n          let inputCol = tileCol;\n          mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${o?", batchIndices":""});\n      }\n      kStart = kStart + tileInner;\n      workgroupBarrier();\n\n      // Compute acc values for a single thread.\n      for (var k = 0; k < tileInner / innerElementSize; k = k + 1) {\n          let BCached0 = mm_Bsub[k * innerElementSize][tileCol];\n          let BCached1 = mm_Bsub[k * innerElementSize + 1][tileCol];\n          let BCached2 = mm_Bsub[k * innerElementSize + 2][tileCol];\n          ${w===3?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"}\n\n          ${bc(i,w)}\n      }\n\n      workgroupBarrier();\n  }\n\n  for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n      mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]);\n  }\n}`},Xs=(e,t)=>e?`\n            mm_Asub[inputRow][inputCol] = mm_readA(batch,\n              kStart + inputRow,\n              globalRowStart + inputCol${t?", batchIndices":""});\n            `:`\n            mm_Asub[inputRow][inputCol] = mm_readA(batch,\n              globalRowStart + inputRow,\n              kStart + inputCol${t?", batchIndices":""});\n            `,wc=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",Lr=(e,t,r="f32",o,i=!1,u=32,a=!1,c=32,p=!1)=>{let h=e[1]*t[1],d=e[0]*t[0],y=i?h:u,w=i?u:h;if(!(w%t[1]===0&&y%t[0]===0&&u%t[1]===0))throw new Error(`tileAHight ${w} must be divisible by workgroupSize[1]${t[1]}, tileAWidth ${y} must be divisible by workgroupSize[0]${t[0]}, tileInner ${u} must be divisible by workgroupSize[1]${t[1]}`);let _=w/t[1],v=y/t[0],S=u/t[1],A=p?`\n    let localRow = i32(localId.y);\n    let localCol = i32(localId.x);\n    let globalRowStart = i32(workgroupId.y) * ${h};\n    let globalColStart = i32(workgroupId.x) * ${d};\n\n    // Loop over shared dimension.\n    for (var t = 0; t < num_tiles; t = t + 1) {\n      // Load one tile of A into local memory.\n      for (var inputRow = localRow; inputRow < ${w}; inputRow = inputRow + ${t[1]}) {\n        for (var inputCol = localCol; inputCol < ${y}; inputCol = inputCol + ${t[0]}) {\n          ${Xs(i,o)}\n        }\n      }\n      // Load one tile of B into local memory.\n      for (var inputRow = localRow; inputRow < ${u}; inputRow = inputRow + ${t[1]}) {\n            for (var inputCol = localCol; inputCol < ${d}; inputCol = inputCol + ${t[0]}) {\n          mm_Bsub[inputRow][inputCol] = mm_readB(batch,\n            kStart + inputRow,\n            globalColStart + inputCol${o?", batchIndices":""});\n        }\n      }\n      kStart = kStart + tileInner;\n      workgroupBarrier();\n\n      // Compute acc values for a single thread.\n      var BCached : array<${r}, colPerThread>;\n      for (var k = 0; k < tileInner; k = k + 1) {\n        for (var inner = 0; inner < colPerThread; inner = inner + 1) {\n          BCached[inner] = mm_Bsub[k][localCol + inner * ${t[0]}];\n        }\n        for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n          let ACached = ${i?`mm_Asub[k][localRow + innerRow * ${t[1]}];`:`mm_Asub[localRow + innerRow * ${t[1]}][k];`}\n          for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n            acc[innerRow][innerCol] = acc[innerRow][innerCol] +\n                ACached * BCached[innerCol];\n          }\n        }\n      }\n      workgroupBarrier();\n    }\n    for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n      let gRow = globalRowStart + localRow + innerRow * ${t[1]};\n      for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n        let gCol = globalColStart + localCol + innerCol * ${t[0]};\n        mm_write(batch, gRow, gCol, acc[innerRow][innerCol]);\n      }\n    }\n    `:`\nlet tileRow = i32(localId.y) * rowPerThread;\nlet tileCol = i32(localId.x) * colPerThread;\n\nlet globalRow = i32(globalId.y) * rowPerThread;\nlet globalCol = i32(globalId.x) * colPerThread;\nlet globalRowStart = i32(workgroupId.y) * ${h};\n\nlet tileRowA = i32(localId.y) * ${_};\nlet tileColA = i32(localId.x) * ${v};\nlet tileRowB = i32(localId.y) * ${S};\n// Loop over shared dimension.\nfor (var t = 0; t < num_tiles; t = t + 1) {\n  // Load one tile of A into local memory.\n  for (var innerRow = 0; innerRow < ${_}; innerRow = innerRow + 1) {\n    for (var innerCol = 0; innerCol < ${v}; innerCol = innerCol + 1) {\n      let inputRow = tileRowA + innerRow;\n      let inputCol = tileColA + innerCol;\n      ${Xs(i,o)}\n    }\n  }\n\n  // Load one tile of B into local memory.\n  for (var innerRow = 0; innerRow < ${S}; innerRow = innerRow + 1) {\n    for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n      let inputRow = tileRowB + innerRow;\n      let inputCol = tileCol + innerCol;\n      mm_Bsub[inputRow][inputCol] = mm_readB(batch,\n        kStart + inputRow,\n        globalCol + innerCol${o?", batchIndices":""});\n    }\n  }\n  kStart = kStart + tileInner;\n  workgroupBarrier();\n\n  // Compute acc values for a single thread.\n  var BCached : array<${r}, colPerThread>;\n  for (var k = 0; k < tileInner; k = k + 1) {\n    for (var inner = 0; inner < colPerThread; inner = inner + 1) {\n      BCached[inner] = mm_Bsub[k][tileCol + inner];\n    }\n\n    for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n      ${wc(i)}\n      for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n        acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol];\n      }\n    }\n  }\n\n  workgroupBarrier();\n}\n\nfor (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n  for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n    mm_write(batch, globalRow + innerRow, globalCol + innerCol,\n        acc[innerRow][innerCol]);\n  }\n}\n`;return`\n  var<workgroup> mm_Asub : array<array<${r}, ${y}>, ${w}>;\n  var<workgroup> mm_Bsub : array<array<${r}, ${d}>, ${u}>;\n  const rowPerThread = ${e[1]};\n  const colPerThread = ${e[0]};\n  const tileInner = ${u};\n\n@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]})\nfn main(@builtin(local_invocation_id) localId : vec3<u32>,\n        @builtin(global_invocation_id) globalId : vec3<u32>,\n        @builtin(workgroup_id) workgroupId : vec3<u32>) {\n    let batch = ${a?"0":"i32(globalId.z)"};\n    ${o?`let batchIndices = ${o.offsetToIndices("u32(batch)")};`:""}\n    let num_tiles = ${a?`${Math.ceil(c/u)}`:"(uniforms.dim_inner - 1) / tileInner + 1"};\n    var kStart = ${a?`i32(globalId.z) * ${c}`:"0"};\n\n    var acc : array<array<${r}, colPerThread>, rowPerThread>;\n\n    // Without this initialization strange values show up in acc.\n    for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n      for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n        acc[innerRow][innerCol] = 0.0;\n      }\n    }\n    ${A}\n  }\n`},vc=(e,t,r,o,i,u=!1)=>{let[a,c,p]=i,[h,d,y,w]=o,_=_r(a,p),v=_r(c,p),S=De(o[0].type.tensor),A=()=>{let E=d.rank,P=h.rank,O=`var aIndices: ${d.type.indices};`;for(let R=E-2-1,L=P-1;R>=0;R--,L--)O+=`\naIndices[${R}] = ${P>1?`batchIndices[${L}]`:"batchIndices"};`;return _.forEach(R=>{O+=`\naIndices[${R}] = 0;`}),O+=`\naIndices[${E-2}] = u32(row);\n                   aIndices[${E-1}] = u32(colIn);`,O},I=()=>{let E=y.rank,P=h.rank,O=`var bIndices: ${y.type.indices};`;for(let R=E-2-1,L=P-1;R>=0;R--,L--)O+=`\nbIndices[${R}] = ${P>1?`batchIndices[${L}]`:"batchIndices"};`;return v.forEach(R=>{O+=`\nbIndices[${R}] = 0;`}),O+=`\nbIndices[${E-2}] = u32(row);\n                   bIndices[${E-1}] = u32(colIn);`,O};return`\n    fn mm_readA(batch: i32, row: i32, colIn: i32, batchIndices: ${h.type.indices}) -> ${tt(e,S)} {\n      var value = ${tt(e,S)}(0.0);\n      let col = colIn * ${e};\n      if(row < uniforms.dim_a_outer && col < uniforms.dim_inner)\n      {\n        ${A()}\n        value = ${d.getByIndices("aIndices")};\n      }\n      return value;\n    }\n\n    fn mm_readB(batch: i32, row: i32, colIn: i32, batchIndices: ${h.type.indices}) -> ${tt(e,S)} {\n      var value = ${tt(e,S)}(0.0);\n      let col = colIn * ${e};\n      if(row < uniforms.dim_inner && col < uniforms.dim_b_outer)\n      {\n        ${I()}\n        value = ${y.getByIndices("bIndices")};\n      }\n      return value;\n    }\n\n    fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${tt(e,S)}) {\n      let col = colIn * ${e};\n      if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) {\n        var value = valueIn;\n        let coords = vec3<i32>(batch, row, colIn);\n        ${t?`value = value + ${u?"bias[colIn]":`${tt(e,S)}(bias[row])`};`:""}\n        ${r}\n        ${w.setByIndices("vec3<u32>(coords)","value")}\n      }\n    }\n    `},Un=(e,t,r,o,i=!1)=>{let u=e[0].dims,a=e[1].dims,c=u.slice(0,-2),p=a.slice(0,-2),h=o?o.slice(0,-2):r.slice(0,-2),d=M.size(h),y=u[u.length-2],w=u[u.length-1],_=a[a.length-1],v=w%4===0&&_%4===0,S=y<=8?[4,1,1]:[4,4,1],A=[8,8,1],I=[Math.ceil(_/A[0]/S[0]),Math.ceil(y/A[1]/S[1]),Math.ceil(d/A[2]/S[2])],x=v?4:1,E=[...c,y,w/x],P=E.length,O=[...p,w,_/x],R=O.length,L=[d,y,_/x],N=[{type:6,data:y},{type:6,data:_},{type:6,data:w}];xt(t,N),N.push(...Z(h,E,O));let K=["rank","rank"],Q=e.length>2;Q&&(N.push(...Z(e[2].dims)),K.push("rank")),N.push(...Z(L));let he=W=>{let se=h.length,Ce=An("batchDims",e[0].dataType,se,1),We=De(e[0].dataType),ee=U("a",e[0].dataType,P,x),ae=U("b",e[1].dataType,R,x),Ae=j("result",e[0].dataType,L.length,x),me=[ee,ae];if(Q){let G=i?x:1;me.push(U("bias",e[2].dataType,e[2].dims.length,G))}let ie=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"}];Ct(t,ie);let ue=De(Ae.type.tensor),le=St(t,Ae.type.value,ue),qe=vc(x,Q,le,[Ce,ee,ae,Ae],[c,p,h],i);return`\n  ${W.registerUniforms(ie).registerInternalVariables(Ce).declareVariables(...me,Ae)}\n  ${qe}\n  ${v?Hr(S,A,We,Ce):Lr(S,A,We,Ce)}\n                   `};return{name:"MatMul",shaderCache:{hint:`${S};${t.activation};${v};${i}`,inputDependencies:K},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:I[0],y:I[1],z:I[2]},programUniforms:N}),getShaderSource:he}}});var $c,Qs,Js=Y(()=>{"use strict";ye();Lt();_e();Ft();zn();zo();Fr();$c=(e,t,r,o,i=!1,u,a=4,c=4,p=4,h="f32")=>{let d=Q=>{switch(Q){case 1:return"resData = x[xIndex];";case 3:return`resData = vec3<${h}>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);`;case 4:return"resData = x[xIndex / 4];";default:throw new Error(`innerElementSize ${Q} is not supported.`)}},y=Q=>{switch(Q){case 1:return"return w[row * i32(uniforms.w_shape[3]) + colIn];";case 4:return"return w[row * i32(uniforms.w_shape[3]) / 4 + colIn];";default:throw new Error(`innerElementSize ${Q} is not supported.`)}},w=e?`\n    let coord = vec4<i32>(batch, xRow, xCol, xCh);\n    `:`\n    let coord = vec4<i32>(batch, xCh, xRow, xCol);\n    `,_=e?`\n    let coords = vec4<i32>(\n      batch,\n      row / outWidth,\n      row % outWidth,\n      col);\n    `:`\n    let coords = vec4<i32>(\n      batch,\n      row,\n      col / outWidth,\n      col % outWidth);\n    `,v=e?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",S=e?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",A=e?"row":"col",I=e?"col":"row",x=`\n    let inChannels = i32(uniforms.w_shape[2]);\n    let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};\n    let outRow = ${A} / outWidth;\n    let outCol = ${A} % outWidth;\n\n    let WRow = ${I} / (i32(uniforms.w_shape[1]) * inChannels);\n    let WCol = ${I} / inChannels % i32(uniforms.w_shape[1]);\n    let xRow = outRow * uniforms.stride[0] + uniforms.dilation[0] * WRow - uniforms.pad[0];\n    let xCol = outCol * uniforms.stride[1] + uniforms.dilation[1] * WCol - uniforms.pad[1];\n    let xCh = ${I} % inChannels;\n    var resData = ${tt(a,h)}(0.0);\n    // The bounds checking is always needed since we use it to pad zero for\n    // the \'same\' padding type.\n    if (xRow >= 0 && xRow < ${v} && xCol >= 0 && xCol < ${S}) {\n      ${w}\n      let xIndex = getIndexFromCoords4D(coord, vec4<i32>(uniforms.x_shape));\n      ${d(a)}\n    }\n    return resData;`,E=e?t&&o?`\n    let col = colIn * ${a};\n    ${x}`:`\n    let col = colIn * ${a};\n    if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) {\n      ${x}\n    }\n    return ${tt(a,h)}(0.0);`:o&&r?`\n    let col = colIn * ${a};\n    ${x}`:`\n    let col = colIn * ${a};\n    if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) {\n      ${x}\n    }\n    return ${tt(a,h)}(0.0);`,P=`${y(c)}`,O=tt(p,h),R=e?tt(a,h):tt(c,h),L=e?tt(c,h):tt(a,h),N=St(u,O,h);return`\n    fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${R} {\n      ${e?E:P}\n    }\n\n    fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${L} {\n      ${e?P:E}\n    }\n\n    fn mm_write(batch: i32, row : i32, colIn : i32, valueIn : ${O}) {\n      let col = colIn * ${p};\n      if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer)\n      {\n      var value = valueIn;\n      let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};\n      ${_}\n      ${Dn(i)}\n      ${N}\n      setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);\n      }\n    }`},Qs=(e,t,r,o,i,u,a,c)=>{let p=t.format==="NHWC",h=p?e[0].dims[3]:e[0].dims[1],d=r[0],y=p?r[2]:r[3],w=p?r[1]:r[2],_=p?r[3]:r[1],v=p&&(h%4===0||h%3===0)&&_%4===0,S=p?_:y*w,A=p?y*w:_,I=[8,8,1],x=o<=8?[4,1,1]:[4,4,1],E=[Math.ceil(S/I[0]/x[0]),Math.ceil(A/I[1]/x[1]),Math.ceil(d/I[2]/x[2])];Ve("verbose",()=>`[conv2d_mm_webgpu] dispatch = ${E}`);let P=v?p&&h%4!==0?3:4:1,O=I[1]*x[1],R=I[0]*x[0],L=Math.max(I[0]*P,I[1]),N=o%O===0,K=i%R===0,Q=u%L===0,he=v?[P,4,4]:[1,1,1],W=[{type:6,data:o},{type:6,data:i},{type:6,data:u},{type:6,data:[t.pads[0],t.pads[1]]},{type:6,data:t.strides},{type:6,data:t.dilations}];xt(t,W),W.push(...Z(e[0].dims,e[1].dims));let se=["rank","rank"];a&&(W.push(...Z(e[2].dims)),se.push("rank")),W.push(...Z(r));let Ce=We=>{let ee=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"},{name:"pad",type:"i32",length:2},{name:"stride",type:"i32",length:2},{name:"dilation",type:"i32",length:2}];Ct(t,ee);let ae=v?4:1,Ae=De(e[0].dataType),me=`\n      fn setOutputAtIndex(flatIndex : i32, value : ${v?`vec4<${Ae}>`:Ae}) {\n        result[flatIndex] = ${v?`vec4<${Ae}>`:Ae}(value);\n      }\n      fn setOutputAtCoords(d0 : i32, d1 : i32, d2 : i32, d3 : i32, value : ${v?`vec4<${Ae}>`:Ae}) {\n        let flatIndex = getOutputIndexFromCoords(vec4<i32>(d0, d1, d2, d3));\n        setOutputAtIndex(flatIndex ${v?"/ 4":""}, value);\n      }`,ie=U("x",e[0].dataType,e[0].dims.length,P===3?1:P),ue=U("w",e[1].dataType,e[1].dims.length,ae),le=[ie,ue],qe=j("result",e[0].dataType,r.length,ae);if(a){let G=U("bias",e[2].dataType,e[2].dims.length,ae);le.push(G),me+=`\n        fn getBiasByOutputCoords(coords : vec4<i32>) -> ${v?`vec4<${Ae}>`:Ae} {\n          return bias[coords.${p?"w":"y"}${v?"/ 4":""}];\n        }`}return`\n        ${Mn("uniforms.result_strides")}\n        //struct Uniforms { xShape : vec4<i32>, wShape : vec4<i32>, outShape : vec4<i32>,\n        //  outShapeStrides: vec3<i32>, filterDims : vec2<i32>, pad : vec2<i32>, stride : vec2<i32>,\n        //  dilation : vec2<i32>, dimAOuter : i32, dimBOuter : i32, dimInner : i32 };\n        ${We.registerUniforms(ee).declareVariables(...le,qe)}\n        ${me}\n        ${$c(p,N,K,Q,a,t,he[0],he[1],he[2],Ae)}\n        ${v?Hr(x,I,Ae,void 0,!p,L):Lr(x,I,Ae,void 0,!p,L,!1,void 0,c)}`};return{name:"Conv2DMatMul",shaderCache:{hint:`${t.cacheKey};${P};${v};${N};${K};${Q};${O};${R};${L}`,inputDependencies:se},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:E[0],y:E[1],z:E[2]},programUniforms:W}),getShaderSource:Ce}}});var Mo,eu,tu=Y(()=>{"use strict";ye();Se();_e();Uo();Ft();Mo=(e,t,r)=>{let o=e.length>2,i=o?"value += b[output_channel];":"",u=e[0].dims,a=e[1].dims,c=a[0]/t.group,p=t.format==="NHWC",h=Vn(u,a,t.dilations,t.pads,t.strides,p),d=M.size(h),y=[{type:12,data:d},{type:12,data:t.dilations},{type:12,data:[t.strides[0],t.strides[1]]},{type:12,data:[t.pads[0],t.pads[1]]},{type:12,data:c}];xt(t,y),y.push(...Z(u,a));let w=["rank","rank"];o&&(y.push(...Z(e[2].dims)),w.push("rank")),y.push(...Z(h));let _=v=>{let S=j("output",e[0].dataType,h.length),A=De(S.type.tensor),I=St(t,S.type.value,A),x=U("x",e[0].dataType,u.length),E=U("w",e[1].dataType,a.length),P=[x,E];o&&P.push(U("b",e[2].dataType,e[2].dims.length));let O=[{name:"output_size",type:"u32"},{name:"dilations",type:"u32",length:t.dilations.length},{name:"strides",type:"u32",length:2},{name:"pads",type:"u32",length:2},{name:"output_channels_per_group",type:"u32"}];return Ct(t,O),`\n  ${v.registerUniforms(O).declareVariables(...P,S)}\n\n  ${v.mainStart()}\n    ${v.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n    let outputIndices = ${S.offsetToIndices("global_idx")};\n    let batch: u32 = outputIndices[0];\n    let output_channel: u32 = outputIndices[${p?3:1}];\n    let xRCCorner: vec2<u32> = vec2<u32>(outputIndices[${p?1:2}], outputIndices[${p?2:3}]) * uniforms.strides - uniforms.pads;\n    let group_id: u32 = output_channel / uniforms.output_channels_per_group;\n\n    var value: ${S.type.value} = ${S.type.value}(0);\n    for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[1]; wInChannel++) {\n      let input_channel = group_id * uniforms.w_shape[1] + wInChannel;\n      for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[2]; wHeight++) {\n        let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0];\n\n        if (xHeight < 0u || xHeight >= uniforms.x_shape[${p?1:2}]) {\n          continue;\n        }\n\n        for (var wWidth: u32 = 0u; wWidth < uniforms.w_shape[3]; wWidth++) {\n          let xWidth = xRCCorner.y + wWidth * uniforms.dilations[1];\n          if (xWidth < 0u || xWidth >= uniforms.x_shape[${p?2:3}]) {\n            continue;\n          }\n\n          let xVal = ${p?x.get("batch","xHeight","xWidth","input_channel"):x.get("batch","input_channel","xHeight","xWidth")};\n          let wVal = ${E.get("output_channel","wInChannel","wHeight","wWidth")};\n          value += xVal*wVal;\n        }\n      }\n    }\n    ${i}\n    ${I}\n    ${S.setByOffset("global_idx","value")}\n  }`};return{name:"GroupedConv",shaderCache:{hint:t.cacheKey,inputDependencies:w},getRunData:()=>({outputs:[{dims:r?r(h):h,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:y}),getShaderSource:_}},eu=(e,t,r)=>{let o=e.length>2,i=Me(r[3]),u=Me(r[2]),a=M.size(r)/i/u,c=[e[0].dims[0],e[0].dims[1],e[0].dims[2],e[0].dims[3]/i],p=[e[1].dims[0],e[1].dims[1],e[1].dims[2],e[1].dims[3]/i],h=[r[0],r[1],r[2],r[3]/i],d=[{type:12,data:a},{type:6,data:[t.strides[0],t.strides[1]]},{type:6,data:[t.pads[0],t.pads[1]]}];xt(t,d),d.push(...Z(c,p,h));let y=(u-1)*t.strides[1]+p[1],w=_=>{let v=j("output",e[0].dataType,h.length,i),S=De(v.type.tensor),A=St(t,v.type.value,S),I=U("x",e[0].dataType,c.length,i),x=U("w",e[1].dataType,p.length,i),E=[I,x];o&&E.push(U("b",e[2].dataType,e[2].dims,i));let P=o?"value += b[output_channel];":"",O=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return Ct(t,O),`\n  ${_.registerUniforms(O).declareVariables(...E,v)}\n  ${_.mainStart()}\n    ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n    let width0 = uniforms.output_shape[3];\n    let output_channel = global_idx % width0;\n    var index1 = global_idx / width0;\n    let width1 = uniforms.output_shape[2] / ${u}u;\n    let col = (index1 % width1) * ${u}u;\n    index1 = index1 / width1;\n    let row = index1 % uniforms.output_shape[1];\n    let batch = index1 / uniforms.output_shape[1];\n\n    let x_corner = vec2<i32>(i32(row), i32(col)) * uniforms.strides - uniforms.pads;\n\n    var x_vals: array<${I.type.value}, ${y}>;\n    var values: array<${v.type.value}, ${u}>;\n    let input_channel = output_channel;\n    // Use constant instead of uniform can give better performance for w\'s height/width.\n    for (var w_height: u32 = 0u; w_height < ${p[0]}; w_height++) {\n      let x_height = x_corner.x + i32(w_height);\n      if (x_height >= 0 && u32(x_height) < uniforms.x_shape[1]) {\n        for (var i = 0; i < ${y}; i++) {\n          let x_width = x_corner.y + i;\n          if (x_width >= 0 && u32(x_width) < uniforms.x_shape[2]) {\n            x_vals[i] = ${I.get("batch","u32(x_height)","u32(x_width)","input_channel")};\n          } else {\n            x_vals[i] = ${I.type.value}(0);\n          }\n        }\n        for (var w_width: u32 = 0u; w_width < ${p[1]}; w_width++) {\n          let w_val = ${x.get("w_height","w_width","0","output_channel")};\n          for (var i = 0u; i < ${u}u; i++) {\n            values[i] = fma(x_vals[i * u32(uniforms.strides[1]) + w_width], w_val, values[i]);\n          }\n        }\n      }\n    }\n\n    for (var i = 0u; i < ${u}u; i++) {\n      var value = values[i];\n      ${P}\n      ${A}\n      ${v.set("batch","row","col + i","output_channel","value")};\n    }\n  }`};return{name:"GroupedConv-Vectorize",shaderCache:{hint:`${t.cacheKey};${i};${u};${y};${p[0]};${p[1]}`,inputDependencies:o?["rank","rank","type"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:d}),getShaderSource:w}}});var Vo,_c,ru,Wo=Y(()=>{"use strict";ye();Se();Fr();_e();Ft();Vo=(e,t,r,o,i=!1)=>{let u=e[0].dims,a=e[1].dims,c=u[u.length-2],p=a[a.length-1],h=u[u.length-1],d=Me(p),y=Me(h),w=Me(c),_=M.size(r)/d/w,v=e.length>2,S=o?o.slice(0,-2):r.slice(0,-2),I=[M.size(S),c,p],x=[{type:12,data:_},{type:12,data:c},{type:12,data:p},{type:12,data:h}];xt(t,x),x.push(...Z(S,u,a)),v&&x.push(...Z(e[2].dims)),x.push(...Z(I));let E=P=>{let O=An("batch_dims",e[0].dataType,S.length),R=U("a",e[0].dataType,u.length,y),L=U("b",e[1].dataType,a.length,d),N=j("output",e[0].dataType,I.length,d),K=De(N.type.tensor),Q=St(t,N.type.value,K),he=[R,L],W="";if(v){let ie=i?d:1;he.push(U("bias",e[2].dataType,e[2].dims.length,ie)),W=`${i?`value += bias[col / ${ie}];`:`value += ${N.type.value}(bias[row + i]);`}`}let se=u.slice(0,-2),Ce=a.slice(0,-2),We=_r(se,S),ee=_r(Ce,S),ae=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];Ct(t,ae);let Ae=(ie,ue)=>{let le=ie.rank,qe=ie.name;if(le===2)return`var ${qe}_indices = ${ie.type.indices}(0u, 0u);`;let G=O.rank,ne=`var ${qe}_indices: ${ie.type.indices};`;for(let xe=le-2-1,Ke=G-1;xe>=0;xe--,Ke--)ne+=`\n${qe}_indices[${xe}] = ${G>1?`batch_indices[${Ke}]`:"batch_indices"};`;return ue.forEach(xe=>{ne+=`\n${qe}_indices[${xe}] = 0;`}),ne+=`${qe}_indices[${le-2}] = 0u;\n                     ${qe}_indices[${le-1}] = 0u;`,ne},me=()=>{let ie=`var a_data: ${R.type.value};`;for(let ue=0;ue<y;ue++)ie+=`\n              let b_data${ue} = b[(b_offset + (k + ${ue}) * uniforms.N + col) / ${d}];`;for(let ue=0;ue<w;ue++){ie+=`a_data = a[(a_offset + (row + ${ue}) * uniforms.K + k) / ${y}];`;for(let le=0;le<y;le++)ie+=`\n            values[${ue}] = fma(${L.type.value}(a_data${y===1?"":`[${le}]`}), b_data${le}, values[${ue}]);\n`}return ie};return`\n  ${P.registerUniforms(ae).registerInternalVariables(O).declareVariables(...he,N)}\n  ${P.mainStart()}\n    ${P.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n    let col = (global_idx % (uniforms.N / ${d})) * ${d};\n    var index1 = global_idx / (uniforms.N / ${d});\n    let stride1 = uniforms.M / ${w};\n    let row = (index1 % stride1) * ${w};\n    let batch = index1 / stride1;\n\n    ${r.length===2?"":`let batch_indices = ${O.offsetToIndices("batch")};`}\n    ${Ae(R,We)}\n    let a_offset = ${R.indicesToOffset("a_indices")};\n    ${Ae(L,ee)}\n    let b_offset = ${L.indicesToOffset("b_indices")};\n    var values: array<${N.type.value}, ${w}>;\n    for (var k: u32 = 0u; k < uniforms.K; k = k + ${y}) {\n      ${me()}\n    }\n    for (var i = 0u; i < ${w}u; i++) {\n      var value = values[i];\n      ${W}\n      ${Q}\n      let cur_indices = ${N.type.indices}(batch, row + i, col);\n      let offset = ${N.indicesToOffset("cur_indices")};\n      ${N.setByOffset(`offset / ${d}`,"value")};\n    }\n  }\n  `};return{name:"MatMulNaive",shaderCache:{hint:`${t.activation};${d};${y};${w};${i}`,inputDependencies:v?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(_/64)},programUniforms:x}),getShaderSource:E}},_c=e=>{if(!e||e.length!==2)throw new Error("MatMul requires 2 inputs.");if(e[0].dims[e[0].dims.length-1]!==e[1].dims[e[1].dims.length-2])throw new Error("shared dimension does not match.")},ru=e=>{_c(e.inputs);let t=It.calcShape(e.inputs[0].dims,e.inputs[1].dims,!0);if(!t)throw new Error("Can\'t use matmul on the given tensors");let r=t[t.length-1],o=e.inputs[0].dims[e.inputs[0].dims.length-1];r<8&&o<8?e.compute(Vo(e.inputs,{activation:""},t)):e.compute(Un(e.inputs,{activation:""},t))}});var Vn,No,Sc,nu,Go,xc,Cc,Ho,Uo=Y(()=>{"use strict";Se();Js();Fr();tu();Ft();Wo();Sr();Vn=(e,t,r,o,i,u)=>{let a=e[0],c=e.slice(u?1:2,u?3:4),p=c.length,h=t[0],y=t.slice(2).map((v,S)=>v+(v-1)*(r[S]-1)),_=c.map((v,S)=>v+o[S]+o[S+p]).map((v,S)=>Math.floor((v-y[S]+i[S])/i[S]));return _.splice(0,0,a),_.splice(u?3:1,0,h),_},No=[2,3,1,0],Sc=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length!==4&&e[0].dims.length!==3)throw new Error("currently only support conv 1D and 2D");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let r=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],o=e[1].dims[1]*t.group;if(r!==o)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");if(e.length===3&&(e[2].dims.length!==1||e[1].dims[0]!==e[2].dims[0]))throw new Error("invalid bias");let i=e[0].dims.length-2;if(t.dilations.length!==i)throw new Error(`dilations should be ${i}D`);if(t.strides.length!==i)throw new Error(`strides should be ${i}D`);if(t.pads.length!==i*2)throw new Error(`pads should be ${i*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape")},nu=(e,t)=>{let r=e.kernelShape.slice();for(let u=2;u<t[1].dims.length;++u)r[u-2]===0&&(r[u-2]=t[1].dims[u]);let o=e.pads.slice();nr.adjustPadsBasedOnAutoPad(t[0].dims,e.strides,e.dilations,r,o,e.format==="NHWC",e.autoPad);let i=Object.assign({},e);return Object.assign(i,{kernelShape:r,pads:o}),i},Go=e=>{let t=Bn(e),r=e.format,o=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],i=e.dilations,u=e.group,a=e.kernel_shape,c=e.pads,p=e.strides,h=e.w_is_const();return{autoPad:o,format:r,dilations:i,group:u,kernelShape:a,pads:c,strides:p,wIsConst:h,...t,cacheKey:`${e.format};${t.activation};`}},xc=(e,t,r)=>{let o=nu(r,t),i=r.format==="NHWC";if(r.group!==1){if(!e.adapterInfo.isArchitecture("ampere")&&i&&t[1].dims[0]===r.group&&t[1].dims[1]===1&&r.dilations[0]===1&&r.dilations[1]===1){let L=Vn(t[0].dims,t[1].dims,r.dilations,o.pads,r.strides,i),N=e.kernelCustomData.wT??e.compute(yt(t[1],No),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=N);let K=[t[0],N];t.length===3&&K.push(t[2]),e.compute(eu(K,o,L),{inputs:K})}else e.compute(Mo(t,o));return}let u=t.length===3,a=t[0].dims[i?1:2],c=t[0].dims[i?2:3],p=t[0].dims[i?3:1],h=t[1].dims[2],d=t[1].dims[3],y=Vn(t[0].dims,t[1].dims,r.dilations,o.pads,r.strides,i),w=y[i?1:2],_=y[i?2:3],v=y[i?3:1],S=i&&h===a&&d===c&&r.pads[0]===0&&r.pads[1]===0;if(S||h===1&&d===1&&r.dilations[0]===1&&r.dilations[1]===1&&r.strides[0]===1&&r.strides[1]===1&&r.pads[0]===0&&r.pads[1]===0){let R=y[0],L,N,K,Q=[];if(i){let se=e.kernelCustomData.wT??e.compute(yt(t[1],No),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=se),S){let Ce=a*c*p;L=t[0].reshape([1,R,Ce]),N=se.reshape([1,Ce,v]),K=[1,R,v]}else L=t[0].reshape([R,a*c,p]),N=se.reshape([1,p,v]),K=[R,w*_,v];Q.push(L),Q.push(N)}else L=t[0].reshape([R,p,a*c]),N=t[1].reshape([1,v,p]),K=[R,v,w*_],Q.push(N),Q.push(L);u&&Q.push(t[2]);let he=K[2],W=Q[0].dims[Q[0].dims.length-1];he<8&&W<8?e.compute(Vo(Q,o,y,K,i),{inputs:Q}):e.compute(Un(Q,o,y,K,i),{inputs:Q});return}let A=!0,I=e.kernelCustomData.wT??e.compute(yt(t[1],No),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=I);let x=[t[0],I];u&&x.push(t[2]);let E=i?w*_:v,P=i?v:w*_,O=h*d*p;e.compute(Qs(x,o,y,E,P,O,u,A),{inputs:x})},Cc=(e,t)=>{let r=t.format==="NHWC",o=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&o.push(e.inputs[2]);let i=[0,t.pads[0],0,t.pads[1]],u=[1].concat(t.strides),a=[1].concat(t.dilations),c=[1].concat(t.kernelShape),p=nu({...t,pads:i,strides:u,dilations:a,kernelShape:c},o);e.compute(Mo(o,p,h=>r?[h[0],h[2],h[3]]:[]))},Ho=(e,t)=>{Sc(e.inputs,t),e.inputs[0].dims.length===3?Cc(e,t):xc(e,e.inputs,t)}});var Ac,ou,iu=Y(()=>{"use strict";ye();Lt();_e();Ft();zn();zo();Fr();Ac=(e,t=!1,r,o,i=4)=>{let u=I=>{switch(I){case 1:return"return w[getIndexFromCoords4D(coord, vec4<i32>(uniforms.w_shape))];";case 4:return`\n            let coord1 = vec4<i32>(coordX, coordY, col + 1, rowInner);\n            let coord2 = vec4<i32>(coordX, coordY, col + 2, rowInner);\n            let coord3 = vec4<i32>(coordX, coordY, col + 3, rowInner);\n            let v0 = w[getIndexFromCoords4D(coord, vec4<i32>(uniforms.w_shape))];\n            let v1 = w[getIndexFromCoords4D(coord1, vec4<i32>(uniforms.w_shape))];\n            let v2 = w[getIndexFromCoords4D(coord2, vec4<i32>(uniforms.w_shape))];\n            let v3 = w[getIndexFromCoords4D(coord3, vec4<i32>(uniforms.w_shape))];\n            return ${o}(v0, v1, v2, v3);\n            `;default:throw new Error(`innerElementSize ${I} is not supported.`)}},a=e?`\n      let coord = vec4<i32>(batch, iXR, iXC, xCh);\n      `:`\n      let coord = vec4<i32>(batch, xCh, iXR, iXC);\n      `,c=e?`\n    let coords = vec4<i32>(\n      batch,\n      row / outWidth,\n      row % outWidth,\n      col);\n    `:`\n    let coords = vec4<i32>(\n      batch,\n      row,\n      col / outWidth,\n      col % outWidth);\n    `,p=e?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",h=e?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",d=e?"row":"col",y=e?"col":"row",w=`\n      let inChannels = ${e?"i32(uniforms.x_shape[3])":"i32(uniforms.x_shape[1])"};\n      let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};\n      let outRow = ${d} / outWidth;\n      let outCol = ${d} % outWidth;\n\n      let WRow = ${y} / (uniforms.filter_dims[1] * inChannels);\n      let WCol = ${y} / inChannels % uniforms.filter_dims[1];\n      let xR = f32(outRow - uniforms.pads[0] + uniforms.dilations[0] * WRow) / f32(uniforms.strides[0]);\n      let xC = f32(outCol - uniforms.pads[1] + uniforms.dilations[1] * WCol) / f32(uniforms.strides[1]);\n      if (xR < 0.0 || xR >= f32(${p}) || fract(xR) > 0.0) {\n        return ${o}(0.0);\n      }\n      if (xC < 0.0 || xC >= f32(${h}) || fract(xC) > 0.0) {\n        return ${o}(0.0);\n      }\n      let iXR = i32(xR);\n      let iXC = i32(xC);\n      let xCh = ${y} % inChannels;\n      ${a}\n      return x[getIndexFromCoords4D(coord, vec4<i32>(uniforms.x_shape))/${i}];`,_=e?`\n      let col = colIn * ${i};\n      if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) {\n        ${w}\n      }\n      return ${o}(0.0);`:`\n      let col = colIn * ${i};\n      if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) {\n        ${w}\n      }\n      return ${o}(0.0);`,v=`\n      let col = colIn * ${i};\n      let inChannels = ${e?"i32(uniforms.x_shape[3])":"i32(uniforms.x_shape[1])"};\n      let coordX = uniforms.filter_dims[0] - 1 - row / (uniforms.filter_dims[1] * inChannels);\n      let coordY = uniforms.filter_dims[1] - 1 - (row / inChannels) % uniforms.filter_dims[1];\n      if (${e?"row < uniforms.dim_inner && col < uniforms.dim_b_outer":"row < uniforms.dim_inner && col < uniforms.dim_a_outer"}  && coordX >= 0 && coordY >= 0) {\n        let rowInner = row % inChannels;\n        let coord = vec4<i32>(coordX, coordY, col, rowInner);\n        ${u(i)}\n      }\n      return ${o}(0.0);\n      `,S=St(r,o);return`\n  fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${o} {\n    ${e?_:v}\n  }\n\n  fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${o} {\n    ${e?v:_}\n  }\n\n  fn mm_write(batch: i32, row : i32, colIn : i32, valueInput : ${o}) {\n    let col = colIn * ${i};\n    if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) {\n      var value = valueInput;\n      let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};\n      ${c}\n      ${Dn(t)}\n      ${S}\n      result[getIndexFromCoords4D(coords, vec4<i32>(uniforms.result_shape))/${i}] = value;\n    }\n  }`},ou=(e,t,r,o,i,u,a,c)=>{let p=t.format==="NHWC",h=p?e[0].dims[3]:e[0].dims[1],d=r[0],y=p?r[2]:r[3],w=p?r[1]:r[2],_=p?r[3]:r[1],v=p&&h%4===0&&h%3&&_%4===0,S=p?_:y*w,A=p?y*w:_,I=[8,8,1],x=o<=8?[4,1,1]:[4,4,1],E=[Math.ceil(S/I[0]/x[0]),Math.ceil(A/I[1]/x[1]),Math.ceil(d/I[2]/x[2])];Ve("verbose",()=>`[conv_backprop_mm_webgpu] dispatch = ${E}`);let P=v?4:1,O=Math.max(I[0]*P,I[1]),R=v?4:1,L=[t.kernelShape[p?1:2],t.kernelShape[p?2:3]],N=[L[0]+(t.dilations[0]<=1?0:(L[0]-1)*(t.dilations[0]-1)),L[1]+(t.dilations[1]<=1?0:(L[1]-1)*(t.dilations[1]-1))],K=[N[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),N[1]-1-Math.floor((t.pads[1]+t.pads[3])/2)],Q=[{type:6,data:o},{type:6,data:i},{type:6,data:u},{type:6,data:t.strides},{type:6,data:t.dilations},{type:6,data:L},{type:6,data:K}];xt(t,Q),Q.push(...Z(e[0].dims,e[1].dims));let he=["rank","rank"];a&&(Q.push(...Z(e[2].dims)),he.push("rank")),Q.push(...Z(r));let W=se=>{let Ce=U("x",e[0].dataType,e[0].dims.length,R),We=U("w",e[1].dataType,e[1].dims.length,1),ee=j("result",e[0].dataType,r.length,R),ae=[Ce,We],Ae="";if(a){let ue=U("bias",e[2].dataType,e[2].dims.length,R);ae.push(ue),Ae+=`\n          fn getBiasByOutputCoords(coords : vec4<i32>) -> ${ue.type.value} {\n            return bias[coords.${p?"w":"y"}${v?"/ 4":""}];\n          }`}let me=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"},{name:"strides",type:"i32",length:2},{name:"dilations",type:"i32",length:2},{name:"filter_dims",type:"i32",length:L.length},{name:"pads",type:"i32",length:K.length}];Ct(t,me);let ie=De(e[0].dataType,1);if(ie!=="f16"&&ie!=="f32")throw new Error(`elemType ${ie} is not supported.`);return`\n        ${Mn("uniforms.result_strides")}\n        ${se.registerUniforms(me).declareVariables(...ae,ee)};\n        ${Ae}\n        ${Ac(p,a,t,Ce.type.value,P)}\n        ${v?Hr(x,I,ie,void 0,!p,O):Lr(x,I,ie,void 0,!p,O,!1,void 0,c)}`};return{name:"Conv2DTransposeMatMul",shaderCache:{hint:`${t.cacheKey};${x};${I};${v}`,inputDependencies:he},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:E[0],y:E[1],z:E[2]},programUniforms:Q}),getShaderSource:W}}});var Ic,Lo,au=Y(()=>{"use strict";ye();Lt();Se();_e();Ic=(e,t,r,o,i,u=!1,a,c,p=!1)=>{let h=p?1:2,d=p?2:3,y=p?3:1,w=u?2:1,_=`\n  fn setOutputAtIndex(flatIndex : u32, value : ${u?`vec4<${a}>`:a}) {\n    result[flatIndex] = ${u?`vec4<${a}>`:a}(value);\n  }`;o&&(_+=`\n    fn getBiasByOutputCoords(coords : vec4<u32>) -> ${u?`vec4<${a}>`:a} {\n      return bias[coords.${p?"w":"y"}${u?"/ 4":""}];\n    }`);let v=u?4:1,S=U("W",t[1].dataType,t[1].dims.length,v),A=U("Dy",t[0].dataType,t[0].dims.length,v),I=[A,S];o&&I.push(U("bias",t[2].dataType,[r[y]].length,v));let x=j("result",t[0].dataType,r.length,v),E=`{\n        let batch: u32 = ${i?"global_id.z":"workgroup_id.z"} / uniforms.result_shape[1];\n        let r = ${i?"global_id.z":"workgroup_id.z"} % uniforms.result_shape[1];\n        let c = ${i?"global_id.y":"workgroup_id.y"} * ${w};\n        let d1: u32 = ${i?"global_id.x":"workgroup_id.x"} * 4;\n\n        let dyCorner = vec2<i32>(i32(r), i32(c)) - vec2<i32>(uniforms.pads);\n\n        // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).\n        // ? = to be determined. : = across all values in that axis.\n        var dotProd: array<vec4<${a}>, ${w}>;\n        for (var i = 0; i < ${w}; i++) {\n          dotProd[i] = vec4<${a}>(0.0);\n        }\n        for (var wR: u32 = 0; wR < uniforms.filter_dims[0]; wR = wR + 1) {\n          var dyR = (${a}(dyCorner.x) + ${a}(wR)) / ${a}(uniforms.strides.x);\n          let wRPerm = uniforms.filter_dims[0] - 1 - wR;\n          if (dyR < 0.0 || dyR >= ${a}(uniforms.Dy_shape[1]) ||\n              fract(dyR) > 0.0 || wRPerm < 0) {\n            continue;\n          }\n          let idyR: u32 = u32(dyR);\n\n          for (var wC: u32 = 0; wC < uniforms.filter_dims[1]; wC = wC + 1) {\n            let dyC = (${a}(dyCorner.y) + ${a}(wC)) / ${a}(uniforms.strides.y);\n            let dyC2 = (${a}(dyCorner.y) + 1.0 + ${a}(wC)) / ${a}(uniforms.strides.y);\n            let wCPerm = uniforms.filter_dims[1] - 1 - wC;\n            if (wCPerm < 0) {\n              continue;\n            }\n            var bDyCVal = true;\n            var bDyCVal2 = true;\n            if (dyC < 0.0 || dyC >= ${a}(uniforms.Dy_shape[2]) ||\n                fract(dyC) > 0.0) {\n              bDyCVal = false;\n            }\n            if (dyC2 < 0.0 || dyC2 >= ${a}(uniforms.Dy_shape[2]) ||\n                fract(dyC2) > 0.0) {\n              bDyCVal2 = false;\n            }\n\n            let idyC: u32 = u32(dyC);\n            let idyC2: u32 = u32(dyC2);\n            if (bDyCVal && bDyCVal2) {\n              let d2Length = uniforms.Dy_shape[3];\n              for (var d2 :u32 = 0; d2 < d2Length; d2 = d2 + 4) {\n                let wValue0 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};\n                let wValue1 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};\n                let wValue2 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};\n                let wValue3 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};\n\n                var xValue = ${A.get("batch","idyR","idyC","d2")};\n                let tmpval = vec4<${a}>(dot(xValue, wValue0),\n                                      dot(xValue, wValue1),\n                                      dot(xValue, wValue2),\n                                      dot(xValue, wValue3));\n                dotProd[0] = dotProd[0] + tmpval;\n\n                xValue =  ${A.get("batch","idyR","idyC2","d2")};\n\n                dotProd[1] = dotProd[1] + vec4<${a}>(dot(xValue, wValue0),\n                                                    dot(xValue, wValue1),\n                                                    dot(xValue, wValue2),\n                                                    dot(xValue, wValue3));\n              }\n            } else if (bDyCVal) {\n              let d2Length = uniforms.Dy_shape[${y}];\n              for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) {\n                let wValue0 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};\n                let wValue1 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};\n                let wValue2 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};\n                let wValue3 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};\n\n                var xValue = ${A.get("batch","idyR","idyC","d2")};\n                let tmpval = vec4<${a}>(dot(xValue, wValue0),\n                                      dot(xValue, wValue1),\n                                      dot(xValue, wValue2),\n                                      dot(xValue, wValue3));\n                dotProd[0] = dotProd[0] + tmpval;\n              }\n            } else if (bDyCVal2) {\n              let d2Length = uniforms.Dy_shape[3];\n              for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) {\n                let wValue0 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};\n                let wValue1 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};\n                let wValue2 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};\n                let wValue3 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};\n\n                var xValue = ${A.get("batch","idyR","idyC2","d2")};\n                let tmpval = vec4<${a}>(dot(xValue, wValue0),\n                                      dot(xValue, wValue1),\n                                      dot(xValue, wValue2),\n                                      dot(xValue, wValue3));\n                dotProd[1] = dotProd[1] + tmpval;\n              }\n            }\n          }\n        }\n\n        for (var i: u32 = 0; i < ${w}; i = i + 1) {\n          let value = dotProd[i] + ${o?"bias[c+i]":`vec4<${a}>(0.0)`};\n          ${x.set("batch","r","c + i","d1","value")};\n        }\n      }`,P=`\n          let outputIndices = ${x.offsetToIndices("global_idx")};\n          let batch = ${x.indicesGet("outputIndices",0)};\n          let d1 = ${x.indicesGet("outputIndices",y)};\n          let r = ${x.indicesGet("outputIndices",h)};\n          let c = ${x.indicesGet("outputIndices",d)};\n          let dyCorner = vec2<i32>(i32(r), i32(c)) - uniforms.pads;\n          let dyRCorner = dyCorner.x;\n          let dyCCorner = dyCorner.y;\n          let groupId = d1 / uniforms.output_channels_per_group;\n          let wOutChannel = d1 - groupId * uniforms.output_channels_per_group;\n          // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).\n          // ? = to be determined. : = across all values in that axis.\n          var dotProd = ${a}(0.0);\n          for (var wR: u32 = 0; wR < uniforms.effective_filter_dims.x; wR = wR + 1) {\n            if (wR % uniforms.dilations.x != 0) {\n              continue;\n            }\n            let dyR = (${a}(dyRCorner) + ${a}(wR)) / ${a}(uniforms.strides[0]);\n            let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x;\n            if (dyR < 0.0 || dyR >= ${a}(uniforms.Dy_shape[${h}]) || fract(dyR) > 0.0 ||\n                wRPerm < 0) {\n              continue;\n            }\n            let idyR: u32 = u32(dyR);\n\n            for (var wC: u32 = 0; wC < uniforms.effective_filter_dims.y; wC = wC + 1) {\n              if (wC % uniforms.dilations.y != 0) {\n                continue;\n              }\n              let dyC = (${a}(dyCCorner) + ${a}(wC)) / ${a}(uniforms.strides.y);\n              let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y;\n              if (dyC < 0.0 || dyC >= ${a}(uniforms.Dy_shape[${d}]) ||\n                  fract(dyC) > 0.0 || wCPerm < 0) {\n                continue;\n              }\n              let idyC: u32 = u32(dyC);\n              var inputChannel = groupId * uniforms.input_channels_per_group;\n              for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group; d2 = d2 + 1) {\n                let xValue = ${p?A.get("batch","idyR","idyC","inputChannel"):A.get("batch","inputChannel","idyR","idyC")};\n                let wValue = ${S.get("inputChannel","wOutChannel","u32(wRPerm)","u32(wCPerm)")};\n                dotProd = dotProd + xValue * wValue;\n                inputChannel = inputChannel + 1;\n              }\n            }\n          }\n          let value = dotProd + ${o?"bias[d1]":`${a}(0.0)`};\n          ${x.setByOffset("global_idx","value")};\n        `;return`\n  ${e.registerUniforms(c).declareVariables(...I,x)}\n  ${_}\n\n    ${e.mainStart()}\n    ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")};\n  ${u?E:P}}`},Lo=(e,t,r)=>{let o=e.length>2,i=t.outputShape,u=M.size(i),a=[Math.ceil(u/64),1,1];Ve("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${a}`);let c=t.format==="NHWC",p=["rank","rank"],h=[t.strides[0],t.strides[1]],d=[t.kernelShape[c?1:2],t.kernelShape[c?2:3]],y=[t.dilations[0],t.dilations[1]],w=[d[0]+(t.dilations[0]<=1?0:(t.kernelShape[c?1:2]-1)*(t.dilations[0]-1)),d[1]+(t.dilations[1]<=1?0:(t.kernelShape[c?2:3]-1)*(t.dilations[1]-1))],_=[w[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),w[1]-1-Math.floor(t.pads[1]+t.pads[3])/2],v=!1,S=t.group,A=e[1].dims,I=A[0]/S,x=A[1],E=[{type:12,data:u},{type:12,data:h},{type:12,data:d},{type:12,data:y},{type:12,data:w},{type:6,data:_},{type:12,data:I},{type:12,data:x},...Z(e[0].dims,e[1].dims)];o&&(E.push(...Z(e[2].dims)),p.push("rank")),E.push(...Z(i));let P=a[1]===1&&a[2]===1,O=R=>{let L=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:h.length},{name:"filter_dims",type:"u32",length:d.length},{name:"dilations",type:"u32",length:d.length},{name:"effective_filter_dims",type:"u32",length:w.length},{name:"pads",type:"i32",length:_.length},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],N=De(e[0].dataType);return`${Ic(R,e,i,o,P,v,N,L,c)}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${t.cacheKey};`,inputDependencies:p},getRunData:()=>({dispatchGroup:{x:a[0],y:a[1],z:a[2]},outputs:[{dims:r?r(i):i,dataType:e[0].dataType}],programUniforms:E}),getShaderSource:O}}});var Tc,Ec,Pc,su,uu,kc,Oc,Rc,Bc,du,lu=Y(()=>{"use strict";iu();au();Ft();Sr();Tc=(e,t,r,o,i,u)=>(e-1)*t+r+(o-1)*i+1-u,Ec=(e,t,r,o,i)=>{let u=Math.floor(e/2);t==="SAME_UPPER"?(r[o]=u,r[i]=e-u):t==="SAME_LOWER"&&(r[o]=e-u,r[i]=u)},Pc=(e,t,r,o,i,u,a,c,p,h)=>{let d=e.length-2,y=h.length===0;if(p.length===0)for(let v=0;v<d;++v)p.push(0);let w=e[0],_=t[c?3:1]*i;for(let v=0,S=e.length-d-(c?1:0);v<d;++v,++S){let A=e[S],I=y?A*a[v]:h[v],x=Tc(A,a[v],u[v],t[S],r[v],I);Ec(x,o,u,v,v+d),y&&h.push(a[v]*(A-1)+p[v]+(t[S]-1)*r[v]+1-u[v]-u[v+d])}h.splice(0,0,w),h.splice(c?3:1,0,_)},su=(e,t)=>{let r=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((y,w)=>y*w,1)===0){r.length=0;for(let y=2;y<t[1].dims.length;++y)r.push(t[1].dims[y])}let o=e.format==="NHWC";r.splice(0,0,t[1].dims[0]),r.splice(o?3:1,0,t[1].dims[1]);let i=e.pads.slice(),u=e.outputShape.slice(),a=e.outputPadding.slice(),c=t[0].dims,p=e.dilations.slice();if(p.reduce((y,w)=>y+w,0)===0){let y=t[0].dims.length-2;p=new Array(y).fill(1)}let h=e.strides.slice();if(h.reduce((y,w)=>y+w,0)===0){let y=t[0].dims.length-2;h=new Array(y).fill(1)}Pc(c,r,p,e.autoPad,e.group,i,h,o,a,u);let d=Object.assign({},e);return Object.assign(d,{kernelShape:r,pads:i,outputPadding:a,outputShape:u,dilations:p,strides:h}),d},uu=e=>{let t=Bn(e),r=e.format,o=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],i=e.dilations,u=e.group,a=e.kernelShape,c=e.pads,p=e.strides,h=e.wIsConst(),d=e.outputPadding,y=e.outputShape;return{autoPad:o,format:r,dilations:i,group:u,kernelShape:a,outputPadding:d,outputShape:y,pads:c,strides:p,wIsConst:h,...t,cacheKey:`${e.format};${t.activation};`}},kc=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length!==4&&e[0].dims.length!==3)throw new Error("currently only support 2-dimensional conv");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let r=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],o=e[1].dims[0];if(r!==o)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let i=e[1].dims[1]*t.group;if(e.length===3&&(e[2].dims.length!==1||e[2].dims[0]!==i))throw new Error("invalid bias");let u=e[0].dims.length-2;if(t.dilations.reduce((d,y)=>d+y,0)>0&&t.dilations.length!==u)throw new Error(`dilations should be ${u}D`);if(t.strides.reduce((d,y)=>d+y,0)>0&&t.strides.length!==u)throw new Error(`strides should be ${u}D`);if(t.pads.reduce((d,y)=>d+y,0)>0&&t.pads.length!==u*2)throw new Error(`pads should be ${u*2}D`);if(t.outputPadding.length!==u&&t.outputPadding.length!==0)throw new Error(`output_padding should be ${u}D`);if(t.kernelShape.reduce((d,y)=>d+y,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape");if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw new Error("invalid output shape")},Oc=[2,3,1,0],Rc=(e,t,r)=>{let o=su(r,t),i=r.format==="NHWC",u=o.outputShape,a=u[i?3:1],c=t[0].dims[i?3:1];if(o.group!==1||a===1&&c===1){e.compute(Lo(t,o));return}let p=u[i?1:2],h=u[i?2:3],d=t[1].dims[2],y=t[1].dims[3],w=i?p*h:a,_=i?a:p*h,v=d*y*c,S=!0,A=e.kernelCustomData.wT??e.compute(yt(t[1],Oc),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=A);let I=[t[0],A],x=t.length===3;x&&(!i&&t[2].dims.length===1?I.push(t[2].reshape([t[2].dims[0],1,1])):I.push(t[2])),e.compute(ou(I,o,u,w,_,v,x,S),{inputs:I})},Bc=(e,t)=>{let r=t.format==="NHWC",o=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&o.push(e.inputs[2]);let i=t.kernelShape;(i.length===0||i[0]===0)&&(i=[e.inputs[1].dims[2]]);let u=t.dilations;(u.length===0||u[0]===0)&&(u=[1]);let a=t.strides;(a.length===0||a[0]===0)&&(a=[1]);let c=t.pads;c.length===0&&(c=[0,0]),c=[0,c[0],0,c[1]],a=[1].concat(a),u=[1].concat(u),i=[1].concat(i);let p=su({...t,pads:c,strides:a,dilations:u,kernelShape:i},o);e.compute(Lo(o,p,h=>r?[h[0],h[2],h[3]]:[h[0],h[1],h[3]]))},du=(e,t)=>{kc(e.inputs,t),e.inputs[0].dims.length===3?Bc(e,t):Rc(e,e.inputs,t)}});var Dc,cu,pu,mu=Y(()=>{"use strict";ye();Se();Ze();_e();Dc=(e,t,r,o)=>{let i=M.size(t),u=t.length,a=U("input",e,u),c=j("output",e,u),p=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),h=M.normalizeAxis(p,u),d=y=>{let w=` i32(${a.indicesGet("inputIndices","uniforms.axis")}) `,_=fe("uniforms.input_shape","uniforms.axis",u),v=o.reverse?w+(o.exclusive?" + 1":""):"0",S=o.reverse?_:w+(o.exclusive?"":" + 1");return`\n                ${y.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(a,c)}\n                ${y.mainStart()}\n                  ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n                  var inputIndices = ${c.offsetToIndices("global_idx")};\n                  var sum = ${c.type.value}(0);\n                  let first : i32 = ${v};\n                  let last : i32 = ${S};\n                  for (var i : i32 = first; i < last; i++) {\n                    ${a.indicesSet("inputIndices","uniforms.axis","u32(i)")};\n                    sum = sum + ${a.getByIndices("inputIndices")};\n                  }\n                  ${c.setByOffset("global_idx","sum")};\n                }`};return{name:"CumSum",shaderCache:{hint:o.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:t,dataType:e}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:[{type:12,data:i},{type:12,data:h},...Z(t,t)]}),getShaderSource:d}},cu=(e,t)=>{let r=e.inputs[0].dims,o=e.inputs[0].dataType,i=e.inputs[1];e.compute(Dc(o,r,i,t),{inputs:[0]})},pu=e=>{let t=e.exclusive===1,r=e.reverse===1;return ve({exclusive:t,reverse:r})}});var zc,Mc,Uc,fu,hu,gu=Y(()=>{"use strict";ye();Se();Ze();_e();zc=e=>{if(!e||e.length!==1)throw new Error("DepthToSpace requires 1 input.");if(e[0].dims.length!==4)throw new Error("DepthToSpace requires 4D input.")},Mc=(e,t,r,o)=>{let i=[];i.push(`fn perm(i: ${o.type.indices}) -> ${r.type.indices} {\n    var a: ${r.type.indices};`);for(let u=0;u<t;++u)i.push(r.indicesSet("a",e[u],`i[${u}]`));return i.push("return a;}"),i.join(`\n`)},Uc=(e,t)=>{let r,o,i,u,a,c,p=t.format==="NHWC",h=t.blocksize,d=t.mode==="DCR";p?([r,o,i,u]=e.dims,a=d?[r,o,i,h,h,u/h**2]:[r,o,i,u/h**2,h,h],c=d?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([r,o,i,u]=[e.dims[0],e.dims[2],e.dims[3],e.dims[1]],a=d?[r,h,h,u/h**2,o,i]:[r,u/h**2,h,h,o,i],c=d?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let y=e.reshape(a),w=y.dims.length,_=e.dataType,v=U("a",_,w),S=j("output",_,w),A=I=>`\n  ${I.registerUniform("output_size","u32").declareVariables(v,S)}\n\n  ${Mc(c,w,v,S)}\n\n  ${I.mainStart()}\n    ${I.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n    let indices = ${S.offsetToIndices("global_idx")};\n    let aIndices = perm(indices);\n\n    ${S.setByOffset("global_idx",v.getByIndices("aIndices"))}\n  }`;return{name:"DepthToSpace",shaderCache:{hint:`${e.dims};${t.blocksize};${t.mode}`,inputDependencies:["rank"]},getRunData:I=>{let x=p?[r,o*h,i*h,u/h**2]:[r,u/h**2,o*h,i*h],E=M.size(x),P=y.dims,O=M.sortBasedOnPerm(P,c);return{outputs:[{dims:x,dataType:I[0].dataType}],dispatchGroup:{x:Math.ceil(E/64)},programUniforms:[{type:12,data:E},...Z(P,O)]}},getShaderSource:A}},fu=(e,t)=>{zc(e.inputs),e.compute(Uc(e.inputs[0],t))},hu=e=>ve({blocksize:e.blocksize,mode:e.mode,format:e.format})});var Fo,Wn,yu,Vc,Wc,qo,jo,bu,Nc,wu,vu,$u=Y(()=>{"use strict";ye();Se();Ze();_e();Fo="[a-zA-Z]|\\\\.\\\\.\\\\.",Wn="("+Fo+")+",yu="^"+Wn+"$",Vc="("+Wn+",)*"+Wn,Wc="^"+Vc+"$",qo=class{constructor(t=-1){this.symbolToIndices=new Map,this.inputIndex=t}addSymbol(t,r){let o=this.symbolToIndices.get(t);o===void 0?o=[r]:o.push(r),this.symbolToIndices.set(t,o)}},jo=class{constructor(t,r){this.equation=r;this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[o,i]=r.includes("->")?r.split("->",2):[r,""];if(!o.match(RegExp(Wc)))throw new Error("Invalid LHS term");if(o.split(",").forEach((c,p)=>{let h=t[p].dims.slice();if(!c.match(RegExp(yu)))throw new Error("Invalid LHS term");let d=this.processTerm(c,!0,h,p);this.lhs.push(d)}),i==="")i+=[...this.symbolToInfo.entries()].filter(([c,p])=>p.count===1||c==="...").map(([c])=>c).join("");else if(!i.match(RegExp(Wn)))throw new Error("Invalid RHS");i.match(RegExp(Fo,"g"))?.forEach(c=>{if(c==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let p=this.symbolToInfo.get(c);if(p===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(p.dimValue)}}),this.rhs=this.processTerm(i,!1,this.outputDims)}addSymbol(t,r,o){let i=this.symbolToInfo.get(t);if(i!==void 0){if(i.dimValue!==r&&i.count!==1)throw new Error("Dimension mismatch");i.count++,i.inputIndices.push(o)}else i={count:1,dimValue:r,inputIndices:[o]};this.symbolToInfo.set(t,i)}processTerm(t,r,o,i=-1){let u=o.length,a=!1,c=[],p=0;if(!t.match(RegExp(yu))&&!r&&t!=="")throw new Error("Invalid LHS term");let h=t.match(RegExp(Fo,"g")),d=new qo(i);return h?.forEach((y,w)=>{if(y==="..."){if(a)throw new Error("Only one ellipsis is allowed per input term");a=!0;let _=u-h.length+1;if(_<0)throw new Error("Ellipsis out of bounds");if(c=o.slice(p,p+_),this.hasEllipsis){if(this.ellipsisDims.length!==c.length||this.ellipsisDims.toString()!==c.toString())throw new Error("Ellipsis dimensions mismatch")}else if(r)this.hasEllipsis=!0,this.ellipsisDims=c;else throw new Error("Ellipsis must be specified in the LHS");for(let v=0;v<c.length;v++){let S=String.fromCharCode("0".charCodeAt(0)+v);d.addSymbol(S,w+v),this.addSymbol(S,o[p++],i)}}else d.addSymbol(y,w+(this.hasEllipsis?this.ellipsisDims.length-1:0)),this.addSymbol(y,o[p++],i)}),d}},bu=e=>e+"_max",Nc=(e,t,r,o)=>{let u=e.map(d=>d.length).map((d,y)=>U(`input${y}`,t,d)),a=M.size(o),c=j("output",t,o.length),p=[...r.symbolToInfo.keys()].filter(d=>!r.rhs.symbolToIndices.has(d)),h=d=>{let y=[],w="var prod = 1.0;",_="var sum = 0.0;",v="sum += prod;",S=[],A=[],I=[],x=[],E=r.symbolToInfo.size===r.rhs.symbolToIndices.size;r.symbolToInfo.forEach((O,R)=>{if(r.rhs.symbolToIndices.has(R)){let L=r.rhs.symbolToIndices.get(R)?.[0];L!==void 0&&r.lhs.forEach((N,K)=>{if(O.inputIndices.includes(K)){let Q=N.symbolToIndices.get(R);if(Q===void 0)throw new Error("Invalid symbol error");Q.forEach(he=>{y.push(`${u[K].indicesSet(`input${K}Indices`,he,c.indicesGet("outputIndices",L))}`)})}})}else r.lhs.forEach((L,N)=>{if(O.inputIndices.includes(N)){let K=L.symbolToIndices.get(R);if(K===void 0)throw new Error("Invalid symbol error");K.forEach(Q=>{S.push(`${u[N].indicesSet(`input${N}Indices`,Q,`${R}`)}`)}),x.push(`prod *= ${u[N].getByIndices(`input${N}Indices`)};`)}}),A.push(`for(var ${R}: u32 = 0; ${R} < uniforms.${bu(R)}; ${R}++) {`),I.push("}")});let P=E?[...y,`let sum = ${u.map((O,R)=>O.getByIndices(`input${R}Indices`)).join(" * ")};`]:[...y,_,...A,...S,w,...x,v,...I];return`\n            ${d.registerUniforms(p.map(O=>({name:`${bu(O)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...u,c)}\n\n            ${d.mainStart()}\n            ${d.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n            var outputIndices = ${c.offsetToIndices("global_idx")};\n            ${u.map((O,R)=>`var input${R}Indices: ${u[R].type.indices};`).join(`\n`)}\n            ${P.join(`\n`)};\n            ${c.setByOffset("global_idx","sum")};\n          }`};return{name:"Einsum",shaderCache:{hint:r.equation,inputDependencies:e.map(()=>"rank")},getRunData:()=>{let d=p.filter(w=>r.symbolToInfo.has(w)).map(w=>({type:12,data:r.symbolToInfo.get(w)?.dimValue||0}));d.push({type:12,data:a});let y=e.map((w,_)=>[...Z(w)]).reduce((w,_)=>w.concat(_),d);return y.push(...Z(o)),{outputs:[{dims:o,dataType:t}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:y}},getShaderSource:h}},wu=(e,t)=>{let r=new jo(e.inputs,t.equation),o=r.outputDims,i=e.inputs.map((u,a)=>u.dims);e.compute(Nc(i,e.inputs[0].dataType,r,o))},vu=e=>{let t=e.equation.replace(/\\s+/g,"");return ve({equation:t})}});var Gc,_u,Hc,Lc,Su,xu=Y(()=>{"use strict";ye();Se();_e();Gc=e=>{if(!e||e.length!==2)throw new Error("Expand requires 2 input.");let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),o=r.length<t.length?0:r.length-t.length,i=t.length<r.length?0:t.length-r.length;for(;o<r.length&&i<t.length;++o,++i)if(r[o]!==t[i]&&r[o]!==1&&t[i]!==1)throw new Error("Expand requires shape to be broadcastable to input")},_u=(e,t)=>{let r=e.length-t.length,o=[];for(let i=0;i<r;++i)o.push(e[i]);for(let i=0;i<t.length;++i)o.push(t[i]===1?e[i+r]:t[i]);return o},Hc=(e,t)=>e.length>t.length?_u(e,t):_u(t,e),Lc=e=>{let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),o=Hc(t,r),i=e[0].dataType,u=i===9?4:1,a=Math.ceil(M.size(o)/u),c=h=>{let d=U("input",i,t.length,u),y=j("output",i,o.length,u),w;if(i===9){let _=(v,S,A="")=>`\n          let outputIndices${S} = ${y.offsetToIndices(`outputOffset + ${S}u`)};\n          let offset${S} = ${d.broadcastedIndicesToOffset(`outputIndices${S}`,y)};\n          let index${S} = offset${S} / 4u;\n          let component${S} = offset${S} % 4u;\n          ${v}[${S}] = ${A}(${d.getByOffset(`index${S}`)}[component${S}]);\n        `;w=`\n        let outputOffset = global_idx * ${u};\n        var data = vec4<u32>(0);\n        ${_("data",0,"u32")}\n        ${_("data",1,"u32")}\n        ${_("data",2,"u32")}\n        ${_("data",3,"u32")}\n        ${y.setByOffset("global_idx","data")}\n      }`}else w=`\n        let outputIndices = ${y.offsetToIndices("global_idx")};\n        let inputOffset = ${d.broadcastedIndicesToOffset("outputIndices",y)};\n        ${y.setByOffset("global_idx",d.getByOffset("inputOffset"))}\n      }`;return`\n    ${h.registerUniform("vec_size","u32").declareVariables(d,y)}\n    ${h.mainStart()}\n    ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}\n    ${w}`},p=[{type:12,data:a},...Z(t,o)];return{name:"Expand",shaderCache:{hint:`${o.length}`,inputDependencies:["rank"]},getShaderSource:c,getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:p})}},Su=e=>{Gc(e.inputs),e.compute(Lc(e.inputs),{inputs:[0]})}});var Fc,Cu,Au=Y(()=>{"use strict";ye();Se();_e();Rn();Fc=e=>{let t=e[0].dataType,r=M.size(e[0].dims),o=M.size(e[1].dims),i=o%4===0,u=a=>{let c=U("x",t,[1],4),p=U("bias",t,[1],4),h=j("y",t,[1],4),d=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],y=_=>`\n      let bias${_}_offset: u32 = (global_idx * 4 + ${_}) % uniforms.bias_size;\n      let bias${_} = ${p.getByOffset(`bias${_}_offset / 4`)}[bias${_}_offset % 4];`,w=i?`\n      let bias = ${p.getByOffset("global_idx % (uniforms.bias_size / 4)")};`:`${y(0)}${y(1)}${y(2)}${y(3)}\n      let bias = ${c.type.value}(bias0, bias1, bias2, bias3);`;return`${a.registerUniforms(d).declareVariables(c,p,h)}\n\n    ${Bo(et(t))}\n\n    ${a.mainStart(or)}\n      ${a.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_vec_size")}\n\n      let x = ${c.getByOffset("global_idx")};\n      ${w}\n      let x_in = x + bias;\n      ${h.setByOffset("global_idx",Do("x_in"))}\n    }`};return{name:"FastGeluWithBias",shaderCache:{hint:`${i}`,inputDependencies:["type","type"]},getShaderSource:u,getRunData:a=>({outputs:[{dims:a[0].dims,dataType:a[0].dataType}],programUniforms:[{type:12,data:Math.ceil(r/4)},{type:12,data:o}],dispatchGroup:{x:Math.ceil(r/or/4)}})}},Cu=e=>{e.inputs.length<2||M.size(e.inputs[1].dims)===0?Bs(e):e.compute(Fc(e.inputs))}});var qc,jc,Iu,Tu,Eu=Y(()=>{"use strict";ye();Se();Ze();_e();qc=e=>{if(!e||e.length!==2)throw new Error("Gather requires 2 inputs.")},jc=(e,t)=>{let r=e[0].dims,o=e[1].dims,i=r.length,u=M.normalizeAxis(t.axis,i),a=r.slice(0);a.splice(u,1,...o);let c=r[u],p=e[0].dataType===9?4:1,h=Math.ceil(M.size(a)/p),d=[{type:12,data:h},{type:6,data:c},{type:12,data:u},...Z(e[0].dims,e[1].dims,a)],y=w=>{let _=U("data",e[0].dataType,e[0].dims.length,p),v=U("inputIndices",e[1].dataType,e[1].dims.length),S=j("output",e[0].dataType,a.length,p),A=x=>{let E=o.length,P=`var indicesIndices${x}  = ${v.type.indices}(0);`;for(let O=0;O<E;O++)P+=`${E>1?`indicesIndices${x}[${O}]`:`indicesIndices${x}`} = ${a.length>1?`outputIndices${x}[uniforms.axis + ${O}]`:`outputIndices${x}`};`;P+=`\n          var idx${x} = ${v.getByIndices(`indicesIndices${x}`)};\n          if (idx${x} < 0) {\n            idx${x} = idx${x} + uniforms.axisDimLimit;\n          }\n          var dataIndices${x} : ${_.type.indices};\n        `;for(let O=0,R=0;O<i;O++)O===u?(P+=`${i>1?`dataIndices${x}[${O}]`:`dataIndices${x}`} = u32(idx${x});`,R+=E):(P+=`${i>1?`dataIndices${x}[${O}]`:`dataIndices${x}`} = ${a.length>1?`outputIndices${x}[${R}]`:`outputIndices${x}`};`,R++);return P},I;if(e[0].dataType===9){let x=(E,P,O="")=>`\n          let outputIndices${P} = ${S.offsetToIndices(`outputOffset + ${P}u`)};\n          ${A(P)};\n          let offset${P} = ${_.indicesToOffset(`dataIndices${P}`)};\n          let index${P} = offset${P} / 4u;\n          let component${P} = offset${P} % 4u;\n          ${E}[${P}] = ${O}(${_.getByOffset(`index${P}`)}[component${P}]);\n        `;I=`\n        let outputOffset = global_idx * ${p};\n        var value = vec4<u32>(0);\n        ${x("value",0,"u32")}\n        ${x("value",1,"u32")}\n        ${x("value",2,"u32")}\n        ${x("value",3,"u32")}\n        ${S.setByOffset("global_idx","value")}\n      `}else I=`\n      let outputIndices = ${S.offsetToIndices("global_idx")};\n      ${A("")};\n      let value = ${_.getByIndices("dataIndices")};\n      ${S.setByOffset("global_idx","value")};\n      `;return`\n      ${w.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(_,v,S)}\n      ${w.mainStart()}\n        ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n        ${I}\n      }`};return{name:"Gather",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:d}),getShaderSource:y}},Iu=e=>ve({axis:e.axis}),Tu=(e,t)=>{let r=e.inputs;qc(r),e.compute(jc(e.inputs,t))}});var Kc,Yc,Pu,ku,Ou=Y(()=>{"use strict";ye();Se();Ze();_e();Kc=e=>{if(!e||e.length!==2)throw new Error("GatherElements requires 2 inputs.");if(e[0].dims.length<1)throw new Error("GatherElements requires that the data input be rank >= 1.");if(e[0].dims.length!==e[1].dims.length)throw new Error(`GatherElements requires that the data input and\n                     indices input tensors be of same rank.`)},Yc=(e,t)=>{let r=e[0].dims,o=e[0].dataType,i=r.length,u=e[1].dims,a=e[1].dataType,c=M.normalizeAxis(t.axis,i),p=r[c],h=u.slice(0),d=M.size(h),y=U("input",o,i),w=U("indicesInput",a,u.length),_=j("output",o,h.length),v=[{type:12,data:d},{type:6,data:p},{type:12,data:c}];return v.push(...Z(r,u,h)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:h,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:v}),getShaderSource:I=>`\n      ${I.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(y,w,_)}\n      ${I.mainStart()}\n      ${I.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n\n      let outputIndices = ${_.offsetToIndices("global_idx")};\n\n      var idx = ${w.getByOffset("global_idx")};\n      if (idx < 0) {\n        idx = idx + uniforms.axisDimLimit;\n      }\n      var inputIndices = ${y.type.indices}(outputIndices);\n      ${y.indicesSet("inputIndices","uniforms.axis","u32(idx)")};\n      let value = ${y.getByIndices("inputIndices")};\n\n      ${_.setByOffset("global_idx","value")};\n  }`}},Pu=e=>ve({axis:e.axis}),ku=(e,t)=>{let r=e.inputs;Kc(r),e.compute(Yc(e.inputs,t))}});var Zc,Xc,Ru,Bu,Du=Y(()=>{"use strict";ye();Se();_e();Zc=e=>{if(!e)throw new Error("Input is missing");if(e.length<2||e.length>3)throw new Error("Invaid input number.");if(e.length===3&&e[2].dims.length>2)throw new Error("Invalid input shape of C");if(e[0].dataType!==e[1].dataType||e.length===3&&e[0].dataType!==e[2].dataType)throw new Error("Input types are mismatched")},Xc=(e,t)=>{let r=e[0].dims.slice(),o=e[1].dims.slice(),[i,u,a]=Sn.getShapeOfGemmResult(r,t.transA,o,t.transB,e.length===3?e[2].dims:void 0),c=[i,u];if(!c)throw new Error("Can\'t use gemm on the given tensors");let p=M.size(c),h=[{type:12,data:p},{type:12,data:i},{type:12,data:u},{type:12,data:a},{type:1,data:t.alpha},{type:1,data:t.beta}],d=["type","type"];e.length===3&&(h.push(...Z(e[2].dims)),d.push("rank")),h.push(...Z(c));let y=w=>{let _="";t.transA&&t.transB?_="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":t.transA&&!t.transB?_="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!t.transA&&t.transB?_="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!t.transA&&!t.transB&&(_="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let v=t.alpha===1?"":"value *= uniforms.alpha;",S=U("a",e[0].dataType,e[0].dims),A=U("b",e[1].dataType,e[1].dims),I=S.type.value,x=null,E=[S,A];e.length===3&&(x=U("c",e[2].dataType,e[2].dims.length),E.push(x));let P=j("output",e[0].dataType,c.length);E.push(P);let O=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}];return`\n  ${w.registerUniforms(O).declareVariables(...E)}\n\n  ${w.mainStart()}\n    ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n    let m = global_idx / uniforms.N;\n    let n = global_idx % uniforms.N;\n\n    var value = ${I}(0);\n    for (var k: u32 = 0u; k < uniforms.K; k++) {\n      ${_}\n    }\n\n    ${v}\n    ${(()=>x!=null?`let cOffset = ${x.broadcastedIndicesToOffset("vec2(m, n)",P)}; value += ${I}(uniforms.beta) * ${x.getByOffset("cOffset")};`:"")()}\n    output[global_idx] = value;\n  }`};return{name:"Gemm",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:c,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:h}),getShaderSource:y}},Ru=e=>{let t=e.transA,r=e.transB,o=e.alpha,i=e.beta;return{transA:t,transB:r,alpha:o,beta:i,cacheKey:`${e.transA};${e.transB};${e.alpha===1}`}},Bu=(e,t)=>{Zc(e.inputs),e.compute(Xc(e.inputs,t))}});var Qc,Jc,ep,zu,Mu=Y(()=>{"use strict";ye();Se();_e();Qc=(e,t)=>{let r=e[0].dims,o=r,i=2,u=M.sizeToDimension(r,i),a=M.sizeFromDimension(r,i),c=Me(a),p=a/c,h=[r[0],r[1],p],d=["rank","type","type"],y=[{type:12,data:a},{type:12,data:p}];y.push(...Z(h,h));let w=_=>{let v=U("x",e[0].dataType,h.length,c),S=U("scale",e[1].dataType,e[1].dims),A=U("bias",e[2].dataType,e[2].dims),I=j("output",e[0].dataType,h.length,c),x=[v,S,A,I],E=v.type.value,P=c===1?"f32":`vec${c}<f32>`,O=64,R=[{name:"normSize",type:"u32"},{name:"normPackedSize",type:"u32"}];return`\n  var<workgroup> meanShared : f32;\n  var<workgroup> squaredNormShared : f32;\n  var<workgroup> workgroupShared : array<${P}, ${O}>;\n  const workgroupSize = ${O}u;\n  ${_.registerUniforms(R).declareVariables(...x)}\n  ${_.mainStart(O)}\n    let norm = global_idx / workgroupSize;\n    let batch = norm / uniforms.x_shape[1];\n    let channel = norm % uniforms.x_shape[1];\n    let localIndex = local_id.x;\n\n    // initialize workgroup memory\n    var initial = ${P}(0);\n    for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) {\n      initial = initial + ${P}(${v.get("batch","channel","h")});\n    }\n    workgroupShared[localIndex] = initial;\n    workgroupBarrier();\n\n    // Calculate the mean of current channel data.\n    for (var currSize = workgroupSize >> 1;  currSize > 0; currSize = currSize >> 1) {\n      if (localIndex < currSize) {\n        workgroupShared[localIndex] = workgroupShared[localIndex] + workgroupShared[localIndex + currSize];\n      }\n      workgroupBarrier();\n    }\n    if (localIndex == 0) {\n      meanShared = ${_t("workgroupShared[0]",c)} / f32(uniforms.normSize);\n    }\n    workgroupBarrier();\n\n    // reinitialize workgroup memory.\n    initial = ${P}(0);\n    for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) {\n      let deviation =  ${P}(${v.get("batch","channel","h")}) - ${P}(meanShared);\n      initial = initial + deviation * deviation;\n    }\n    workgroupShared[localIndex] = initial;\n    workgroupBarrier();\n\n    // Calculate the sum of square of deviation of current channel data.\n    for (var currSize = workgroupSize >> 1;  currSize > 0; currSize = currSize >> 1) {\n      if (localIndex < currSize) {\n        workgroupShared[localIndex] = workgroupShared[localIndex] + workgroupShared[localIndex + currSize];\n      }\n      workgroupBarrier();\n    }\n    if (localIndex == 0) {\n      squaredNormShared = ${_t("workgroupShared[0]",c)};\n    }\n    workgroupBarrier();\n\n    let invStdDev = inverseSqrt(squaredNormShared / f32(uniforms.normSize) + f32(${t.epsilon}));\n    let channelScale = invStdDev * f32(${S.getByOffset("channel")});\n    let channelShift = f32(${A.getByOffset("channel")}) - meanShared * channelScale;\n    for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) {\n      let value = ${v.get("batch","channel","h")} * ${E}(${P}(channelScale)) + ${E}(${P}(channelShift));\n      ${I.set("batch","channel","h","value")};\n    }\n  }`};return{name:"InstanceNormalization",shaderCache:{hint:`${t.epsilon};${c}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:u},programUniforms:y}),getShaderSource:w}},Jc=(e,t,r,o,i,u,a,c)=>{let p=Me(a),h=64,d=p===1?"vec2f":`mat2x${p}f`,y=p===1?"f32":`vec${p}f`,w=(R,L)=>`${d}(${R}, ${L})`,_=i*a/p,v=Math.ceil(u/h),S=["type"],A=[{type:12,data:v},{type:12,data:u},{type:12,data:Math.floor(a/p)},{type:12,data:Math.floor(u*a/p)}],I=R=>{let L=U("input",t.dataType,t.dims,p);return`\n  ${R.declareVariables(L)}\n  @group(0) @binding(1) var<storage, read_write> output : array<${d}>;\n  struct Uniforms {wg_size:u32, H:u32, C:u32, image_size:u32};\n  @group(0) @binding(2) var<uniform> uniforms: Uniforms;\n\n  ${R.mainStart(h)}\n    let currentImageNumber = global_idx / ${h} / uniforms.C;\n    let currentChannelNumber = (global_idx / ${h}) % uniforms.C;\n    let wgOffset = local_id.x * uniforms.wg_size;\n    if (wgOffset >= uniforms.H) {\n        return;\n    }\n    let wgMax = min(wgOffset + uniforms.wg_size, uniforms.H);\n\n    let offset = currentImageNumber * uniforms.image_size + currentChannelNumber;\n    var sum = ${$t("f32",p)};\n    var squaredSum = ${$t("f32",p)};\n    for (var i: u32 = wgOffset; i < wgMax; i++) {\n        let value = ${y}(input[offset + i * uniforms.C]);\n        sum += value;\n        squaredSum += value * value;\n    }\n    output[global_idx] = ${w("sum","squaredSum")};\n  }`},x=e.compute({name:"InstanceNormComputeMean",shaderCache:{hint:`${p}`,inputDependencies:S},getRunData:()=>({outputs:[{dims:[i,a,h,2],dataType:1}],dispatchGroup:{x:i*a/p},programUniforms:A}),getShaderSource:I},{inputs:[t],outputs:[-1]})[0],E=[{type:12,data:_},{type:12,data:u},{type:12,data:Math.floor(a/p)},{type:12,data:Math.floor(h*a/p)}],P=["type","type","type"],O=R=>{let L=U("scale",r.dataType,r.dims,p),N=U("bias",o.dataType,o.dims,p);return`\n  @group(0) @binding(0) var<storage, read> input : array<${d}>;\n  @group(0) @binding(1) var<storage, read> scale : array<${L.type.storage}>;\n  @group(0) @binding(2) var<storage, read> bias : array<${N.type.storage}>;\n  @group(0) @binding(3) var<storage, read_write> output : array<${d}>;\n  struct Uniforms {units_of_work : u32, H: u32, C : u32, image_size : u32};\n  @group(0) @binding(4) var<uniform> uniforms: Uniforms;\n\n  ${R.mainStart()}\n    ${R.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.units_of_work")}\n    let currentImageNumber = global_idx / uniforms.C;\n    let currentChannelNumber = global_idx % uniforms.C;\n\n    let offset = currentImageNumber * uniforms.image_size;\n    var sum = ${$t("f32",p)};\n    var squaredSum = ${$t("f32",p)};\n    for (var i: u32 = 0; i < min(${h}, uniforms.H); i++) {\n        let value = input[offset + i + currentChannelNumber * ${h}];\n        sum += value[0];\n        squaredSum += value[1];\n    }\n    sum = sum / f32(uniforms.H);\n    squaredSum = squaredSum / f32(uniforms.H);\n    let invStdDev = inverseSqrt(squaredSum - sum * sum + f32(${c}));\n    let channelScale = invStdDev * ${y}(scale[currentChannelNumber]);\n    let channelShift = ${y}(bias[currentChannelNumber]) - sum * channelScale;\n\n    output[global_idx] = ${w("channelScale","channelShift")};\n  }`};return e.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${p};${c}`,inputDependencies:P},getRunData:()=>({outputs:[{dims:[i,a,2],dataType:1}],dispatchGroup:{x:Math.ceil(_/64)},programUniforms:E}),getShaderSource:O},{inputs:[x,r,o],outputs:[-1]})[0]},ep=(e,t,r)=>{let o=t[0].dims,i=o,u=o[0],a=o[o.length-1],c=M.sizeFromDimension(o,1)/a,p=Me(a),h=M.size(i)/p,d=[{type:12,data:c},{type:12,data:Math.floor(a/p)}],y=["type","type"],w=Jc(e,t[0],t[1],t[2],u,c,a,r.epsilon),_=v=>{let S=De(t[0].dataType),A=p===1?"vec2f":`mat2x${p}f`,I=p===1?S:`vec${p}<${S}>`,x=U("input",t[0].dataType,t[0].dims,p),E=j("output",t[0].dataType,i,p);return`\n  @group(0) @binding(0) var<storage, read> input : array<${x.type.storage}>;\n  @group(0) @binding(1) var<storage, read> scaleInput : array<${A}>;\n  @group(0) @binding(2) var<storage, read_write> output : array<${E.type.storage}>;\n  struct Uniforms {H: u32, C : u32};\n  @group(0) @binding(3) var<uniform> uniforms: Uniforms;\n\n  ${v.mainStart()}\n    let currentImageNumber = global_idx / (uniforms.C * uniforms.H);\n    let currentChannelNumber = global_idx % uniforms.C;\n\n    let scaleOffset = currentImageNumber * uniforms.C + currentChannelNumber;\n    let scale = scaleInput[scaleOffset];\n    output[global_idx] = fma(input[global_idx], ${I}(scale[0]), ${I}(scale[1]));\n  }`};e.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${p}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:i,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:d}),getShaderSource:_},{inputs:[t[0],w]})},zu=(e,t)=>{t.format==="NHWC"?ep(e,e.inputs,t):e.compute(Qc(e.inputs,t))}});var tp,rp,Uu,Vu=Y(()=>{"use strict";ye();Se();_e();tp=e=>{if(!e||e.length<2)throw new Error("layerNorm requires at least 2 inputs.")},rp=(e,t,r)=>{let o=t.simplified,i=e[0].dims,u=e[1],a=!o&&e[2],c=i,p=M.normalizeAxis(t.axis,i.length),h=M.sizeToDimension(i,p),d=M.sizeFromDimension(i,p),y=M.size(u.dims),w=a?M.size(a.dims):0;if(y!==d||a&&w!==d)throw new Error(`Size of X.shape()[axis:] == ${d}.\n       Size of scale and bias (if provided) must match this.\n       Got scale size of ${y} and bias size of ${w}`);let _=[];for(let O=0;O<i.length;++O)O<p?_.push(i[O]):_.push(1);let v=Me(d),S=["type","type"],A=[{type:12,data:h},{type:1,data:d},{type:12,data:Math.floor(d/v)},{type:1,data:t.epsilon}];a&&S.push("type");let I=r>1,x=r>2,E=O=>{let R=De(e[0].dataType),L=[U("x",e[0].dataType,e[0].dims,v),U("scale",u.dataType,u.dims,v)];a&&L.push(U("bias",a.dataType,a.dims,v)),L.push(j("output",e[0].dataType,c,v)),I&&L.push(j("mean_data_output",1,_)),x&&L.push(j("inv_std_output",1,_));let N=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return`\n  ${O.registerUniforms(N).declareVariables(...L)}\n  ${O.mainStart()}\n    ${O.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")}\n    let offset = global_idx * uniforms.norm_size_vectorized;\n    var mean_vector = ${$t("f32",v)};\n    var mean_square_vector = ${$t("f32",v)};\n\n    for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) {\n      let value = ${ir(R,v,"x[h + offset]")};\n      mean_vector += value;\n      mean_square_vector += value * value;\n    }\n    let mean = ${_t("mean_vector",v)} / uniforms.norm_size;\n    let inv_std_dev = inverseSqrt(${_t("mean_square_vector",v)} / uniforms.norm_size ${o?"":"- mean * mean"} + uniforms.epsilon);\n\n    for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) {\n      let f32input = ${ir(R,v,"x[j + offset]")};\n      let f32scale = ${ir(R,v,"scale[j]")};\n      output[j + offset] = ${L[0].type.value}((f32input ${o?"":"- mean"}) * inv_std_dev * f32scale\n        ${a?`+ ${ir(R,v,"bias[j]")}`:""}\n      );\n    }\n\n    ${I?"mean_data_output[global_idx] = mean":""};\n    ${x?"inv_std_output[global_idx] = inv_std_dev":""};\n  }`},P=[{dims:c,dataType:e[0].dataType}];return I&&P.push({dims:_,dataType:1}),x&&P.push({dims:_,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${v};${r};${o}`,inputDependencies:S},getRunData:()=>({outputs:P,dispatchGroup:{x:Math.ceil(h/64)},programUniforms:A}),getShaderSource:E}},Uu=(e,t)=>{tp(e.inputs),e.compute(rp(e.inputs,t,e.outputCount))}});var np,op,Wu,Nu,Gu=Y(()=>{"use strict";ye();Se();Ze();_e();np=(e,t)=>{if(e.length<3||e.length>4)throw new Error("MatMulNBits requires 3 or 4 inputs");let r=e[0],o=r.dims.length;if(r.dims[o-1]!==t.k)throw new Error("The last dim of input shape does not match the k value");let i=Math.floor((t.k+t.blockSize-1)/t.blockSize),u=t.blockSize/8*t.bits,a=e[1];if(!M.areEqual(a.dims,[t.n,i,u]))throw new Error("The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize");let p=e[2].dims;if(M.size(p)!==t.n*i)throw new Error("scales input size error.");if(e.length===4){let d=e[3].dims,y=t.bits>4?t.n*i:t.n*Math.floor((i+1)/2);if(M.size(d)!==y)throw new Error("zeroPoints input size error.")}},op=(e,t,r,o)=>{let i=e[0].dims,u=i.length,a=Math.floor((t.k+t.blockSize-1)/t.blockSize),c=i[u-2],p=t.k,h=t.n,d=i.slice(0,u-2),y=M.size(d),_=t.blockSize/8*t.bits/4,v=e[0].dataType,S=Me(c),A=Me(t.k),I=Me(_),x=tr(v),E=c*a*x,P=Math.floor(o/E),O=a<=r[0]&&P>0,R=!O||P>=4?Me(h):P>=2&&Me(h)>=2?2:1,L=d.concat([c,h]),N=M.size(L)/R/S,K=O?[]:[{type:12,data:N},{type:12,data:t.blockSize}],Q=[y,c,p/A],he=M.convertShape(e[1].dims).slice();he.splice(-1,1,_/I),K.push(...Z(Q)),K.push(...Z(he)),K.push(...Z(e[2].dims)),e.length===4&&K.push(...Z(M.convertShape(e[3].dims)));let W=[y,c,h/R];K.push(...Z(W));let se=Ce=>{let We=Q.length,ee=U("a",e[0].dataType,We,A),ae=U("b",12,he.length,I),Ae=U("scales",e[2].dataType,e[2].dims.length),me=[ee,ae,Ae],ie=e.length===4?U("zero_points",12,e[3].dims.length):void 0;ie&&me.push(ie);let ue=W.length,le=j("output",e[0].dataType,ue,R),qe=[{name:"output_size",type:"u32"},{name:"block_size",type:"u32"}],G=De(e[0].dataType),ne=(()=>{switch(A){case 1:return`array<${G}, 8>`;case 2:return`mat4x2<${G}>`;case 4:return`mat2x4<${G}>`;default:throw new Error(`${A}-component is not supported.`)}})(),xe=`\n        for (var word: u32 = 0; word < ${_}; word += ${I}) {\n          ${ae.indicesSet("b_indices","2","word")};\n          let b_data = ${ae.getByIndices("b_indices")};\n          for (var i: u32 = 0; i < ${I}; i++) {\n            let b_value: u32 = ${I===1?"b_data":"b_data[word + i]"};\n            let b_mask: u32 = 0x0F0F0F0Fu;\n            let b_value_lower: vec4<u32> = unpack4xU8(b_value & b_mask);\n            let b_value_upper: vec4<u32> = unpack4xU8((b_value >> 4) & b_mask);\n            let b_quantized_values = ${ne}(${Array.from({length:4},(Be,Ge)=>`${G}(b_value_lower[${Ge}]), ${G}(b_value_upper[${Ge}])`).join(", ")});\n            let b_dequantized_values = ${(()=>A===1?`${ne}(${Array.from({length:8},(Be,Ge)=>`(b_quantized_values[${Ge}] - zero_point) * scale`).join(", ")});`:`(b_quantized_values - ${ne}(${Array(8).fill("zero_point").join(",")})) * scale;`)()};\n            // Number of B elements per 32-bit word is 32/bits = 32/4 = 8\n            for (var m: u32 = 0; m < ${O?c:S}u; m++) {\n              ${ee.indicesSet("a_indices",We-2,O?"m":`row * ${S} + m`)};\n              ${ee.indicesSet("a_indices",We-1,"word_offset")};\n              var input_offset = ${ee.indicesToOffset("a_indices")};\n              var a_data: ${ne};\n              for (var j: u32 = 0; j < ${8/A}; j++) {\n                a_data[j] = ${ee.getByOffset("input_offset")};\n                input_offset++;\n              }\n              ${O?"workgroup_shared[workgroup_shared_offset + m]":"output_values[m]"}${R>1?"[c]":""} += ${Array.from({length:8/A},(Be,Ge)=>`${A===1?`a_data[${Ge}] * b_dequantized_values[${Ge}]`:`dot(a_data[${Ge}], b_dequantized_values[${Ge}])`}`).join(" + ")};\n            }\n            word_offset += ${8/A};\n          }\n        }`,Ke=ie?`\n          zero_point_offset += 4;\n          if (zero_point_offset == 32) {\n            zero_point_offset = 0;\n            zero_point_index++;\n            zero_point_word = ${ie.getByOffset("zero_point_index")};\n          }`:"";return O?`\n        var<workgroup> workgroup_shared: array<${le.type.value}, ${c*a}>;\n        ${Ce.declareVariables(...me,le)}\n        ${Ce.mainStart([a,1,1])}\n          var a_indices: ${ee.type.indices};\n          var block = local_id.x;\n          var col = workgroup_id.y;\n          var batch = workgroup_id.z;\n          ${ee.indicesSet("a_indices","0","batch")};\n          // Two zero points are packed into one byte when uniforms.bits is 4.\n          for (var c: u32 = 0; c < ${R}; c++) {\n            let col_times_components_plus_c = col * ${R} + c;\n              ${ie?`\n            var zero_point_bytes_per_col: u32 = (${a} + 1) / 2;\n            var zero_point_byte_count: u32 = col_times_components_plus_c * zero_point_bytes_per_col + (block >> 0x1u);\n            var zero_point_word_index: u32 = zero_point_byte_count >> 0x2u;\n            var zero_point_byte_offset: u32 = zero_point_byte_count & 0x3u;\n            var zero_point_nibble_offset: u32 = block & 0x1u;\n            var zero_point_bits_offset: u32 = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2);\n            var zero_point_word: u32 = ${ie.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;`:""}\n            var b_indices: ${ae.type.indices};\n            ${ae.indicesSet("b_indices","0","col_times_components_plus_c")};\n            // The scale and zero points are computed per block.\n            var scales_index = col_times_components_plus_c * ${a} + block;\n            let scale = ${Ae.getByOffset("scales_index")};\n            // The default zero point is 8 for unsigned 4-bit quantization.\n            let zero_point = ${G}(${ie?"(zero_point_word) & 0xFu":8});\n            ${ae.indicesSet("b_indices","1","block")};\n            var word_offset: u32 = block * ${t.blockSize/A};\n            var workgroup_shared_offset: u32 = block * ${c};\n            ${xe}\n          }\n          workgroupBarrier();\n          if (local_id.x == 0u) {\n            var output_indices: ${le.type.indices};\n            ${le.indicesSet("output_indices","0","batch")};\n            ${le.indicesSet("output_indices",ue-1,"col")};\n            ${le.indicesSet("output_indices",ue-2,"0")};\n            var output_offset = ${le.indicesToOffset("output_indices")};\n            for (var m: u32 = 0u; m < ${c}u; m++) {\n              var output_value: ${le.type.value} = ${le.type.value}(0);\n              var workgroup_shared_offset: u32 = m;\n              for (var b: u32 = 0u; b < ${a}u; b++) {\n                output_value += workgroup_shared[workgroup_shared_offset];\n                workgroup_shared_offset += ${c};\n              }\n              ${le.setByOffset("output_offset","output_value")};\n              output_offset += ${h/R};\n            }\n          }\n        }`:`\n        ${Ce.registerUniforms(qe).declareVariables(...me,le)}\n        ${Ce.mainStart()}\n          ${Ce.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n          var output_values: array<${le.type.value}, ${S}>;\n          var output_indices = ${le.offsetToIndices("global_idx")};\n          var col = ${le.indicesGet("output_indices",ue-1)};\n          var row = ${le.indicesGet("output_indices",ue-2)};\n          var a_indices: ${ee.type.indices} = output_indices;\n          // Two zero points are packed into one byte because uniforms.bits <= 4.\n          // zero_point_offset is either 0 or 4. It is bit offset within one byte.\n          // TODO support zero_point_offset for bits > 4\n          ${ie?`\n          var zero_point_abs_offset = col * ${R} * ((${a} + 1) / 2);\n          var zero_point_index: u32 = zero_point_abs_offset / 4;\n          var zero_point_word: u32 = ${ie.getByOffset("zero_point_index")};\n          var zero_point_offset: u32 = (zero_point_abs_offset % 4) * 8;`:""}\n          var scale_index = col * ${a*R};\n          var b_indices: ${ae.type.indices};\n          for (var c: u32 = 0; c < ${R}; c++) {\n            ${ae.indicesSet("b_indices","0",`col * ${R} + c`)};\n            var block_offset: u32 = 0;\n            for (var block: u32 = 0; block < ${a}; block++) {\n              // The scale and zero points are computed per block.\n              let scale = ${Ae.getByOffset("scale_index")};\n              // The default zero point is 8 for unsigned 4-bit quantization.\n              let zero_point = ${G}(${ie?"extractBits(zero_point_word, zero_point_offset, 4)":8});\n              ${ae.indicesSet("b_indices","1","block")};\n              var word_offset: u32 = block_offset;\n              ${xe}\n              scale_index++;\n              ${Ke}\n              block_offset += uniforms.block_size / ${A};\n            }\n            // Drop the trailing 4 bits if the zero_poit_offset is not a byte boundary to align with the next byte.\n            ${ie?`if (zero_point_offset % 8 > 0) {\n                ${Ke}\n              }`:""}\n            }\n            for (var k: u32 = 0u; k < ${S}u; k++) {\n              ${le.indicesSet("output_indices",ue-2,`${S} * row + k`)};\n              ${le.setByIndices("output_indices","output_values[k]")}\n            }\n        }`};return{name:O?"BlockwiseMatMulNBits":"MatMulNBits",shaderCache:{hint:`${t.cacheKey};${c};${v};${e.length}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:L,dataType:v}],name:O?"BlockwiseMatMulNBits":"MatMulNBits",dispatchGroup:O?{x:1,y:Math.ceil(h/R),z:y}:{x:Math.ceil(N/64)},programUniforms:K}),getShaderSource:se}},Wu=(e,t)=>{np(e.inputs,t);let r=e.getMaxComputeWorkgroupSizes(),o=e.getMaxComputeWorkgroupStoragesize();e.compute(op(e.inputs,t,r,o))},Nu=e=>ve(e)});var it,ip,Lu,Hu,ap,Ko,Fu,qu=Y(()=>{"use strict";ye();Se();Ze();_n();Ro();_e();Sr();it=(e,t)=>e.length>t&&e[t].dims.length>0&&M.size(e[t].dims)>0?e[t]:void 0,ip=(e,t)=>{let r=e[0],o=it(e,1),i=it(e,2),u=it(e,3),a=it(e,4),c=it(e,5),p=it(e,6),h=it(e,7);if(r.dims.length!==3&&r.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let d=!1,y=r.dims[0],w=r.dims[1],_=r.dims.length===3?d?r.dims[2]/3:r.dims[2]:t.numHeads*r.dims[4],v=w,S=0,A=0,I=Math.floor(_/t.numHeads);if(p&&h){if(p.dims.length!==4)throw new Error(\'Input "past_key" is expected to have 4 dimensions\');if(p.dims[0]!==y||p.dims[1]!==t.numHeads||p.dims[3]!==I)throw new Error(\'Input "past_key" shape (batch_size, num_heads, past_sequence_length, head_size)\');if(h.dims[0]!==y||h.dims[1]!==t.numHeads||h.dims[3]!==I)throw new Error(\'Input "past_value" shape (batch_size, num_heads, past_sequence_length, head_size)\');if(p.dims[2]!==h.dims[2])throw new Error(\'Input "past_key" and "past_value" shall have same dim 2 (past_sequence_length)\');if(h.dims.length!==4)throw new Error(\'Input "past_value" is expected to have 4 dimensions\');S=p.dims[2],A=p.dims[2]}else if(p||h)throw new Error(\'Input "past_key" and "past_value" shall be both present or both absent\');let x;if(o){if(r.dims.length!==3)throw new Error(\'Input "query" is expected to have 3 dimensions when key is given\');if(o.dims.length<3||o.dims.length>5)throw new Error(\'Input "key" is expected to have 3, 4, or 5 dimensions\');if(r.dims[0]!==o.dims[0])throw new Error(\'Input "query" and "key" shall have same dim 0 (batch size)\');if(o.dims.length===3){if(o.dims[2]!==r.dims[2])throw new Error(\'Input "query" and "key" shall have same dim 2 (hidden_size)\');x=2,v=o.dims[1]}else if(o.dims.length===5){if(o.dims[2]!==t.numHeads||o.dims[3]!==2||o.dims[4]!==I)throw new Error(\'Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv\');if(i)throw new Error(\'Expect "value" be none when "key" has packed kv format.\');x=5,v=o.dims[1]}else{if(o.dims[1]!==t.numHeads||o.dims[3]!==I)throw new Error(\'Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key\');x=0,v=o.dims[2]}}else{if(r.dims.length!==3&&r.dims.length!==5)throw new Error(\'Input "query" is expected to have 3 or 5 dimensions when key is empty\');if(r.dims.length===5&&(r.dims[2]!==t.numHeads||r.dims[3]!==3))throw new Error(\'Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv\');x=3}if(u){if(u.dims.length!==1)throw new Error(\'Input "bias" is expected to have 1 dimension\');if(i&&r.dims.length===5&&r.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let E=0;if(a){E=8;let N=a.dims;throw N.length===1?N[0]===y?E=1:N[0]===3*y+2&&(E=3):N.length===2&&N[0]===y&&N[1]===v&&(E=5),E===8?new Error(\'Input "key_padding_mask" shape shall be (batch_size) or (batch_size, kv_sequence_length)\'):new Error("Mask not supported")}let P=!1,O=_;if(i){if(i.dims.length!==3&&i.dims.length!==4)throw new Error(\'Input "value" is expected to have 3 or 4 dimensions\');if(r.dims[0]!==i.dims[0])throw new Error(\'Input "query" and "value" shall have same dim 0 (batch_size)\');if(i.dims.length===3){if(v!==i.dims[1])throw new Error(\'Input "key" and "value" shall have the same dim 1 (kv_sequence_length)\');O=i.dims[2]}else{if(v!==i.dims[2])throw new Error(\'Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)\');O=i.dims[1]*i.dims[3],P=!0}}let R=S+v,L=!1;if(a)throw new Error("Key padding mask is not supported");if(c){if(c.dims.length!==4)throw new Error(\'Input "relative_position_bias" is expected to have 4 dimensions\');if(c.dims[0]!==y&&c.dims[0]!==1||c.dims[1]!==t.numHeads||c.dims[2]!==w||c.dims[3]!==R)throw new Error(\'Input "relative_position_bias" shape (batch_size, 1, sequence_length, kv_sequence_length)\')}return{batchSize:y,sequenceLength:w,pastSequenceLength:S,kvSequenceLength:v,totalSequenceLength:R,maxSequenceLength:A,inputHiddenSize:0,hiddenSize:_,vHiddenSize:O,headSize:I,vHeadSize:Math.floor(O/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:E,scale:t.scale,broadcastResPosBias:L,passPastInKv:P,qkvFormat:x}},Lu=e=>ve({...e}),Hu=ve({perm:[0,2,1,3]}),ap=(e,t,r,o,i,u,a)=>{let c=[o,i,u],p=M.size(c),h=[{type:12,data:p},{type:12,data:a},{type:12,data:u}],d=y=>{let w=j("qkv_with_bias",t.dataType,c),_=U("qkv",t.dataType,c),v=U("bias",r.dataType,c),S=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`\n  ${y.registerUniforms(S).declareVariables(_,v,w)}\n  ${y.mainStart()}\n    ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n    let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset;\n\n    qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx];\n  }`};return e.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:c,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:h}),getShaderSource:d},{inputs:[t,r],outputs:[-1]})[0]},Ko=(e,t,r,o,i,u,a,c)=>{let p=u;if(a){if(o===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return p=ap(e,u,a,t,o,r*i,c),p=p.reshape([t,o,r,i]),e.compute(yt(p,Hu.perm),{inputs:[p],outputs:[-1]})[0]}else return u.dims.length===3&&(p=u.reshape([t,o,r,i])),e.compute(yt(p,Hu.perm),{inputs:[p],outputs:[-1]})[0]},Fu=(e,t)=>{let r=ip(e.inputs,t),o=e.inputs[0],i=it(e.inputs,1),u=it(e.inputs,2),a=it(e.inputs,3),c=it(e.inputs,4),p=it(e.inputs,5),h=it(e.inputs,6),d=it(e.inputs,7);if(o.dims.length===5)throw new Error("Packed QKV is not implemented");if(i?.dims.length===5)throw new Error("Packed KV is not implemented");let y=i&&u&&i.dims.length===4&&u.dims.length===4,w=Ko(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,o,a,0);if(y)return Pn(e,w,i,u,c,void 0,h,d,p,r,t);if(!i||!u)throw new Error("key and value must be provided");let _=Ko(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,i,a,r.hiddenSize),v=Ko(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,u,a,2*r.hiddenSize);Pn(e,w,_,v,c,void 0,h,d,p,r,t)}});var sp,up,dp,lp,cp,pp,mp,fp,ju,Ku=Y(()=>{"use strict";ye();Se();_e();sp=e=>{if(!e||e.length<1)throw new Error("Too few inputs");if(e[0].dataType!==1&&e[0].dataType!==10)throw new Error("Input type must be float or float16.");if(e.length>=2){let t=e[0].dims.length*2===e[1].dims[0];if(e.length===4&&(t=e[3].dims[0]*2===e[1].dims[0]),!t)throw new Error("The pads should be a 1D tensor of shape [2 * input_rank] or [2 * num_axes].")}},up=(e,t,r)=>{let o="";for(let i=t-1;i>=0;--i)o+=`\n            k = i32(${e.indicesGet("indices",i)}) - ${fe("uniforms.pads",i,r)};\n            if (k < 0) {\n              break;\n            }\n            if (k >= i32(${fe("uniforms.x_shape",i,t)})) {\n              break;\n            }\n            offset += k * i32(${fe("uniforms.x_strides",i,t)});\n        `;return`\n          value = ${e.type.value}(uniforms.constant_value);\n          for (var i = 0; i < 1; i++) {\n            var offset = 0;\n            var k = 0;\n            ${o}\n            value = x[offset];\n          }\n      `},dp=(e,t,r)=>{let o="";for(let i=t-1;i>=0;--i)o+=`\n                k = i32(${e.indicesGet("indices",i)}) - ${fe("uniforms.pads",i,r)};\n                if (k < 0) {\n                  k = -k;\n                }\n                {\n                  let _2n_1 = 2 * (i32(${fe("uniforms.x_shape",i,t)}) - 1);\n                  k = k % _2n_1;\n                  if(k >= i32(${fe("uniforms.x_shape",i,t)})) {\n                    k = _2n_1 - k;\n                  }\n                }\n                offset += k * i32(${fe("uniforms.x_strides",i,t)});\n            `;return`\n              var offset = 0;\n              var k = 0;\n              ${o}\n              value = x[offset];\n          `},lp=(e,t,r)=>{let o="";for(let i=t-1;i>=0;--i)o+=`\n                k = i32(${e.indicesGet("indices",i)}) - ${fe("uniforms.pads",i,r)};\n                if (k < 0) {\n                  k = 0;\n                }\n                if (k >= i32(${fe("uniforms.x_shape",i,t)})) {\n                  k = i32(${fe("uniforms.x_shape",i,t)}) - 1;\n                }\n                offset += k * i32(${fe("uniforms.x_strides",i,t)});\n            `;return`\n              var offset = 0;\n              var k = 0;\n              ${o}\n              value = x[offset];\n          `},cp=(e,t,r)=>{let o="";for(let i=t-1;i>=0;--i)o+=`\n                k = i32(${e.indicesGet("indices",i)}) - ${fe("uniforms.pads",i,r)};\n                if (k < 0)  {\n                  k += i32(${fe("uniforms.x_shape",i,t)}]);\n                }\n                if (k >= i32(${fe("uniforms.x_shape",i,t)})) {\n                  k -= i32(${fe("uniforms.x_shape",i,t)});\n                }\n                offset += k * i32(${fe("uniforms.x_strides",i,t)});\n            `;return`\n              var offset = 0;\n              var k = 0;\n              ${o}\n              value = x[offset];\n          `},pp=(e,t,r)=>{switch(r.mode){case 0:return up(e,t,r.pads.length);case 1:return dp(e,t,r.pads.length);case 2:return lp(e,t,r.pads.length);case 3:return cp(e,t,r.pads.length);default:throw new Error("Invalid mode")}},mp=(e,t)=>{let r=M.padShape(e[0].dims.slice(),t.pads),o=e[0].dims,i=M.size(r),u=[{type:12,data:i},{type:6,data:t.pads}];t.mode===0&&u.push({type:e[0].dataType,data:t.value}),u.push(...Z(e[0].dims,r));let a=["rank"],c=p=>{let h=j("output",e[0].dataType,r.length),d=U("x",e[0].dataType,o.length),y=d.type.value,w=pp(h,o.length,t),_=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:t.pads.length}];return t.mode===0&&_.push({name:"constant_value",type:y}),`\n            ${p.registerUniforms(_).declareVariables(d,h)}\n            ${p.mainStart()}\n            ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n            let indices = ${h.offsetToIndices("global_idx")};\n\n            var value = ${y}(0);\n            ${w}\n            output[global_idx] = value;\n        }`};return{name:"Pad",shaderCache:{hint:`${t.mode}`,inputDependencies:a},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(M.size(r)/64)},programUniforms:u}),getShaderSource:c}},fp=(e,t)=>{if(e.length>1){let r=e[1].getBigInt64Array(),o=e.length>=3&&e[2].data?e[2].getFloat32Array()[0]:0,i=e[0].dims.length,u=new Int32Array(2*i).fill(0);if(e.length>=4){let c=e[3].getBigInt64Array();for(let p=0;p<c.length;p++)u[Number(c[p])]=Number(r[p]),u[Number(c[p])+i]=Number(r[p+c.length])}else r.forEach((c,p)=>u[Number(p)]=Number(c));let a=[];return u.forEach(c=>a.push(c)),{mode:t.mode,value:o,pads:a}}else return t},ju=(e,t)=>{sp(e.inputs);let r=fp(e.inputs,t);e.compute(mp(e.inputs,r),{inputs:[0]})}});var Nn,Yu,Zu,Xu,Qu,hp,gp,Ju,ed,td,rd,nd,od,id,ad,sd,ud,dd,ld,cd=Y(()=>{"use strict";$r();ye();Se();_e();Nn=e=>{if(vr.webgpu.validateInputContent&&(!e||e.length!==1))throw new Error("Pool ops requires 1 input.")},Yu=(e,t,r)=>{let o=t.format==="NHWC",i=e.dims.slice();o&&i.splice(1,0,i.pop());let u=Object.hasOwnProperty.call(t,"dilations"),a=t.kernelShape.slice(),c=t.strides.slice(),p=u?t.dilations.slice():[],h=t.pads.slice();nr.adjustPoolAttributes(r,i,a,c,p,h);let d=nr.computePoolOutputShape(r,i,c,p,a,h,t.autoPad),y=Object.assign({},t);u?Object.assign(y,{kernelShape:a,strides:c,pads:h,dilations:p,cacheKey:t.cacheKey}):Object.assign(y,{kernelShape:a,strides:c,pads:h,cacheKey:t.cacheKey});let w=d.slice();return w.push(w.splice(1,1)[0]),[y,o?w:d]},Zu=(e,t)=>{let r=t.format==="NHWC",o=M.size(e),i=M.size(t.kernelShape),u=[{type:12,data:o},{type:12,data:i}],a=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(t.kernelShape.length<=2){let c=t.kernelShape[t.kernelShape.length-1],p=t.strides[t.strides.length-1],h=t.pads[t.pads.length/2-1],d=t.pads[t.pads.length-1],y=!!(h+d);u.push({type:12,data:c},{type:12,data:p},{type:12,data:h},{type:12,data:d}),a.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let w=!1;if(t.kernelShape.length===2){let _=t.kernelShape[t.kernelShape.length-2],v=t.strides[t.strides.length-2],S=t.pads[t.pads.length/2-2],A=t.pads[t.pads.length-2];w=!!(S+A),u.push({type:12,data:_},{type:12,data:v},{type:12,data:S},{type:12,data:A}),a.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[u,a,!0,y,w]}else{if(r)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let c=M.computeStrides(t.kernelShape);u.push({type:12,data:c},{type:12,data:t.pads},{type:12,data:t.strides}),a.push({name:"kernelStrides",type:"u32",length:c.length},{name:"pads",type:"u32",length:t.pads.length},{name:"strides",type:"u32",length:t.strides.length});let p=t.pads.reduce((h,d)=>h+d);return[u,a,!!p,!1,!1]}},Xu=(e,t,r,o,i,u,a,c,p,h,d,y)=>{let w=i.format==="NHWC",_=t.type.value,v=j("output",t.type.tensor,o);if(i.kernelShape.length<=2){let S="",A="",I="",x=r-(w?2:1);if(d?S=`\n                for (var i: u32 = 0u; i < uniforms.kw; i++) {\n                  xIndices[${x}] = indices[${x}] * uniforms.sw - uniforms.pwStart + i;\n                  if (xIndices[${x}] < 0 || xIndices[${x}]\n                      >= uniforms.x_shape[${x}]) {\n                    pad++;\n                    continue;\n                  }\n                  let x_val = x[${t.indicesToOffset("xIndices")}];\n                  ${u}\n                }`:S=`\n                for (var i: u32 = 0u; i < uniforms.kw; i++) {\n                  xIndices[${x}] = indices[${x}] * uniforms.sw - uniforms.pwStart + i;\n                  let x_val = x[${t.indicesToOffset("xIndices")}];\n                  ${u}\n                }`,i.kernelShape.length===2){let P=r-(w?3:2);y?A=`\n                for (var j: u32 = 0u; j < uniforms.kh; j++) {\n                  xIndices[${P}] = indices[${P}] * uniforms.sh - uniforms.phStart + j;\n                  if (xIndices[${P}] < 0 || xIndices[${P}] >= uniforms.x_shape[${P}]) {\n                    pad += i32(uniforms.kw);\n                    continue;\n                  }\n              `:A=`\n                for (var j: u32 = 0u; j < uniforms.kh; j++) {\n                  xIndices[${P}] = indices[${P}] * uniforms.sh - uniforms.phStart + j;\n                `,I=`\n              }\n            `}return`\n            ${e.registerUniforms(p).declareVariables(t,v)}\n\n            ${e.mainStart()}\n              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n\n              let indices = ${v.offsetToIndices("global_idx")};\n              var xIndices = ${v.offsetToIndices("global_idx")};\n\n              var value = ${_}(${c});\n              var pad = 0;\n              ${A}\n              ${S}\n              ${I}\n              ${a}\n\n              output[global_idx] = value;\n            }`}else{if(w)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let S=i.kernelShape.length,A=i.pads.length,I="";return h?I=`\n                if (xIndices[j] >= uniforms.x_shape[j]) {\n                  pad++;\n                  isPad = true;\n                  break;\n                }\n              }\n              if (!isPad) {\n                let x_val = x[${t.indicesToOffset("xIndices")}];\n                ${u}\n              }`:I=`\n              }\n              let x_val = x[${t.indicesToOffset("xIndices")}];\n              ${u}\n            `,`\n            ${e.registerUniforms(p).declareVariables(t,v)}\n\n            ${e.mainStart()}\n              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n              let indices = ${v.offsetToIndices("global_idx")};\n              var xIndices = ${v.offsetToIndices("global_idx")};\n\n              var offsets: array<u32, ${S}>;\n\n              var value = ${_}(${c});\n              var pad = 0;\n              var isPad = false;\n\n              for (var i: u32 = 0u; i < uniforms.kernelSize; i++) {\n                var offset = i;\n                for (var j = 0u; j < ${S-1}u; j++) {\n                  offsets[j] = offset / ${fe("uniforms.kernelStrides","j",S)};\n                  offset -= offsets[j] * ${fe("uniforms.kernelStrides","j",S)};\n                }\n                offsets[${S-1}] = offset;\n\n                isPad = false;\n                for (var j = ${r-S}u; j < ${r}u; j++) {\n                  xIndices[j] = indices[j] * ${fe("uniforms.strides",`j - ${r-S}u`,S)}\n                    + offsets[j - ${r-S}u] - ${fe("uniforms.pads","j - 2u",A)};\n                  ${I}\n              }\n              ${a}\n\n              output[global_idx] = value;\n            }`}},Qu=e=>`${e.format};${e.ceilMode};${e.autoPad};${e.kernelShape.length}`,hp=e=>`${Qu(e)};${e.countIncludePad}`,gp=e=>`${Qu(e)};${e.storageOrder};${e.dilations}`,Ju=e=>({format:e.format,autoPad:["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],ceilMode:e.ceil_mode,kernelShape:e.kernel_shape,strides:e.strides,pads:e.pads}),ed=(e,t,r,o)=>{let[i,u]=Yu(t,o,r),a=U("x",t.dataType,t.dims.length),c=a.type.value,p="value += x_val;",h="";i.countIncludePad?h+=`value /= ${c}(uniforms.kernelSize);`:h+=`value /= ${c}(i32(uniforms.kernelSize) - pad);`;let[d,y,w,_,v]=Zu(u,i);d.push(...Z(t.dims,u));let S=["rank"];return{name:e,shaderCache:{hint:`${o.cacheKey};${w};${_};${v}`,inputDependencies:S},getRunData:()=>({outputs:[{dims:u,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(M.size(u)/64)},programUniforms:d}),getShaderSource:A=>Xu(A,a,t.dims.length,u.length,i,p,h,0,y,w,_,v)}},td=e=>{let t=e.count_include_pad!==0,r=Ju(e);if(r.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for AveragePool");let o={countIncludePad:t,...r,cacheKey:""};return{...o,cacheKey:hp(o)}},rd=(e,t)=>{Nn(e.inputs),e.compute(ed("AveragePool",e.inputs[0],!1,t))},nd={autoPad:"",ceilMode:0,countIncludePad:!1,kernelShape:[],strides:[],pads:[],storageOrder:0,dilations:[]},od=e=>{let t=e.format;return{format:t,...nd,cacheKey:t}},id=(e,t)=>{Nn(e.inputs),e.compute(ed("GlobalAveragePool",e.inputs[0],!0,t))},ad=(e,t,r,o)=>{let[i,u]=Yu(t,o,r),a=`\n      value = max(x_val, value);\n    `,c="",p=U("x",t.dataType,t.dims.length),h=["rank"],[d,y,w,_,v]=Zu(u,i);return d.push(...Z(t.dims,u)),{name:e,shaderCache:{hint:`${o.cacheKey};${w};${_};${v}`,inputDependencies:h},getRunData:()=>({outputs:[{dims:u,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(M.size(u)/64)},programUniforms:d}),getShaderSource:S=>Xu(S,p,t.dims.length,u.length,i,a,c,t.dataType===10?-65504:-1e5,y,w,_,v)}},sd=(e,t)=>{Nn(e.inputs),e.compute(ad("MaxPool",e.inputs[0],!1,t))},ud=e=>{let t=e.storage_order,r=e.dilations,o=Ju(e);if(t!==0)throw new Error("column major storage order is not yet supported for MaxPool");if(o.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for MaxPool");let i={storageOrder:t,dilations:r,...o,cacheKey:""};return{...i,cacheKey:gp(i)}},dd=e=>{let t=e.format;return{format:t,...nd,cacheKey:t}},ld=(e,t)=>{Nn(e.inputs),e.compute(ad("GlobalMaxPool",e.inputs[0],!0,t))}});var bp,wp,pd,md=Y(()=>{"use strict";$r();ye();_e();bp=(e,t,r)=>{let o=e===t,i=e<t&&r<0,u=e>t&&r>0;if(o||i||u)throw new Error("Range these inputs\' contents are invalid.")},wp=(e,t,r,o)=>{let i=Math.abs(Math.ceil((t-e)/r)),u=[i],a=i,c=[{type:12,data:a},{type:o,data:e},{type:o,data:r},...Z(u)],p=h=>{let d=j("output",o,u.length),y=d.type.value,w=[{name:"outputSize",type:"u32"},{name:"start",type:y},{name:"delta",type:y}];return`\n        ${h.registerUniforms(w).declareVariables(d)}\n        ${h.mainStart()}\n        ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n        output[global_idx] = uniforms.start + ${y}(global_idx) * uniforms.delta;\n      }`};return{name:"Range",shaderCache:{hint:`${o}`},getShaderSource:p,getRunData:()=>({outputs:[{dims:u,dataType:o}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:c})}},pd=e=>{let t=0,r=0,o=0;e.inputs[0].dataType===6?(t=e.inputs[0].getInt32Array()[0],r=e.inputs[1].getInt32Array()[0],o=e.inputs[2].getInt32Array()[0]):e.inputs[0].dataType===1&&(t=e.inputs[0].getFloat32Array()[0],r=e.inputs[1].getFloat32Array()[0],o=e.inputs[2].getFloat32Array()[0]),vr.webgpu.validateInputContent&&bp(t,r,o),e.compute(wp(t,r,o,e.inputs[0].dataType),{inputs:[]})}});var vp,$p,_p,Sp,xp,Cp,Ap,Ip,Tp,Ep,Pp,fd,kp,Op,Rp,Bp,Dp,hd,gd,yd=Y(()=>{"use strict";ye();Se();Ze();_e();vp=(e,t)=>{if(e.every(r=>r>0||(()=>{throw new Error("Resize requires scales input values to be positive")})),e.length>0){if(t.mode==="linear"){if(!(e.length===2||e.length===3||e.length===4&&e[0]===1&&e[1]===1||e.length===4&&e[0]===1&&e[3]===1||e.length===5&&e[0]===1&&e[1]===1))throw new Error(`For linear mode, Resize requires scales to be 2D, 3D, 4D with either two outermost or one innermost and\n            one outermost scale values equal to 1, or 5D with two outermost scale values equal to 1`)}else if(t.mode==="cubic"&&!(e.length===2||e.length===4&&e[0]===1&&e[1]===1||e.length===4&&e[0]===1&&e[3]===1))throw new Error("Resize requires scales input size to be 2 or 4 for cubic mode")}},$p=(e,t,r)=>{t.every(i=>i>=0&&i<r||(()=>{throw new Error("Resize requires axes input values to be positive and less than rank")}));let o=new Array(r).fill(1);return t.forEach((i,u)=>o[i]=e[u]),o},_p=(e,t,r,o,i,u)=>{let[a,c,p]=r>10?[1,2,3]:[-1,e.length>1?1:-1,-1],h=e[0].dims.length;if(a>0&&e.length>a&&e[a].dims.length>0)e[a].getFloat32Array().forEach(d=>u.push(d));else if(t.coordinateTransformMode==="tf_crop_and_resize")throw new Error("Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize");if(c>0&&e.length>c&&e[c].dims.length>0){if(e[c].getFloat32Array().forEach(d=>o.push(d)),o.length!==0&&o.length!==h&&r>=18&&o.length!==t.axes.length)throw new Error("Resize requires scales input size to be same as input rank or axes size for opset 18 and up");vp(o,t),t.axes.length>0&&$p(o,t.axes,h).forEach((d,y)=>o[y]=d)}if(p>0&&e.length>p&&(e[p].getBigInt64Array().forEach(d=>i.push(Number(d))),i.length!==h||r>=18&&i.length===t.axes.length))throw new Error("Resize requires sizes input size to be same as input rank or axes size for opset 18 and up");if(t.axes.length>0){if(o.length!==t.axes.length)throw new Error(\'Resize requires "scales" input size to be of axes rank when axes attributes is specified\');if(i.length!==t.axes.length)throw new Error(\'Resize requires "sizes" input size to be of rank axes rank when axes attributes is specified\')}if(typeof o<"u"&&typeof i<"u"&&o.length>0&&i.length>h)throw new Error("Resize requires only of scales or sizes to be specified")},Sp=(e,t)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32,\n     lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${t} { `+(()=>{switch(e){case"asymmetric":return`return ${t}(xResized) / ${t}(xScale);`;case"pytorch_half_pixel":return`if (lengthResized > 1) {\n                    return (${t}(xResized) + 0.5) / ${t}(xScale) - 0.5;\n                  } else {\n                    return 0.0;\n                  }`;case"tf_half_pixel_for_nn":return`return (${t}(xResized) + 0.5) / ${t}(xScale);`;case"align_corners":return`if (lengthResized == 1) {\n                    return 0.0;\n                  } else {\n                    // The whole part and the fractional part are calculated separately due to inaccuracy of floating\n                    // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an\n                    // offset-by-one error later in floor().\n                    let whole = ${t}(xResized * (lengthOriginal - 1) / (lengthResized - 1));\n                    let fract =\n                        ${t}(xResized * (lengthOriginal - 1) % (lengthResized - 1)) / ${t}(lengthResized - 1);\n                    return whole + fract;\n                  }`;case"tf_crop_and_resize":return`if (lengthResized > 1) {\n                    return ${t}(roiStart) * ${t}(lengthOriginal - 1) +\n                        (${t}(xResized) * ${t}(roiEnd - roiStart) * ${t}(lengthOriginal - 1)) /\n                        ${t}(lengthResized - 1);\n                  } else {\n                    return 0.5 * ${t}(roiStart + roiEnd) * ${t}(lengthOriginal - 1);\n                  }`;case"half_pixel_symmetric":return`const outputWidth = ${t}xScale * ${t}(lengthResized);\n                  const adjustment = ${t}(lengthResized) / outputWidth;\n                  const center = ${t}(lengthOriginal) / 2;\n                  const offset = center * (1 - adjustment);\n                  return offset + ((${t}(xResized) + 0.5) / ${t}(xScale)) - 0.5;`;case"half_pixel":return`return ((${t}(xResized) + 0.5) / ${t}(xScale)) - 0.5;`;default:throw new Error(`Coordinate transform mode ${e} is not supported`)}})()+"}",xp=(e,t,r)=>`fn getNearestPixelFromOriginal(xOriginal: ${r}, isDownSample: bool) -> ${r} {`+(()=>{switch(e){case"round_prefer_ceil":return"if (fract(xOriginal) == 0.5) {             return ceil(xOriginal);           } else {             return round(xOriginal);           }";case"floor":return"return floor(xOriginal);";case"ceil":return"return ceil(xOriginal);";case"round_prefer_floor":return"if (fract(xOriginal) == 0.5) {                     return floor(xOriginal);                   } else {                     return round(xOriginal);                   }";case"simple":default:if(t<11)return"if (isDownSample)                     {                       return ceil(xOriginal);                     } else {                       return xOriginal;                     }";throw new Error(`Nearest mode ${e} is not supported`)}})()+"}",Cp=(e,t,r)=>{let o=new Array(r).fill(0).concat(new Array(r).fill(1)),i=e.length===0?o:e.slice();return t.length>0?(t.forEach((u,a)=>{o[u]=i[a],o[a+r]=i[t.length+a]}),o):i},Ap=(e,t,r,o)=>{let i=[];if(r.length>0)if(o.length>0){if(e.forEach(u=>i.push(u)),Math.max(...o)>e.length)throw new Error("axes is out of bound");o.forEach((u,a)=>i[u]=r[a])}else r.forEach(u=>i.push(u));else{if(t.length===0)throw new Error("Resize requires either scales or sizes.");i=e.map((u,a)=>Math.round(u*t[a]))}return i},Ip=(e,t,r)=>{let o=(()=>{switch(r.keepAspectRatioPolicy){case"not_larger":return r.axes.length>0?Math.min(...r.axes.map(u=>t[u]),Number.MAX_VALUE):Math.min(...t,Number.MAX_VALUE);case"not_smaller":return r.axes.length>0?Math.max(...r.axes.map(u=>t[u]),Number.MIN_VALUE):Math.max(...t,Number.MIN_VALUE);default:throw new Error(`Keep aspect ratio policy ${r.keepAspectRatioPolicy} is not supported`)}})();t.fill(1,0,t.length);let i=e.slice();return r.axes.length>0?(r.axes.forEach(u=>t[u]=o),r.axes.forEach(u=>i[u]=Math.round(e[u]*t[u]))):(t.fill(o,0,t.length),i.forEach((u,a)=>i[a]=Math.round(u*t[a]))),i},Tp=(e,t,r,o,i)=>`\n    fn calculateOriginalIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> array<${e.type.value}, ${r.length}> {\n      var original_indices: array<${e.type.value}, ${r.length}>;\n      for (var i:u32 = 0; i < ${r.length}; i++) {\n        var output_index = ${e.indicesGet("output_indices","i")};\n        var scale = ${fe("uniforms.scales","i",o)};\n        var roi_low = ${fe("uniforms.roi","i",i)};\n        var roi_hi = ${fe("uniforms.roi",`i + ${t.length}`,i)};\n        if (scale == 1.0) {\n          original_indices[i] = ${e.type.value}(output_index);\n        } else {\n          var input_shape_i = ${fe("uniforms.input_shape","i",t.length)};\n          var output_shape_i = ${fe("uniforms.output_shape","i",r.length)};\n          original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,\n                                                                           input_shape_i, roi_low, roi_hi);\n        }\n      }\n      return original_indices;\n    }`,Ep=(e,t,r,o,i,u,a)=>`\n    fn calculateInputIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {\n      var input_indices: ${e.type.indices};\n      for (var i:u32 = 0; i < ${o.length}; i++) {\n        var output_index = ${t.indicesGet("output_indices","i")};\n        var input_index: u32;\n        var scale = ${fe("uniforms.scales","i",i)};\n        if (scale == 1.0) {\n          input_index = output_index;\n        } else {\n          var roi_low = ${fe("uniforms.roi","i",u)};\n          var roi_hi = ${fe("uniforms.roi",`i + ${r.length}`,u)};\n          var input_shape_i = ${fe("uniforms.input_shape","i",r.length)};\n          var output_shape_i = ${fe("uniforms.output_shape","i",o.length)};\n          var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,\n                                                                        input_shape_i, roi_low, roi_hi);\n          if (!${a} || (original_idx >= 0 && original_idx < ${t.type.value}(input_shape_i))) {\n            if (original_idx < 0) {\n              input_index = 0;\n            } else if (original_idx > ${t.type.value}(input_shape_i - 1)) {\n              input_index = input_shape_i - 1;\n            } else {\n              input_index = u32(getNearestPixelFromOriginal(original_idx, scale < 1));\n            }\n          } else {\n            input_index = u32(original_idx);\n          }\n        }\n        ${e.indicesSet("input_indices","i"," input_index")}\n      }\n      return input_indices;\n    }`,Pp=(e,t)=>`\n    fn checkInputIndices(input_indices: ${e.type.indices}) -> bool {\n      for (var i:u32 = 0; i < ${t.length}; i++) {\n        var input_index = ${e.indicesGet("input_indices","i")};\n        if (input_index < 0 || input_index >= ${fe("uniforms.input_shape","i",t.length)}) {\n          return false;\n        }\n      }\n      return true;\n    }`,fd=(e,t,r,o)=>e.rank>o?`\n    ${e.indicesSet("input_indices",t,"channel")};\n    ${e.indicesSet("input_indices",r,"batch")};\n`:"",kp=(e,t,r,o,i)=>{let[a,c,p,h]=r.length===2?[-1,0,1,-1]:[0,2,3,1],d=e.type.value;return`\n    fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${d} {\n      var input_indices: ${e.type.indices};\n      ${e.indicesSet("input_indices",c,`max(0, min(row, ${r[c]} - 1))`)};\n      ${e.indicesSet("input_indices",p,`max(0, min(col, ${r[p]} - 1))`)};\n      ${fd(e,h,a,2)}\n      return ${e.getByIndices("input_indices")};\n    }\n\n    fn bilinearInterpolation(output_indices: ${t.type.indices}) -> ${d} {\n      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);\n      var row:${d} = originalIndices[${c}];\n      var col:${d} = originalIndices[${p}];\n      ${o?`if (row < 0 || row > (${r[c]} - 1) || col < 0 || col > (${r[p]} - 1)) {\n        return ${i};\n      }`:""};\n      row = max(0, min(row, ${r[c]} - 1));\n      col = max(0, min(col, ${r[p]} - 1));\n      var row1: u32 = u32(row);\n      var col1: u32 = u32(col);\n      var row2: u32 = u32(row + 1);\n      var col2: u32 = u32(col + 1);\n      var channel: u32 = ${r.length>2?`u32(originalIndices[${h}])`:"0"};\n      var batch: u32 =  ${r.length>2?`u32(originalIndices[${a}])`:"0"};\n      var x11: ${d} = getInputValue(batch, channel, row1, col1);\n      var x12: ${d} = getInputValue(batch, channel, row1, col2);\n      var x21: ${d} = getInputValue(batch, channel, row2, col1);\n      var x22: ${d} = getInputValue(batch, channel, row2, col2);\n      var dx1: ${d} = abs(row - ${d}(row1));\n      var dx2: ${d} = abs(${d}(row2) - row);\n      var dy1: ${d} = abs(col - ${d}(col1));\n      var dy2: ${d} = abs(${d}(col2) - col);\n      if (row1 == row2) {\n        dx1 = 0.5;\n        dx2 = 0.5;\n      }\n      if (col1 == col2) {\n        dy1 = 0.5;\n        dy2 = 0.5;\n      }\n      return (x11 * dx2 * dy2 + x12 * dx2 * dy1 + x21 * dx1 * dy2 + x22 * dx1 * dy1);\n    }`},Op=(e,t,r,o,i,u,a,c,p,h)=>{let d=r.length===2,y=!0,[w,_]=d?[0,1]:y?[2,3]:[1,2],v=e.type.value,S=A=>{let I=A===w?"row":"col";return`\n      fn ${I}CubicInterpolation(input_indices: ${e.type.indices}, output_indices: ${t.type.indices}) -> ${v} {\n        var output_index = ${t.indicesGet("output_indices",A)};\n        var originalIdx: ${v} = getOriginalCoordinateFromResizedCoordinate(output_index, ${i[A]},\n        ${o[A]}, ${r[A]}, ${u[A]}, ${u[A]} + ${r.length});\n        var fractOriginalIdx: ${v} = originalIdx - floor(originalIdx);\n        var coefs = getCubicInterpolationCoefs(fractOriginalIdx);\n\n        if (${c} && (originalIdx < 0 || originalIdx > (${r[A]} - 1))) {\n          return ${p};\n        }\n        var data: array<${v}, 4> = array<${v}, 4>(0.0, 0.0, 0.0, 0.0);\n        for (var i: i32 = -1; i < 3; i++) {\n          var ${I}: ${v} = originalIdx + ${v}(i);\n          if (${I} < 0 || ${I} >= ${r[A]}) {\n            ${(()=>h?`coefs[i + 1] = 0.0;\n                        continue;`:c?`return ${p};`:`${I} = max(0, min(${I}, ${r[A]} - 1));`)()};\n          }\n        var input_indices_copy: ${e.type.indices} = input_indices;\n          ${e.indicesSet("input_indices_copy",A,`u32(${I})`)};\n          data[i + 1] = ${A===w?e.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"};\n        }\n        return cubicInterpolation1D(data, coefs);\n      }`};return`\n    ${S(w)};\n    ${S(_)};\n  fn getCubicInterpolationCoefs(s: ${v}) -> array<${v}, 4> {\n    var absS = abs(s);\n    var coeffs: array<${v}, 4> = array<${v}, 4>(0.0, 0.0, 0.0, 0.0);\n    var oneMinusAbsS: ${v} = 1.0 - absS;\n    var twoMinusAbsS: ${v} = 2.0 - absS;\n    var onePlusAbsS: ${v} = 1.0 + absS;\n    coeffs[0] = ((${a} * onePlusAbsS - 5 * ${a}) * onePlusAbsS + 8 * ${a}) * onePlusAbsS - 4 * ${a};\n    coeffs[1] = ((${a} + 2) * absS - (${a} + 3)) * absS * absS + 1;\n    coeffs[2] = ((${a} + 2) * oneMinusAbsS - (${a} + 3)) * oneMinusAbsS * oneMinusAbsS + 1;\n    coeffs[3] = ((${a} * twoMinusAbsS - 5 * ${a}) * twoMinusAbsS + 8 * ${a}) * twoMinusAbsS - 4 * ${a};\n    return coeffs;\n  }\n\n  fn cubicInterpolation1D(x: array<${v}, 4>, coefs: array<${v}, 4>) -> ${v} {\n    var coefsSum: ${v} = coefs[0] + coefs[1] + coefs[2] + coefs[3];\n    return (x[0] * coefs[0] + x[1] * coefs[1]+ x[2] * coefs[2]+ x[3] * coefs[3]) / coefsSum;\n  }\n\n  fn bicubicInterpolation(output_indices: ${t.type.indices}) -> ${v} {\n    var input_indices: ${e.type.indices} = output_indices;\n    return colCubicInterpolation(input_indices, output_indices);\n  }\n    `},Rp=(e,t,r,o,i)=>{let[a,c,p,h,d]=r.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],y=e.type.value;return`\n    fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${y} {\n      var input_indices: ${e.type.indices};\n      ${e.indicesSet("input_indices",c,`max(0, min(depth, ${r[c]} - 1))`)};\n      ${e.indicesSet("input_indices",p,`max(0, min(height, ${r[p]} - 1))`)};\n      ${e.indicesSet("input_indices",h,`max(0, min(width, ${r[h]} - 1))`)};\n      ${fd(e,d,a,3)}\n      return ${e.getByIndices("input_indices")};\n    }\n\n    fn trilinearInterpolation(output_indices: ${t.type.indices}) -> ${y} {\n      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);\n      var depth:${y} = originalIndices[${c}];\n      var height:${y} = originalIndices[${p}];\n      var width:${y} = originalIndices[${h}];\n      ${o?`if (depth < 0 || depth > (${r[c]} - 1) || height < 0 || height > (${r[p]} - 1) || width < 0 || (width > ${r[h]} - 1)) {\n      return ${i};\n        }`:""};\n\n    depth = max(0, min(depth, ${r[c]} - 1));\n      height = max(0, min(height, ${r[p]} - 1));\n      width = max(0, min(width, ${r[h]} - 1));\n      var depth1: u32 = u32(depth);\n      var height1: u32 = u32(height);\n      var width1: u32 = u32(width);\n      var depth2: u32 = u32(depth + 1);\n      var height2: u32 = u32(height + 1);\n      var width2: u32 = u32(width + 1);\n      var channel: u32 = ${r.length>3?`u32(originalIndices[${d}])`:"0"};\n      var batch: u32 =  ${r.length>3?`u32(originalIndices[${a}])`:"0"};\n\n      var x111: ${y} = getInputValue(batch, channel, depth1, height1, width1);\n      var x112: ${y} = getInputValue(batch, channel, depth1, height1, width2);\n      var x121: ${y} = getInputValue(batch, channel, depth1, height2, width1);\n      var x122: ${y} = getInputValue(batch, channel, depth1, height2, width2);\n      var x211: ${y} = getInputValue(batch, channel, depth2, height1, width1);\n      var x212: ${y} = getInputValue(batch, channel, depth2, height1, width2);\n      var x221: ${y} = getInputValue(batch, channel, depth2, height2, width1);\n      var x222: ${y} = getInputValue(batch, channel, depth2, height2, width2);\n      var dx1: ${y} = abs(depth - ${y}(depth1));\n      var dx2: ${y} = abs(${y}(depth2) - depth);\n      var dy1: ${y} = abs(height - ${y}(height1));\n      var dy2: ${y} = abs(${y}(height2) - height);\n      var dz1: ${y} = abs(width - ${y}(width1));\n      var dz2: ${y} = abs(${y}(width2) - width);\n      if (depth1 == depth2) {\n        dx1 = 0.5;\n        dx2 = 0.5;\n      }\n      if (height1 == height2) {\n        dy1 = 0.5;\n        dy2 = 0.5;\n      }\n      if (width1 == width2) {\n        dz1 = 0.5;\n        dz2 = 0.5;\n      }\n      return (x111 * dx2 * dy2 * dz2 + x112 * dx2 * dy2 * dz1 + x121 * dx2 * dy1 *dz2 + x122 * dx2 * dy1 * dz1 +\n              x211 * dx1 * dy2 * dz2 + x212 * dx1 * dy2 * dz1 + x221 * dx1 * dy1 *dz2 + x222 * dx1 * dy1 * dz1);\n    }`},Bp=(e,t,r,o,i,u)=>{let a=e.dims,c=Cp(u,t.axes,a.length),p=Ap(a,o,i,t.axes),h=o.slice();o.length===0&&(h=a.map((x,E)=>x===0?1:p[E]/x),t.keepAspectRatioPolicy!=="stretch"&&(p=Ip(a,h,t)));let d=j("output",e.dataType,p.length),y=U("input",e.dataType,a.length),w=M.size(p),_=a.length===p.length&&a.every((x,E)=>x===p[E]),v=t.coordinateTransformMode==="tf_crop_and_resize",S=t.extrapolationValue,A=y.type.value,I=x=>`\n      ${_?"":`\n      ${Sp(t.coordinateTransformMode,A)};\n      ${(()=>{switch(t.mode){case"nearest":return`\n              ${Pp(y,a)};\n              ${xp(t.nearestMode,r,A)};\n              ${Ep(y,d,a,p,h.length,c.length,v)};\n              `;case"linear":return`\n              ${Tp(d,a,p,h.length,c.length)};\n              ${(()=>{if(a.length===2||a.length===4)return`${kp(y,d,a,v,S)}`;if(a.length===3||a.length===5)return`${Rp(y,d,a,v,S)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()};\n            `;case"cubic":return`\n            ${(()=>{if(a.length===2||a.length===4)return`${Op(y,d,a,p,h,c,t.cubicCoeffA,v,t.extrapolationValue,t.excludeOutside)}`;throw Error("Cubic mode only supports input dims 2 and 4 are supported in linear mode.")})()};\n            `;default:throw Error("Invalid resize mode")}})()};\n      `}\n      ${x.registerUniform("output_size","u32").registerUniform("scales","f32",h.length).registerUniform("roi","f32",c.length).declareVariables(y,d)}\n      ${x.mainStart()}\n        ${x.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n        ${_?"output[global_idx] = input[global_idx];":`\n        let output_indices = ${d.offsetToIndices("global_idx")};\n        var input_indices: ${y.type.indices};\n        ${(()=>{switch(t.mode){case"nearest":return`input_indices = calculateInputIndicesFromOutputIndices(output_indices);\n                if (checkInputIndices(input_indices)) {\n                  output[global_idx] = ${y.getByIndices("input_indices")};\n                } else {\n                  output[global_idx] = ${t.extrapolationValue};\n                }`;case"linear":return`output[global_idx] = ${a.length===2||a.length===4?"bilinearInterpolation":"trilinearInterpolation"}(output_indices);`;case"cubic":return"output[global_idx] = bicubicInterpolation(output_indices);";default:throw Error(`Unsupported resize mode: ${t.mode}`)}})()};\n`}\n      }`;return{name:"Resize",shaderCache:{hint:`${t.cacheKey}|${r}|${h.length>0?h:""}|${i.length>0?i:""}|${c.length>0?c:""}|${_}|${a}`,inputDependencies:["rank"]},getShaderSource:I,getRunData:()=>({outputs:[{dims:p,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(w/64)},programUniforms:[{type:12,data:w},{type:1,data:h},{type:1,data:c},...Z(a,p)]})}},Dp=e=>{let t=e.customDataBuffer;return new Uint32Array(t,t.byteOffset,1)[0]},hd=(e,t)=>{let r=[],o=[],i=[],u=Dp(e);if(t.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");_p(e.inputs,t,u,r,o,i),e.compute(Bp(e.inputs[0],t,u,r,o,i),{inputs:[0]})},gd=e=>{let t=e.antialias,r=e.axes,o=e.coordinateTransformMode,i=e.cubicCoeffA,u=e.excludeOutside!==0,a=e.extrapolationValue,c=e.keepAspectRatioPolicy,p=e.mode,h=e.nearestMode===""?"simple":e.nearestMode;return ve({antialias:t,axes:r,coordinateTransformMode:o,cubicCoeffA:i,excludeOutside:u,extrapolationValue:a,keepAspectRatioPolicy:c,mode:p,nearestMode:h})}});var zp,Mp,bd,wd=Y(()=>{"use strict";ye();Se();Ze();_e();zp=(e,t)=>{let[r,o,i,u]=e,{numHeads:a,rotaryEmbeddingDim:c}=t;if(r.dims.length!==3&&r.dims.length!==4)throw new Error(`Input \'x\' is expected to have 3 or 4 dimensions, got ${r.dims.length}`);if(!M.areEqual(o.dims,[])&&!M.areEqual(o.dims,[1])&&o.dims.length!==2)throw new Error(`Input \'position_ids\' is expected to have 0, 1, or 2 dimensions, got ${o.dims.length}`);if(i.dims.length!==2)throw new Error(`Input \'cos_cache\' is expected to have 2 dimensions, got ${i.dims.length}`);if(u.dims.length!==2)throw new Error(`Input \'sin_cache\' is expected to have 2 dimensions, got ${u.dims.length}`);if(!M.areEqual(i.dims,u.dims))throw new Error("Inputs \'cos_cache\' and \'sin_cache\' are expected to have the same shape");if(c>0&&a===0)throw new Error("num_heads must be provided if rotary_embedding_dim is specified");let p=r.dims[0],h=r.dims[r.dims.length-2],d=i.dims[0],y=M.sizeFromDimension(r.dims,1)/h,w=c===0?i.dims[1]*2:y/a;if(c>w)throw new Error("rotary_embedding_dim must be less than or equal to head_size");if(o.dims.length===2){if(p!==o.dims[0])throw new Error(`Input \'position_ids\' dimension 0 should be of size batch_size, got ${o.dims[0]}`);if(h!==o.dims[1])throw new Error(`Input \'position_ids\' dimension 1 should be of size sequence_length, got ${o.dims[1]}`)}if(w/2!==i.dims[1]&&c/2!==i.dims[1])throw new Error(`Input \'cos_cache\' dimension 1 should be same as head_size / 2 or rotary_embedding_dim / 2, got ${i.dims[1]}`);if(h>d)throw new Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported")},Mp=(e,t)=>{let{interleaved:r,numHeads:o,rotaryEmbeddingDim:i,scale:u}=t,a=e[0].dims[0],c=M.sizeFromDimension(e[0].dims,1),p=e[0].dims[e[0].dims.length-2],h=c/p,d=e[2].dims[1],y=i===0?d*2:h/o,w=new Array(a,p,h/y,y-d),_=M.computeStrides(w),v=[{type:1,data:u},{type:12,data:w},{type:12,data:_},...e[0].dims.length===3?new Array({type:12,data:[c,h,y,1]}):[],...e[0].dims.length===4?new Array({type:12,data:[c,y,p*y,1]}):[],...Z(e[0].dims,e[1].dims,e[2].dims,e[3].dims,e[0].dims)],S=A=>{let I=U("input",e[0].dataType,e[0].dims.length),x=U("position_ids",e[1].dataType,e[1].dims.length),E=U("cos_cache",e[2].dataType,e[2].dims.length),P=U("sin_cache",e[3].dataType,e[3].dims.length),O=j("output",e[0].dataType,e[0].dims.length);return A.registerUniforms([{name:"scale",type:"f32"},{name:"global_shape",type:"u32",length:w.length},{name:"global_strides",type:"u32",length:_.length},{name:"input_output_strides",type:"u32",length:_.length}]),`\n        ${A.declareVariables(I,x,E,P,O)}\n\n        ${A.mainStart(or)}\n          let half_rotary_emb_dim = uniforms.${E.name}_shape[1];\n          let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape;\n          let size = uniforms.global_shape[0] * uniforms.global_strides[0];\n          ${A.guardAgainstOutOfBoundsWorkgroupSizes("size")}\n\n          if (bsnh[3] < half_rotary_emb_dim) {\n            let position_ids_idx =\n                ${x.broadcastedIndicesToOffset("bsnh.xy",j("",x.type.tensor,2))};\n            let position_id =\n                u32(${x.getByOffset("position_ids_idx")}) + select(0, bsnh[1], position_ids_idx == 0);\n            let i = dot(bsnh, uniforms.input_output_strides) + select(0, bsnh[3], ${r});\n            let j = i + select(half_rotary_emb_dim, 1, ${r});\n            let re = ${I.getByOffset("i")} * ${E.get("position_id","bsnh[3]")} -\n                ${I.getByOffset("j")} * ${P.get("position_id","bsnh[3]")};\n            ${O.setByOffset("i","re")}\n            let im = ${I.getByOffset("i")} * ${P.get("position_id","bsnh[3]")} +\n                ${I.getByOffset("j")} * ${E.get("position_id","bsnh[3]")};\n            ${O.setByOffset("j","im")}\n          } else {\n            let k = dot(bsnh, uniforms.input_output_strides) + half_rotary_emb_dim;\n            ${O.setByOffset("k",I.getByOffset("k"))}\n          }\n        }`};return{name:"RotaryEmbedding",shaderCache:{hint:ve({interleaved:r}).cacheKey,inputDependencies:["rank","rank","rank","rank"]},getShaderSource:S,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(M.size(w)/or)},programUniforms:v})}},bd=(e,t)=>{zp(e.inputs,t),e.compute(Mp(e.inputs,t))}});var Up,Vp,vd,$d=Y(()=>{"use strict";ye();Se();_e();Up=e=>{if(!e||e.length<3)throw new Error("layerNorm requires at least 3 inputs.");let t=e[0],r=e[1],o=e[2];if(t.dataType!==r.dataType||t.dataType!==o.dataType)throw new Error("All inputs must have the same data type");if(t.dims.length!==3&&t.dims.length!==2)throw new Error("Input must be 2D or 3D");if(r.dims.length!==3&&r.dims.length!==2)throw new Error("Skip must be 2D or 3D");let i=t.dims[t.dims.length-1],u=t.dims[t.dims.length-2];if(r.dims[r.dims.length-1]!==i)throw new Error("Skip must have the same hidden size as input");if(r.dims[r.dims.length-2]!==u)throw new Error("Skip must have the same sequence length as input");if(o.dims.length!==1)throw new Error("Gamma must be 1D");if(o.dims[o.dims.length-1]!==i)throw new Error("Gamma must have the same hidden size as input");if(e.length>3){let a=e[3];if(a.dims.length!==1)throw new Error("Beta must be 1D");if(a.dims[a.dims.length-1]!==i)throw new Error("Beta must have the same hidden size as input")}if(e.length>4){let a=e[4];if(a.dims.length!==1)throw new Error("Bias must be 1D");if(a.dims[a.dims.length-1]!==i)throw new Error("Bias must have the same hidden size as input")}},Vp=(e,t,r,o)=>{let i=t.simplified,u=e[0].dims,a=M.size(u),c=u,p=a,h=u.slice(-1)[0],d=o?u.slice(0,-1).concat(1):[],y=!i&&e.length>3,w=e.length>4,_=o&&r>1,v=o&&r>2,S=r>3,A=Me(h),I=[{type:12,data:p},{type:12,data:A},{type:12,data:h},{type:1,data:t.epsilon}],x=P=>{let O=[{name:"output_size",type:"u32"},{name:"components",type:"u32"},{name:"hidden_size",type:"u32"},{name:"epsilon",type:"f32"}],R=[U("x",e[0].dataType,e[0].dims,A),U("skip",e[1].dataType,e[1].dims,A),U("gamma",e[2].dataType,e[2].dims,A)];y&&R.push(U("beta",e[3].dataType,e[3].dims,A)),w&&R.push(U("bias",e[4].dataType,e[4].dims,A)),R.push(j("output",e[0].dataType,c,A)),_&&R.push(j("mean_output",1,d)),v&&R.push(j("inv_std_output",1,d)),S&&R.push(j("input_skip_bias_sum",e[0].dataType,c,A));let L=De(e[0].dataType);return`\n\n      ${P.registerUniforms(O).declareVariables(...R)}\n\n      ${P.mainStart()}\n        ${P.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size / uniforms.hidden_size")}\n        let hidden_size_vectorized: u32 = uniforms.hidden_size / uniforms.components;\n        let offset = global_idx * hidden_size_vectorized;\n        var sum = ${$t("f32",A)};\n        var squareSum = ${$t("f32",A)};\n        for (var i: u32 = 0; i < hidden_size_vectorized; i++) {\n          let skip_value = skip[offset + i];\n          let bias_value = ${w?"bias[i]":L+"(0.0)"};\n          let input_value = x[offset + i];\n          let value = input_value + skip_value + bias_value;\n          ${S?"input_skip_bias_sum[offset + i] = value;":""}\n          output[offset + i] = value;\n          let f32_value = ${ir(L,A,"value")};\n          sum += f32_value;\n          squareSum += f32_value * f32_value;\n        }\n        let mean = ${_t("sum",A)} / f32(uniforms.hidden_size);\n        let inv_std_dev = inverseSqrt(${_t("squareSum",A)} / f32(uniforms.hidden_size) ${i?"":"- mean * mean"} + uniforms.epsilon);\n        ${_?"mean_output[global_idx] = mean;":""}\n        ${v?"inv_std_output[global_idx] = inv_std_dev;":""}\n        for (var i: u32 = 0; i < hidden_size_vectorized; i++) {\n          output[offset + i] = (output[offset + i] ${i?"":`- ${L}(mean)`}) * ${L}(inv_std_dev) * gamma[i] ${y?"+ beta[i]":""};\n        }\n      }`},E=[{dims:c,dataType:e[0].dataType}];return r>1&&E.push({dims:d,dataType:1}),r>2&&E.push({dims:d,dataType:1}),r>3&&E.push({dims:u,dataType:e[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:`${A};${_};${v};${S}`,inputDependencies:e.map((P,O)=>"type")},getShaderSource:x,getRunData:()=>({outputs:E,dispatchGroup:{x:Math.ceil(p/h/64)},programUniforms:I})}},vd=(e,t)=>{Up(e.inputs);let o=[0];e.outputCount>1&&o.push(-3),e.outputCount>2&&o.push(-3),e.outputCount>3&&o.push(3),e.compute(Vp(e.inputs,t,e.outputCount,!1),{outputs:o})}});var Wp,Gn,Np,_d,Gp,Hp,Sd,xd,Cd=Y(()=>{"use strict";ye();Se();Ze();_e();Wp=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");if(t.axes.length!==0){if(t.axes.length!==t.starts.length||t.axes.length!==t.ends.length)throw new Error("axes, starts and ends must have the same length")}else if(t.starts.length!==t.ends.length)throw new Error("starts and ends must have the same length");e.slice(1).forEach((r,o)=>{if(e[o+1].dataType!==6&&e[o+1].dataType!==7)throw new Error(`Input ${o} must be an array of int32 or int64`)})},Gn=(e,t)=>{let r=[];if(e.length>t)if(e[t].dataType===7)e[t].getBigInt64Array().forEach(o=>r.push(Number(o)));else if(e[t].dataType===6)e[t].getInt32Array().forEach(o=>r.push(Number(o)));else throw new Error(`Input ${t} must be an array of int32 or int64`);return r},Np=(e,t)=>{if(e.length>1){let r=Gn(e,1),o=Gn(e,2),i=Gn(e,3);return i.length===0&&(i=[...Array(e[0].dims.length).keys()]),ve({starts:r,ends:o,axes:i})}else return t},_d=(e,t,r,o,i)=>{let u=e;return e<0&&(u+=r[o[t]]),i[t]<0?Math.max(0,Math.min(u,r[o[t]]-1)):Math.max(0,Math.min(u,r[o[t]]))},Gp=(e,t,r)=>`fn calculateInputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {\n          var input_indices: ${e.type.indices};\n          var carry = 0u;\n          for (var i = ${r.length}; i >= 0; i--) {\n            let input_shape_i = ${fe("uniforms.input_shape","i",r.length)};\n            let steps_i = ${fe("uniforms.steps","i",r.length)};\n            let signs_i = ${fe("uniforms.signs","i",r.length)};\n            let starts_i = ${fe("uniforms.starts","i",r.length)};\n            var output_index = ${t.indicesGet("output_indices","i")};\n            var input_index = output_index * steps_i + starts_i + carry;\n            carry = input_index / input_shape_i;\n            input_index = input_index % input_shape_i;\n            if (signs_i < 0) {\n              input_index = input_shape_i - input_index - 1u + starts_i;\n            }\n            ${e.indicesSet("input_indices","i","input_index")};\n          }\n          return input_indices;\n      }`,Hp=(e,t)=>{let r=e[0].dims,o=M.size(r),i=t.axes.length>0?M.normalizeAxes(t.axes,r.length):[...Array(r.length).keys()],u=Gn(e,4);u.forEach(I=>I!==0||(()=>{throw new Error("step cannot be 0")})),u.length===0&&(u=Array(i.length).fill(1));let a=t.starts.map((I,x)=>_d(I,x,r,i,u)),c=t.ends.map((I,x)=>_d(I,x,r,i,u));if(i.length!==a.length||i.length!==c.length)throw new Error("start, ends and axes should have the same number of elements");if(i.length!==r.length)for(let I=0;I<r.length;++I)i.includes(I)||(a.splice(I,0,0),c.splice(I,0,r[I]),u.splice(I,0,1));let p=u.map(I=>Math.sign(I));u.forEach((I,x,E)=>{if(I<0){let P=(c[x]-a[x])/I,O=a[x],R=O+P*u[x];a[x]=R,c[x]=O,E[x]=-I}});let h=r.slice(0);i.forEach((I,x)=>{h[I]=Math.ceil((c[I]-a[I])/u[I])});let d={dims:h,dataType:e[0].dataType},y=j("output",e[0].dataType,h.length),w=U("input",e[0].dataType,e[0].dims.length),_=M.size(h),v=[{name:"outputSize",type:"u32"},{name:"starts",type:"u32",length:a.length},{name:"signs",type:"i32",length:p.length},{name:"steps",type:"u32",length:u.length}],S=[{type:12,data:_},{type:12,data:a},{type:6,data:p},{type:12,data:u},...Z(e[0].dims,h)],A=I=>`\n      ${I.registerUniforms(v).declareVariables(w,y)}\n        ${Gp(w,y,r)}\n        ${I.mainStart()}\n          ${I.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n          let output_indices = ${y.offsetToIndices("global_idx")};\n          let input_indices = calculateInputIndices(output_indices);\n          ${y.setByOffset("global_idx",w.getByIndices("input_indices"))}\n      }`;return{name:"Slice",shaderCache:{hint:`${p.length}_${a.length}_${u.length}`,inputDependencies:["rank"]},getShaderSource:A,getRunData:()=>({outputs:[d],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:S})}},Sd=(e,t)=>{Wp(e.inputs,t);let r=Np(e.inputs,t);e.compute(Hp(e.inputs,r),{inputs:[0]})},xd=e=>{let t=e.starts,r=e.ends,o=e.axes;return ve({starts:t,ends:r,axes:o})}});var Lp,Fp,Ad,Id,Td=Y(()=>{"use strict";ye();Se();Ze();_e();Lp=e=>{if(!e||e.length!==1)throw new Error("Softmax op requires 1 input.")},Fp=(e,t)=>{let r=e.dims,o=M.size(r),i=64,u=t.axis;if(u<0&&(u=r.length+u),u<r.length-1)throw new Error("softmax only supports last axis for now.");let a=r[u],c=o/a,p=Me(a),h=a/p,d=(A,I)=>I===4?`max(max(${A}.x, ${A}.y), max(${A}.z, ${A}.w))`:I===2?`max(${A}.x, ${A}.y)`:I===3?`max(max(${A}.x, ${A}.y), ${A}.z)`:A,y=U("x",e.dataType,e.dims,p),w=j("result",e.dataType,e.dims,p),_=y.type.value,v=De(e.dataType)==="f32"?`var threadMax = ${_}(-3.402823e+38f);`:`var threadMax = ${_}(-65504.0h);`,S=A=>`\n      var<workgroup> rowMaxShared : ${_};\n      var<workgroup> rowSumShared : ${_};\n      var<workgroup> threadShared : array<${_}, ${i}>;\n\n      fn getValue(row: i32, col: i32, row_stride: i32) -> ${_} {\n        let index = row * row_stride + col;\n        return x[index];\n      }\n\n      fn setValue(row: i32, col: i32, row_stride: i32, value: ${_}) {\n        let index = row * row_stride + col;\n        result[index] = value;\n      }\n      ${A.registerUniform("packedCols","i32").declareVariables(y,w)}\n      ${A.mainStart()}\n        let gindex = i32(global_idx);\n        let lindex = i32(local_idx);\n        const wg = ${i};\n        let row = gindex / wg;\n        let cols = uniforms.packedCols;\n        let row_stride : i32 = uniforms.packedCols;\n\n        // find the rows max\n        ${v}\n        for (var col = lindex; col < cols; col += wg) {\n          let value = getValue(row, col, row_stride);\n          threadMax = max(threadMax, value);\n        }\n        if (lindex < cols) {\n          threadShared[lindex] = threadMax;\n        }\n        workgroupBarrier();\n\n        var reduceSize = min(cols, wg);\n        for (var currSize = reduceSize >> 1;  currSize > 0; currSize = reduceSize >> 1) {\n          reduceSize = currSize + (reduceSize & 1);\n          if (lindex < currSize) {\n            threadShared[lindex] = max(threadShared[lindex], threadShared[lindex + reduceSize]);\n          }\n          workgroupBarrier();\n        }\n        if (lindex == 0) {\n          rowMaxShared = ${_}(${d("threadShared[0]",p)});\n        }\n        workgroupBarrier();\n\n        // find the rows sum\n        var threadSum = ${_}(0.0);\n        for (var col = lindex; col < cols; col += wg) {\n          let subExp = exp(getValue(row, col, row_stride) - rowMaxShared);\n          threadSum += subExp;\n        }\n        threadShared[lindex] = threadSum;\n        workgroupBarrier();\n\n        for (var currSize = wg >> 1;  currSize > 0; currSize = currSize >> 1) {\n          if (lindex < currSize) {\n            threadShared[lindex] = threadShared[lindex] + threadShared[lindex + currSize];\n          }\n          workgroupBarrier();\n        }\n        if (lindex == 0) {\n          rowSumShared = ${_}(${_t("threadShared[0]",p)});\n        }\n        workgroupBarrier();\n\n        // calculate final value for each element in the row\n        for (var col = lindex; col < cols; col += wg) {\n          let value = exp(getValue(row, col, row_stride) - rowMaxShared) / rowSumShared;\n          setValue(row, col, row_stride, value);\n        }\n      }`;return{name:"Softmax",shaderCache:{hint:`${p}`,inputDependencies:["type"]},getRunData:()=>({outputs:[{dims:r,dataType:e.dataType}],dispatchGroup:{x:c},programUniforms:[{type:6,data:h}]}),getShaderSource:S}},Ad=(e,t)=>{Lp(e.inputs),e.compute(Fp(e.inputs[0],t))},Id=e=>ve({axis:e.axis})});var qp,jp,Kp,Yp,Zp,Ed,Pd,kd=Y(()=>{"use strict";ye();Se();Ze();_e();qp=e=>{if(!e||e.length<1)throw new Error("too few inputs")},jp=(e,t)=>{let r=[],o=t.numOutputs;return e[1].dims[0]>0&&(e[1].getBigInt64Array().forEach(i=>r.push(Number(i))),o=r.length),ve({numOutputs:o,axis:t.axis,splitSizes:r})},Kp=e=>`\nfn calculateOutputIndex(index: u32) -> u32 {\n    for (var i: u32 = 0u; i < ${e}u; i += 1u ) {\n    if (index < ${fe("uniforms.size_in_split_axis","i",e)}) {\n        return i;\n    }\n    }\n    return ${e}u;\n}`,Yp=e=>{let t=e.length,r=[];for(let o=0;o<t;++o){let i=e[o].setByIndices("indices","input[global_idx]");t===1?r.push(i):o===0?r.push(`if (output_number == ${o}u) { ${i} }`):o===t-1?r.push(`else { ${i} }`):r.push(`else if (output_number == ${o}) { ${i} }`)}return`\n      fn writeBufferData(output_number: u32, indices: ${e[0].type.indices}, global_idx: u32) {\n        ${r.join(`\n`)}\n      }`},Zp=(e,t)=>{let r=e[0].dims,o=M.size(r),i=e[0].dataType,u=M.normalizeAxis(t.axis,r.length),a=new Array(t.numOutputs),c=U("input",i,r.length),p=new Array(t.numOutputs),h=[],d=[],y=0,w=[{type:12,data:o}];for(let v=0;v<t.numOutputs;v++){y+=t.splitSizes[v],p[v]=y;let S=r.slice();S[t.axis]=t.splitSizes[v],d.push(S),a[v]=j(`output${v}`,i,S.length),h.push({dims:d[v],dataType:e[0].dataType})}w.push({type:12,data:p},...Z(r,...d));let _=v=>`\n  ${v.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",p.length).declareVariables(c,...a)}\n  ${Kp(p.length)}\n  ${Yp(a)}\n\n  ${v.mainStart()}\n    ${v.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size")}\n\n    var indices = ${c.offsetToIndices("global_idx")};\n    var index = ${c.indicesGet("indices",u)};\n    let output_number = calculateOutputIndex(index);\n    if (output_number != 0) {\n      index -= ${fe("uniforms.size_in_split_axis","output_number - 1u",p.length)};\n      ${c.indicesSet("indices",u,"index")};\n    }\n    writeBufferData(output_number, indices, global_idx);\n  }`;return{name:"Split",shaderCache:{hint:t.cacheKey,inputDependencies:["rank"]},getShaderSource:_,getRunData:()=>({outputs:h,dispatchGroup:{x:Math.ceil(o/64)},programUniforms:w})}},Ed=(e,t)=>{qp(e.inputs);let r=e.inputs.length===1?t:jp(e.inputs,t);e.compute(Zp(e.inputs,r),{inputs:[0]})},Pd=e=>{let t=e.axis,r=e.splitSizes,o=e.numOutputs<0?r.length:e.numOutputs;if(o!==r.length)throw new Error("numOutputs and splitSizes lengh must be equal");return ve({axis:t,numOutputs:o,splitSizes:r})}});var Od,Xp,Qp,Jp,Rd,Bd=Y(()=>{"use strict";ye();Se();_e();Od=e=>Array.from(e.getBigInt64Array(),Number),Xp=e=>{if(!e||e.length!==2)throw new Error("Tile requires 2 inputs.");if(e[0].dataType!==1&&e[0].dataType!==6&&e[0].dataType!==12)throw new Error("Tile only support float, int32, and uint32 data types");if(e[1].dataType!==7)throw new Error("Tile `repeats` input should be of int64 data type");if(e[1].dims.length!==1)throw new Error("Tile `repeats` input should be 1-D");if(Od(e[1]).length!==e[0].dims.length)throw new Error("Tile `repeats` input should have same number of elements as rank of input data tensor")},Qp=(e,t)=>{let r=[];for(let o=0;o<e.length;++o)r.push(e[o]*t[o]);return r},Jp=e=>{let t=e[0].dims,r=Od(e[1]),o=Qp(t,r),i=M.size(o),u=e[0].dataType,a=U("input",u,t.length),c=j("output",u,o.length),p=h=>`\n      const inputShape = ${a.indices(...t)};\n      ${h.registerUniform("output_size","u32").declareVariables(a,c)}\n      ${h.mainStart()}\n      ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n      let output_indices = ${c.offsetToIndices("global_idx")};\n      var input_indices: ${a.type.indices};\n      for (var i = 0; i < ${t.length}; i++) {\n        let input_dim_i = ${a.indicesGet("uniforms.input_shape","i")};\n        let input_dim_value = ${c.indicesGet("output_indices","i")}  % input_dim_i;\n\n        ${a.indicesSet("input_indices","i","input_dim_value")}\n      }\n      ${c.setByOffset("global_idx",a.getByIndices("input_indices"))}\n    }`;return{name:"Tile",shaderCache:{hint:`${r}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:[{type:12,data:i},...Z(e[0].dims,o)]}),getShaderSource:p}},Rd=e=>{Xp(e.inputs),e.compute(Jp(e.inputs),{inputs:[0]})}});var em,tm,Dd,zd=Y(()=>{"use strict";ye();Se();_e();em=(e,t,r,o,i)=>{let u=j("output_data",i,r.length,4),a=U("a_data",t[1].dataType,t[1].dims.length,4),c=U("b_data",t[2].dataType,t[2].dims.length,4),p=U("c_data",t[0].dataType,t[0].dims.length,4),h,d=(y,w,_)=>`select(${w}, ${y}, ${_})`;if(!o)h=u.setByOffset("global_idx",d(a.getByOffset("global_idx"),c.getByOffset("global_idx"),p.getByOffset("global_idx")));else{let y=(w,_,v="")=>{let S=`a_data[index_a${_}][component_a${_}]`,A=`b_data[index_b${_}][component_b${_}]`,I=`bool(c_data[index_c${_}] & (0xffu << (component_c${_} * 8)))`;return`\n            let output_indices${_} = ${u.offsetToIndices(`global_idx * 4u + ${_}u`)};\n            let offset_a${_} = ${a.broadcastedIndicesToOffset(`output_indices${_}`,u)};\n            let offset_b${_} = ${c.broadcastedIndicesToOffset(`output_indices${_}`,u)};\n            let offset_c${_} = ${p.broadcastedIndicesToOffset(`output_indices${_}`,u)};\n            let index_a${_} = offset_a${_} / 4u;\n            let index_b${_} = offset_b${_} / 4u;\n            let index_c${_} = offset_c${_} / 4u;\n            let component_a${_} = offset_a${_} % 4u;\n            let component_b${_} = offset_b${_} % 4u;\n            let component_c${_} = offset_c${_} % 4u;\n            ${w}[${_}] = ${v}(${d(S,A,I)});\n          `};i===9?h=`\n            var data = vec4<u32>(0);\n            ${y("data",0,"u32")}\n            ${y("data",1,"u32")}\n            ${y("data",2,"u32")}\n            ${y("data",3,"u32")}\n            output_data[global_idx] = dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(data));`:h=`\n            ${y("output_data[global_idx]",0)}\n            ${y("output_data[global_idx]",1)}\n            ${y("output_data[global_idx]",2)}\n            ${y("output_data[global_idx]",3)}\n          `}return`\n        ${e.registerUniform("vec_size","u32").declareVariables(p,a,c,u)}\n        ${e.mainStart()}\n        ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}\n        ${h}\n      }`},tm=e=>{let t=e[1].dims,r=e[2].dims,o=e[0].dims,i=e[1].dataType,u=!(M.areEqual(t,r)&&M.areEqual(r,o)),a=t,c=M.size(t);if(u){let h=It.calcShape(It.calcShape(t,r,!1),o,!1);if(!h)throw new Error("Can\'t perform where op on the given tensors");a=h,c=M.size(a)}let p=Math.ceil(c/4);return{name:"Where",shaderCache:{inputDependencies:["rank","rank","rank"]},getShaderSource:h=>em(h,e,a,u,i),getRunData:()=>({outputs:[{dims:a,dataType:i}],dispatchGroup:{x:Math.ceil(c/64/4)},programUniforms:[{type:12,data:p},...Z(o,t,r,a)]})}},Dd=e=>{e.compute(tm(e.inputs))}});var Md,Ud=Y(()=>{"use strict";Ka();Ro();Ja();ts();Vs();Zs();Oo();Uo();lu();mu();gu();$u();xu();Au();Eu();Ou();Du();Mu();Vu();Wo();Gu();qu();Ku();cd();md();In();yd();wd();$d();Cd();Td();kd();Bd();Sr();Rn();zd();Md=new Map([["Abs",[rs]],["Acos",[ns]],["Acosh",[os]],["Add",[Ws]],["ArgMax",[ja,ko]],["ArgMin",[qa,ko]],["Asin",[is]],["Asinh",[as]],["Atan",[ss]],["Atanh",[us]],["Attention",[Xa]],["AveragePool",[rd,td]],["BatchNormalization",[Qa]],["BiasAdd",[es]],["BiasSplitGelu",[Us]],["Cast",[ls,ds]],["Ceil",[ps]],["Clip",[cs]],["Concat",[Ya,Za]],["Conv",[Ho,Go]],["ConvTranspose",[du,uu]],["Cos",[ms]],["Cosh",[fs]],["CumSum",[cu,pu]],["DepthToSpace",[fu,hu]],["Div",[Ns]],["Einsum",[wu,vu]],["Elu",[hs,kn]],["Equal",[Gs]],["Erf",[gs]],["Exp",[ys]],["Expand",[Su]],["FastGelu",[Cu]],["Floor",[bs]],["FusedConv",[Ho,Go]],["Gather",[Tu,Iu]],["GatherElements",[ku,Pu]],["Gelu",[ws]],["Gemm",[Bu,Ru]],["GlobalAveragePool",[id,od]],["GlobalMaxPool",[ld,dd]],["Greater",[qs]],["GreaterOrEqual",[Ks]],["HardSigmoid",[Is,As]],["InstanceNormalization",[zu]],["LayerNormalization",[Uu]],["LeakyRelu",[vs,kn]],["Less",[js]],["LessOrEqual",[Ys]],["Log",[zs]],["MatMul",[ru]],["MatMulNBits",[Wu,Nu]],["MaxPool",[sd,ud]],["Mul",[Hs]],["MultiHeadAttention",[Fu,Lu]],["Neg",[_s]],["Not",[$s]],["Pad",[ju]],["Pow",[Ls]],["Range",[pd]],["Reciprocal",[Ss]],["ReduceMin",[Wa]],["ReduceMean",[Da]],["ReduceMax",[Va]],["ReduceSum",[Ga]],["ReduceProd",[Na]],["ReduceL1",[za]],["ReduceL2",[Ma]],["ReduceLogSum",[La]],["ReduceLogSumExp",[Ua]],["ReduceSumSquare",[Ha]],["Relu",[xs]],["Resize",[hd,gd]],["RotaryEmbedding",[bd]],["Sigmoid",[Cs]],["Sin",[Ts]],["Sinh",[Es]],["Slice",[Sd,xd]],["SkipLayerNormalization",[vd]],["Split",[Ed,Pd]],["Sqrt",[Ps]],["Softmax",[Ad,Id]],["Sub",[Fs]],["Tan",[ks]],["Tanh",[Rs]],["ThresholdedRelu",[Ds,kn]],["Tile",[Rd]],["Transpose",[_a,Sa]],["Where",[Dd]]])});var Hn,Vd=Y(()=>{"use strict";$r();Lt();_e();Hn=class{constructor(t){this.backend=t;this.repo=new Map,this.attributesBound=!1}getArtifact(t){return this.repo.get(t)}setArtifact(t,r){this.repo.set(t,r)}run(t,r,o,i,u){rr(t.programInfo.name);let a=this.backend.device,c=this.backend.getComputePassEncoder();this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2);let p=[];for(let d of r)p.push({binding:p.length,resource:{buffer:d.buffer}});for(let d of o)p.push({binding:p.length,resource:{buffer:d.buffer}});u&&p.push({binding:p.length,resource:u});let h=a.createBindGroup({layout:t.computePipeline.getBindGroupLayout(0),entries:p,label:t.programInfo.name});if(this.backend.sessionStatus==="capturing"){let d={kernelId:this.backend.currentKernelId,computePipeline:t.computePipeline,bindGroup:h,dispatchGroup:i};this.backend.capturedCommandList.get(this.backend.currentSessionId).push(d)}c.setPipeline(t.computePipeline),c.setBindGroup(0,h),c.dispatchWorkgroups(...i),this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2+1),this.backend.pendingDispatchNumber++,(this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber||this.backend.queryType==="at-passes")&&this.backend.endComputePass(),this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber&&this.backend.flush(),Ht(t.programInfo.name)}dispose(){}build(t,r){rr(t.name);let o=this.backend.device,i=[];o.features.has("shader-f16")&&i.push("enable f16;");let u=va(r,this.backend.device.limits),a=t.getShaderSource(u),c=`${i.join(`\n`)}\n${u.additionalImplementations}\n${a}`,p=o.createShaderModule({code:c,label:t.name});Ve("verbose",()=>`[WebGPU] ${t.name} shader code: ${c}`);let h=o.createComputePipeline({compute:{module:p,entryPoint:"main"},layout:"auto",label:t.name});return Ht(t.name),{programInfo:t,computePipeline:h,uniformVariablesInfo:u.variablesInfo}}normalizeDispatchGroupSize(t){let r=typeof t=="number"?t:t.x,o=typeof t=="number"?1:t.y||1,i=typeof t=="number"?1:t.z||1,u=this.backend.device.limits.maxComputeWorkgroupsPerDimension;if(r<=u&&o<=u&&i<=u)return[r,o,i];let a=r*o*i,c=Math.ceil(Math.sqrt(a));if(c>u){if(c=Math.ceil(Math.cbrt(a)),c>u)throw new Error("Total dispatch size exceeds WebGPU maximum.");return[c,c,c]}else return[c,c,1]}}});var rm,nm,Yo,Ln,Wd=Y(()=>{"use strict";$r();ye();Lt();fa();wa();Ud();Vd();rm=(e,t)=>{if(t.length!==e.length)throw new Error(`inputDependencies length ${t.length} is not equal to inputTensors length ${e.length}.`);let r=[];for(let o=0;o<e.length;++o){let i=e[o].dataType;switch(t[o]){case"none":{r.push("");break}case"type":{r.push(`${i}`);break}case"rank":{let u=e[o].dims.length;r.push(`${i};${u}`);break}case"dims":{let u=e[o].dims.join(",");r.push(`${i};${u}`);break}default:throw new Error(`unsupported input dependency: ${t[o]}`)}}return r.join("|")},nm=(e,t,r)=>{let o=e.name;return e.shaderCache?.hint&&(o+="["+e.shaderCache.hint+"]"),o+=":"+r+`:${rm(t,e.shaderCache?.inputDependencies??new Array(t.length).fill("dims"))}`,o},Yo=class{constructor(t){t&&(this.architecture=t.architecture,this.vendor=t.vendor)}isArchitecture(t){return this.architecture===t}isVendor(t){return this.vendor===t}},Ln=class{constructor(){this.currentSessionId=null;this.currentKernelId=null;this.commandEncoder=null;this.computePassEncoder=null;this.maxDispatchNumber=16;this.pendingDispatchNumber=0;this.pendingKernels=[];this.pendingQueries=new Map;this.sessionStatus="default";this.capturedCommandList=new Map;this.capturedPendingKernels=new Map;this.sessionExternalDataMapping=new Map}get currentKernelCustomData(){if(this.currentKernelId===null)throw new Error("currentKernelCustomData(): currentKernelId is null. (should not happen)");let t=this.kernelCustomData.get(this.currentKernelId);return t||(t={},this.kernelCustomData.set(this.currentKernelId,t)),t}async initialize(t,r){this.env=t;let o=[],i={requiredLimits:{maxComputeWorkgroupStorageSize:r.limits.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:r.limits.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:r.limits.maxStorageBufferBindingSize,maxBufferSize:r.limits.maxBufferSize,maxComputeInvocationsPerWorkgroup:r.limits.maxComputeInvocationsPerWorkgroup,maxComputeWorkgroupSizeX:r.limits.maxComputeWorkgroupSizeX,maxComputeWorkgroupSizeY:r.limits.maxComputeWorkgroupSizeY,maxComputeWorkgroupSizeZ:r.limits.maxComputeWorkgroupSizeZ},requiredFeatures:o};r.features.has("chromium-experimental-timestamp-query-inside-passes")?o.push("chromium-experimental-timestamp-query-inside-passes"):r.features.has("timestamp-query")&&o.push("timestamp-query"),r.features.has("shader-f16")&&o.push("shader-f16"),this.device=await r.requestDevice(i),this.adapterInfo=new Yo(await r.requestAdapterInfo()),this.gpuDataManager=ba(this),this.programManager=new Hn(this),this.kernels=new Map,this.kernelPersistentData=new Map,this.kernelCustomData=new Map,pa(t.logLevel,!!t.debug),this.device.onuncapturederror=u=>{u.error instanceof GPUValidationError&&console.error(`An uncaught WebGPU validation error was raised: ${u.error.message}`)},Object.defineProperty(this.env.webgpu,"device",{value:this.device,writable:!1,enumerable:!0,configurable:!1}),Object.defineProperty(this.env.webgpu,"adapter",{value:r,writable:!1,enumerable:!0,configurable:!1}),this.setQueryType()}dispose(){typeof this.querySet<"u"&&this.querySet.destroy(),this.gpuDataManager.dispose()}getCommandEncoder(){return this.commandEncoder||(this.commandEncoder=this.device.createCommandEncoder()),this.commandEncoder}getComputePassEncoder(){if(!this.computePassEncoder){let t=this.getCommandEncoder(),r={};this.queryType==="at-passes"&&(r.timestampWrites={querySet:this.querySet,beginningOfPassWriteIndex:this.pendingDispatchNumber*2,endOfPassWriteIndex:this.pendingDispatchNumber*2+1}),this.computePassEncoder=t.beginComputePass(r)}return this.computePassEncoder}endComputePass(){this.computePassEncoder&&(this.computePassEncoder.end(),this.computePassEncoder=null)}flush(){if(!this.commandEncoder)return;rr(),this.endComputePass();let t;this.queryType!=="none"&&(this.commandEncoder.resolveQuerySet(this.querySet,0,this.pendingDispatchNumber*2,this.queryResolveBuffer,0),t=this.device.createBuffer({size:this.pendingDispatchNumber*2*8,usage:GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST}),this.pendingQueries.set(t,this.pendingKernels),this.pendingKernels=[],this.commandEncoder.copyBufferToBuffer(this.queryResolveBuffer,0,t,0,this.pendingDispatchNumber*2*8)),this.device.queue.submit([this.commandEncoder.finish()]),this.gpuDataManager.refreshPendingBuffers(),this.commandEncoder=null,this.pendingDispatchNumber=0,this.queryType!=="none"&&t.mapAsync(GPUMapMode.READ).then(()=>{let r=new BigUint64Array(t.getMappedRange()),o=this.pendingQueries.get(t);for(let i=0;i<r.length/2;i++){let u=o[i],a=u.kernelId,c=this.kernels.get(a),p=c.kernelType,h=c.kernelName,d=u.programName,y=u.inputTensorViews,w=u.outputTensorViews,_=r[i*2],v=r[i*2+1];typeof this.queryTimeBase>"u"&&(this.queryTimeBase=_);let S=Number(_-this.queryTimeBase),A=Number(v-this.queryTimeBase);if(!Number.isSafeInteger(S)||!Number.isSafeInteger(A))throw new RangeError("incorrect timestamp range");if(this.env.webgpu.profiling?.ondata)this.env.webgpu.profiling.ondata({version:1,inputsMetadata:y.map(I=>({dims:I.dims,dataType:Gt(I.dataType)})),outputsMetadata:w.map(I=>({dims:I.dims,dataType:Gt(I.dataType)})),kernelId:a,kernelType:p,kernelName:h,programName:d,startTime:S,endTime:A});else{let I="";y.forEach((E,P)=>{I+=`input[${P}]: [${E.dims}] | ${Gt(E.dataType)}, `});let x="";w.forEach((E,P)=>{x+=`output[${P}]: [${E.dims}] | ${Gt(E.dataType)}, `}),console.log(`[profiling] kernel "${a}|${p}|${h}|${d}" ${I}${x}execution time: ${A-S} ns`)}wo("GPU",`${d}::${_}::${v}`)}t.unmap(),this.pendingQueries.delete(t)}),Ht()}run(t,r,o,i,u,a){rr(t.name);let c=[];for(let E=0;E<r.length;++E){let P=r[E].data;if(P===0)continue;let O=this.gpuDataManager.get(P);if(!O)throw new Error(`no GPU data for input: ${P}`);c.push(O)}let{outputs:p,dispatchGroup:h,programUniforms:d}=t.getRunData(r),y=o.length===0?p.map((E,P)=>P):o;if(y.length!==p.length)throw new Error(`Output size ${y.length} must be equal to ${p.length}.`);let w=[],_=[];for(let E=0;E<p.length;++E){if(!Number.isInteger(y[E])||y[E]<-3||y[E]>=a)throw new Error(`Invalid output index: ${y[E]}`);if(y[E]===-3)continue;let P=y[E]===-1,O=y[E]===-2,R=P||O?u(p[E].dataType,p[E].dims):i(y[E],p[E].dataType,p[E].dims);if(w.push(R),R.data===0)continue;let L=this.gpuDataManager.get(R.data);if(!L)throw new Error(`no GPU data for output: ${R.data}`);if(P&&this.temporaryData.push(L),O){let N=this.kernelPersistentData.get(this.currentKernelId);N||(N=[],this.kernelPersistentData.set(this.currentKernelId,N)),N.push(L)}_.push(L)}if(c.length!==r.length||_.length!==w.length){if(_.length===0)return Ht(t.name),w;throw new Error(`Program ${t.name} has zero-sized tensor(s) in inputs or outputs. This is not supported now.`)}let v;if(d){let E=0,P=[];d.forEach(N=>{let K=typeof N.data=="number"?[N.data]:N.data;if(K.length===0)return;let Q=N.type===10?2:4,he,W;N.type===10?(W=K.length>4?16:K.length>2?8:K.length*Q,he=K.length>4?16:Q*K.length):(W=K.length<=2?K.length*Q:16,he=16),E=Math.ceil(E/W)*W,P.push(E);let se=N.type===10?8:4;E+=K.length>4?Math.ceil(K.length/se)*he:K.length*Q});let O=16;E=Math.ceil(E/O)*O;let R=new ArrayBuffer(E);d.forEach((N,K)=>{let Q=P[K],he=typeof N.data=="number"?[N.data]:N.data;if(N.type===6)new Int32Array(R,Q,he.length).set(he);else if(N.type===12)new Uint32Array(R,Q,he.length).set(he);else if(N.type===10)new Uint16Array(R,Q,he.length).set(he);else if(N.type===1)new Float32Array(R,Q,he.length).set(he);else throw new Error(`Unsupported uniform type: ${Gt(N.type)}`)});let L=this.gpuDataManager.create(E,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.device.queue.writeBuffer(L.buffer,0,R,0,E),this.gpuDataManager.release(L.id),v={offset:0,size:E,buffer:L.buffer}}let S=this.programManager.normalizeDispatchGroupSize(h),A=S[1]===1&&S[2]===1,I=nm(t,r,A),x=this.programManager.getArtifact(I);if(x||(x=this.programManager.build(t,S),this.programManager.setArtifact(I,x),Ve("info",()=>`[artifact] key: ${I}, programName: ${t.name}`)),d&&x.uniformVariablesInfo){if(d.length!==x.uniformVariablesInfo.length)throw new Error(`Uniform variables count mismatch: expect ${x.uniformVariablesInfo.length}, got ${d.length} in program "${x.programInfo.name}".`);for(let E=0;E<d.length;E++){let P=d[E],O=P.type,R=typeof P.data=="number"?1:P.data.length,[L,N]=x.uniformVariablesInfo[E];if(O!==L||R!==N)throw new Error(`Uniform variable ${E} mismatch: expect type ${L} with size ${N}, got type ${O} with size ${R} in program "${x.programInfo.name}".`)}}if(Ve("info",()=>`[ProgramManager] run "${t.name}" (key=${I}) with ${S[0]}x${S[1]}x${S[2]}`),this.queryType!=="none"||this.sessionStatus==="capturing"){let E={kernelId:this.currentKernelId,programName:x.programInfo.name,inputTensorViews:r,outputTensorViews:w};this.pendingKernels.push(E),this.sessionStatus==="capturing"&&this.capturedPendingKernels.get(this.currentSessionId).push(E)}return this.programManager.run(x,c,_,S,v),Ht(t.name),w}upload(t,r){this.gpuDataManager.upload(t,r)}memcpy(t,r){this.gpuDataManager.memcpy(t,r)}async download(t,r){await this.gpuDataManager.download(t,r)}alloc(t){return this.gpuDataManager.create(t).id}free(t){return this.gpuDataManager.release(t)}createKernel(t,r,o,i){let u=Md.get(t);if(!u)throw new Error(`kernel not implemented: ${t}`);let a={kernelType:t,kernelName:i,kernelEntry:u[0],attributes:[u[1],o]};this.kernels.set(r,a)}releaseKernel(t){let r=this.kernelPersistentData.get(t);if(r){for(let o of r)this.gpuDataManager.release(o.id);this.kernelPersistentData.delete(t)}this.kernelCustomData.delete(t),this.kernels.delete(t)}computeKernel(t,r,o){let i=this.kernels.get(t);if(!i)throw new Error(`kernel not created: ${t}`);let u=i.kernelType,a=i.kernelName,c=i.kernelEntry,p=i.attributes;if(this.currentKernelId!==null)throw new Error(`kernel "[${u}] ${a}" is not allowed to be called recursively`);this.currentKernelId=t,p[0]&&(p[1]=p[0](p[1]),p[0]=void 0),Ve("info",()=>`[WebGPU] Start to run kernel "[${u}] ${a}"...`);let h=this.env.debug;this.temporaryData=[];try{return h&&this.device.pushErrorScope("validation"),c(r,p[1]),0}catch(d){return o.push(Promise.resolve(`[WebGPU] Kernel "[${u}] ${a}" failed. ${d}`)),1}finally{h&&o.push(this.device.popErrorScope().then(d=>d?`GPU validation error for kernel "[${u}] ${a}": ${d.message}`:null));for(let d of this.temporaryData)this.gpuDataManager.release(d.id);this.temporaryData=[],this.currentKernelId=null}}registerBuffer(t,r,o,i){let u=this.sessionExternalDataMapping.get(t);u||(u=new Map,this.sessionExternalDataMapping.set(t,u));let a=u.get(r),c=this.gpuDataManager.registerExternalBuffer(o,i,a?.[1]);return u.set(r,[c,o]),c}unregisterBuffers(t){let r=this.sessionExternalDataMapping.get(t);r&&(r.forEach(o=>this.gpuDataManager.unregisterExternalBuffer(o[1])),this.sessionExternalDataMapping.delete(t))}getBuffer(t){let r=this.gpuDataManager.get(t);if(!r)throw new Error(`no GPU data for buffer: ${t}`);return r.buffer}createDownloader(t,r,o){return async()=>{let i=await xo(this,t,r);return ma(i.buffer,o)}}writeTimestamp(t){this.queryType==="inside-passes"&&this.computePassEncoder.writeTimestamp(this.querySet,t)}setQueryType(){this.queryType="none",(this.env.webgpu.profiling?.mode==="default"||(typeof this.env.trace>"u"?this.env.wasm.trace:this.env.trace))&&(this.device.features.has("chromium-experimental-timestamp-query-inside-passes")?this.queryType="inside-passes":this.device.features.has("timestamp-query")&&(this.queryType="at-passes"),this.queryType!=="none"&&typeof this.querySet>"u"&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:this.maxDispatchNumber*2}),this.queryResolveBuffer=this.device.createBuffer({size:this.maxDispatchNumber*2*8,usage:GPUBufferUsage.COPY_SRC|GPUBufferUsage.QUERY_RESOLVE})))}captureBegin(){Ve("info","captureBegin"),this.capturedCommandList.get(this.currentSessionId)||this.capturedCommandList.set(this.currentSessionId,[]),this.capturedPendingKernels.get(this.currentSessionId)||this.capturedPendingKernels.set(this.currentSessionId,[]),this.flush(),this.sessionStatus="capturing"}captureEnd(){Ve("info","captureEnd"),this.flush(),this.sessionStatus="default"}replay(){Ve("info","replay"),this.sessionStatus="replaying";let t=this.capturedCommandList.get(this.currentSessionId),r=this.capturedPendingKernels.get(this.currentSessionId),o=t.length;this.pendingKernels=[];for(let i=0;i<o;i++){let u=this.getComputePassEncoder(),a=t[i];this.writeTimestamp(this.pendingDispatchNumber*2),u.setPipeline(a.computePipeline),u.setBindGroup(0,a.bindGroup),u.dispatchWorkgroups(...a.dispatchGroup),this.writeTimestamp(this.pendingDispatchNumber*2+1),this.pendingDispatchNumber++,this.queryType!=="none"&&this.pendingKernels.push(r[i]),(this.pendingDispatchNumber>=this.maxDispatchNumber||this.queryType==="at-passes")&&this.endComputePass(),this.pendingDispatchNumber>=this.maxDispatchNumber&&this.flush()}this.flush(),this.sessionStatus="default"}onReleaseSession(t){this.unregisterBuffers(t),this.capturedCommandList.has(t)&&this.capturedCommandList.delete(t),this.capturedPendingKernels.has(t)&&this.capturedPendingKernels.delete(t),this.gpuDataManager.onReleaseSession(t)}onRunStart(t){this.currentSessionId=t,this.setQueryType()}}});var Nd={};gn(Nd,{init:()=>om});var qr,Zo,om,Gd=Y(()=>{"use strict";ye();Wd();Lt();Se();qr=class e{constructor(t,r,o,i){this.module=t;this.dataType=r;this.data=o;this.dims=i}getFloat32Array(){if(this.dataType!==1)throw new Error("Invalid data type");let t=M.size(this.dims);return t===0?new Float32Array:new Float32Array(this.module.HEAP8.buffer,this.data,t)}getBigInt64Array(){if(this.dataType!==7)throw new Error("Invalid data type");let t=M.size(this.dims);return t===0?new BigInt64Array:new BigInt64Array(this.module.HEAP8.buffer,this.data,t)}getInt32Array(){if(this.dataType!==6)throw new Error("Invalid data type");let t=M.size(this.dims);return t===0?new Int32Array:new Int32Array(this.module.HEAP8.buffer,this.data,t)}reshape(t){if(M.size(t)!==M.size(this.dims))throw new Error("Invalid new shape");return new e(this.module,this.dataType,this.data,t)}},Zo=class{constructor(t,r,o){this.module=t;this.backend=r;this.customDataOffset=0;this.customDataSize=0;this.adapterInfo=r.adapterInfo;let i=t.HEAPU32,u=o>>>2;this.opKernelContext=i[u++];let a=i[u++];this.outputCount=i[u++],this.customDataOffset=i[u++],this.customDataSize=i[u++];let c=[];for(let p=0;p<a;p++){let h=i[u++],d=i[u++],y=i[u++],w=[];for(let _=0;_<y;_++)w.push(i[u++]);c.push(new qr(t,h,d,w))}this.inputs=c}get kernelCustomData(){return this.backend.currentKernelCustomData}get customDataBuffer(){return this.module.HEAPU8.subarray(this.customDataOffset,this.customDataOffset+this.customDataSize)}getMaxComputeWorkgroupSizes(){return[this.backend.device.limits.maxComputeWorkgroupSizeX,this.backend.device.limits.maxComputeWorkgroupSizeY,this.backend.device.limits.maxComputeWorkgroupSizeZ]}getMaxComputeWorkgroupStoragesize(){return this.backend.device.limits.maxComputeWorkgroupStorageSize}compute(t,r){let o=r?.inputs?.map(c=>typeof c=="number"?this.inputs[c]:c)??this.inputs,i=r?.outputs??[],u=(c,p,h)=>new qr(this.module,p,this.output(c,h),h),a=(c,p)=>{let h=tr(c);if(!h)throw new Error(`Unsupported data type: ${c}`);let d=h*M.size(p),y=d>0?this.backend.gpuDataManager.create(d).id:0;return new qr(this.module,c,y,p)};return this.backend.run(t,o,i,u,a,this.outputCount)}output(t,r){let o=this.module.stackSave();try{let i=this.module.stackAlloc((1+r.length)*4),u=i>>2;this.module.HEAPU32[u++]=r.length;for(let a=0;a<r.length;a++)this.module.HEAPU32[u++]=r[a];return this.module._JsepOutput(this.opKernelContext,t,i)}catch(i){throw new Error(`Failed to generate kernel\'s output[${t}] with dims [${r}]. If you are running with pre-allocated output, please make sure the output type/dims are correct. Error: ${i}`)}finally{this.module.stackRestore(o)}}},om=async(e,t,r,o)=>{let i=t.jsepInit;if(!i)throw new Error("Failed to initialize JSEP. The WebAssembly module is not built with JSEP support.");if(e==="webgpu"){let u=new Ln;await u.initialize(r,o),i("webgpu",[u,a=>u.alloc(a),a=>u.free(a),(a,c,p,h=!1)=>{if(h)Ve("verbose",()=>`[WebGPU] jsepCopyGpuToGpu: src=${a}, dst=${c}, size=${p}`),u.memcpy(a,c);else{Ve("verbose",()=>`[WebGPU] jsepCopyCpuToGpu: dataOffset=${a}, gpuDataId=${c}, size=${p}`);let d=t.HEAPU8.subarray(a>>>0,(a>>>0)+p);u.upload(c,d)}},async(a,c,p)=>{Ve("verbose",()=>`[WebGPU] jsepCopyGpuToCpu: gpuDataId=${a}, dataOffset=${c}, size=${p}`),await u.download(a,()=>t.HEAPU8.subarray(c>>>0,(c>>>0)+p))},(a,c,p)=>u.createKernel(a,c,p,t.UTF8ToString(t._JsepGetNodeName(c))),a=>u.releaseKernel(a),(a,c,p,h)=>{Ve("verbose",()=>`[WebGPU] jsepRun: sessionHandle=${p}, kernel=${a}, contextDataOffset=${c}`);let d=new Zo(t,u,c);return u.computeKernel(a,d,h)},()=>u.captureBegin(),()=>u.captureEnd(),()=>u.replay()])}else i("webnn")}});var Vi;Vi=Ei();var pl=zi(),fo,ho=!1,yn=!1,Ui=!1,ml=e=>{if(e===1)return!1;if(typeof SharedArrayBuffer>"u")return typeof self<"u"&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+e+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),!1;typeof process<"u"&&process.versions&&process.versions.node&&console.warn("env.wasm.numThreads is set to "+e+", however, currently onnxruntime-web does not support multi-threads in Node.js. Please consider using onnxruntime-node for performance critical scenarios.");try{return typeof MessageChannel<"u"&&new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)),WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11]))}catch{return!1}},fl=()=>{try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,10,30,1,28,0,65,0,253,15,253,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,253,186,1,26,11]))}catch{return!1}},hl=(e,t)=>e?t?"ort-wasm-simd-threaded.wasm":"ort-wasm-simd.wasm":t?"ort-wasm-threaded.wasm":"ort-wasm.wasm",Wi=async e=>{if(ho)return Promise.resolve();if(yn)throw new Error("multiple calls to \'initializeWebAssembly()\' detected.");if(Ui)throw new Error("previous call to \'initializeWebAssembly()\' failed.");yn=!0;let t=e.initTimeout,r=e.numThreads,o=e.simd,i=ml(r),u=o&&fl(),a=e.wasmPaths,c=typeof a=="string"?a:void 0,p=hl(u,i),h=typeof a=="object"?a[p]:void 0,d=!1,y=[];if(t>0&&y.push(new Promise(w=>{setTimeout(()=>{d=!0,w()},t)})),y.push(new Promise((w,_)=>{let v=i?pl:Vi,S={locateFile:(A,I)=>{if(i&&A.endsWith(".worker.js")&&typeof Blob<"u")return URL.createObjectURL(new Blob([Mi()],{type:"text/javascript"}));if(A.endsWith(".wasm")){if(h)return h;let x=c??I;return p==="ort-wasm-simd.wasm"?x+"ort-wasm-simd.jsep.wasm":p==="ort-wasm-simd-threaded.wasm"?x+"ort-wasm-simd-threaded.jsep.wasm":x+p}return I+A}};if(i)if(S.numThreads=r,typeof Blob>"u")S.mainScriptUrlOrBlob=(void 0)(__dirname,"ort-wasm-threaded.js");else{let A=`var ortWasmThreaded=${v.toString()};`;S.mainScriptUrlOrBlob=new Blob([A],{type:"text/javascript"})}v(S).then(A=>{yn=!1,ho=!0,fo=A,w()},A=>{yn=!1,Ui=!0,_(A)})})),await Promise.race(y),d)throw new Error(`WebAssembly backend initializing failed due to timeout: ${t}ms`)},Le=()=>{if(ho&&fo)return fo;throw new Error("WebAssembly is not initialized yet.")};var Fe=(e,t)=>{let r=Le(),o=r.lengthBytesUTF8(e)+1,i=r._malloc(o);return r.stringToUTF8(e,i,o),t.push(i),i},Nr=(e,t,r,o)=>{if(typeof e=="object"&&e!==null){if(r.has(e))throw new Error("Circular reference in options");r.add(e)}Object.entries(e).forEach(([i,u])=>{let a=t?t+i:i;if(typeof u=="object")Nr(u,a+".",r,o);else if(typeof u=="string"||typeof u=="number")o(a,u.toString());else if(typeof u=="boolean")o(a,u?"1":"0");else throw new Error(`Can\'t handle extra config type: ${typeof u}`)})},Ue=e=>{let t=Le(),r=t.stackSave();try{let o=t.stackAlloc(8);t._OrtGetLastError(o,o+4);let i=t.HEAP32[o/4],u=t.HEAPU32[o/4+1],a=u?t.UTF8ToString(u):"";throw new Error(`${e} ERROR_CODE: ${i}, ERROR_MESSAGE: ${a}`)}finally{t.stackRestore(r)}};var Ni=e=>{let t=Le(),r=0,o=[],i=e||{};try{if(e?.logSeverityLevel===void 0)i.logSeverityLevel=2;else if(typeof e.logSeverityLevel!="number"||!Number.isInteger(e.logSeverityLevel)||e.logSeverityLevel<0||e.logSeverityLevel>4)throw new Error(`log serverity level is not valid: ${e.logSeverityLevel}`);if(e?.logVerbosityLevel===void 0)i.logVerbosityLevel=0;else if(typeof e.logVerbosityLevel!="number"||!Number.isInteger(e.logVerbosityLevel))throw new Error(`log verbosity level is not valid: ${e.logVerbosityLevel}`);e?.terminate===void 0&&(i.terminate=!1);let u=0;return e?.tag!==void 0&&(u=Fe(e.tag,o)),r=t._OrtCreateRunOptions(i.logSeverityLevel,i.logVerbosityLevel,!!i.terminate,u),r===0&&Ue("Can\'t create run options."),e?.extra!==void 0&&Nr(e.extra,"",new WeakSet,(a,c)=>{let p=Fe(a,o),h=Fe(c,o);t._OrtAddRunConfigEntry(r,p,h)!==0&&Ue(`Can\'t set a run config entry: ${a} - ${c}.`)}),[r,o]}catch(u){throw r!==0&&t._OrtReleaseRunOptions(r),o.forEach(a=>t._free(a)),u}};var gl=e=>{switch(e){case"disabled":return 0;case"basic":return 1;case"extended":return 2;case"all":return 99;default:throw new Error(`unsupported graph optimization level: ${e}`)}},yl=e=>{switch(e){case"sequential":return 0;case"parallel":return 1;default:throw new Error(`unsupported execution mode: ${e}`)}},bl=e=>{e.extra||(e.extra={}),e.extra.session||(e.extra.session={});let t=e.extra.session;t.use_ort_model_bytes_directly||(t.use_ort_model_bytes_directly="1"),e.executionProviders&&e.executionProviders.some(r=>(typeof r=="string"?r:r.name)==="webgpu")&&(e.enableMemPattern=!1)},wl=(e,t,r)=>{for(let o of t){let i=typeof o=="string"?o:o.name;switch(i){case"webnn":if(i="WEBNN",typeof o!="string"){let a=o;if(a?.deviceType){let c=Fe("deviceType",r),p=Fe(a.deviceType,r);Le()._OrtAddSessionConfigEntry(e,c,p)!==0&&Ue(`Can\'t set a session config entry: \'deviceType\' - ${a.deviceType}.`)}if(a?.numThreads){let c=a.numThreads;(typeof c!="number"||!Number.isInteger(c)||c<0)&&(c=0);let p=Fe("numThreads",r),h=Fe(c.toString(),r);Le()._OrtAddSessionConfigEntry(e,p,h)!==0&&Ue(`Can\'t set a session config entry: \'numThreads\' - ${a.numThreads}.`)}if(a?.powerPreference){let c=Fe("powerPreference",r),p=Fe(a.powerPreference,r);Le()._OrtAddSessionConfigEntry(e,c,p)!==0&&Ue(`Can\'t set a session config entry: \'powerPreference\' - ${a.powerPreference}.`)}}break;case"webgpu":if(i="JS",typeof o!="string"){let a=o;if(a?.preferredLayout){if(a.preferredLayout!=="NCHW"&&a.preferredLayout!=="NHWC")throw new Error(`preferredLayout must be either \'NCHW\' or \'NHWC\': ${a.preferredLayout}`);let c=Fe("preferredLayout",r),p=Fe(a.preferredLayout,r);Le()._OrtAddSessionConfigEntry(e,c,p)!==0&&Ue(`Can\'t set a session config entry: \'preferredLayout\' - ${a.preferredLayout}.`)}}break;case"wasm":case"cpu":continue;default:throw new Error(`not supported execution provider: ${i}`)}let u=Fe(i,r);Le()._OrtAppendExecutionProvider(e,u)!==0&&Ue(`Can\'t append execution provider: ${i}.`)}},Gi=e=>{let t=Le(),r=0,o=[],i=e||{};bl(i);try{let u=gl(i.graphOptimizationLevel??"all"),a=yl(i.executionMode??"sequential"),c=typeof i.logId=="string"?Fe(i.logId,o):0,p=i.logSeverityLevel??2;if(!Number.isInteger(p)||p<0||p>4)throw new Error(`log serverity level is not valid: ${p}`);let h=i.logVerbosityLevel??0;if(!Number.isInteger(h)||h<0||h>4)throw new Error(`log verbosity level is not valid: ${h}`);let d=typeof i.optimizedModelFilePath=="string"?Fe(i.optimizedModelFilePath,o):0;if(r=t._OrtCreateSessionOptions(u,!!i.enableCpuMemArena,!!i.enableMemPattern,a,!!i.enableProfiling,0,c,p,h,d),r===0&&Ue("Can\'t create session options."),i.executionProviders&&wl(r,i.executionProviders,o),i.enableGraphCapture!==void 0){if(typeof i.enableGraphCapture!="boolean")throw new Error(`enableGraphCapture must be a boolean value: ${i.enableGraphCapture}`);let y=Fe("enableGraphCapture",o),w=Fe(i.enableGraphCapture.toString(),o);t._OrtAddSessionConfigEntry(r,y,w)!==0&&Ue(`Can\'t set a session config entry: \'enableGraphCapture\' - ${i.enableGraphCapture}.`)}if(i.freeDimensionOverrides)for(let[y,w]of Object.entries(i.freeDimensionOverrides)){if(typeof y!="string")throw new Error(`free dimension override name must be a string: ${y}`);if(typeof w!="number"||!Number.isInteger(w)||w<0)throw new Error(`free dimension override value must be a non-negative integer: ${w}`);let _=Fe(y,o);t._OrtAddFreeDimensionOverride(r,_,w)!==0&&Ue(`Can\'t set a free dimension override: ${y} - ${w}.`)}return i.extra!==void 0&&Nr(i.extra,"",new WeakSet,(y,w)=>{let _=Fe(y,o),v=Fe(w,o);t._OrtAddSessionConfigEntry(r,_,v)!==0&&Ue(`Can\'t set a session config entry: ${y} - ${w}.`)}),[r,o]}catch(u){throw r!==0&&t._OrtReleaseSessionOptions(r),o.forEach(a=>t._free(a)),u}};ye();var Li=async e=>{if(typeof e=="string")if(typeof process<"u"&&process.versions&&process.versions.node)try{return new Uint8Array(await(void 0)(e))}catch(t){if(t.code==="ERR_FS_FILE_TOO_LARGE"){let r=(void 0)(e),o=[];for await(let i of r)o.push(i);return new Uint8Array(Buffer.concat(o))}throw t}else{let t=await fetch(e);if(!t.ok)throw new Error(`failed to load external data file: ${e}`);let r=t.headers.get("Content-Length"),o=r?parseInt(r,10):0;if(o<1073741824)return new Uint8Array(await t.arrayBuffer());{if(!t.body)throw new Error(`failed to load external data file: ${e}, no response body.`);let i=t.body.getReader(),u;try{u=new ArrayBuffer(o)}catch(c){if(c instanceof RangeError){let p=Math.ceil(o/65536);u=new WebAssembly.Memory({initial:p,maximum:p}).buffer}else throw c}let a=0;for(;;){let{done:c,value:p}=await i.read();if(c)break;let h=p.byteLength;new Uint8Array(u,a,h).set(p),a+=h}return new Uint8Array(u,0,o)}}else return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)};var im=(e,t)=>{Le()._OrtInit(e,t)!==0&&Ue("Can\'t initialize onnxruntime.")},Ld=async e=>{im(e.wasm.numThreads,Gr(e.logLevel))},Fd=async(e,t)=>{{let r=(Gd(),wr(Nd)).init;if(t==="webgpu"){if(typeof navigator>"u"||!navigator.gpu)throw new Error("WebGPU is not supported in current environment");let o=e.webgpu.adapter;if(o){if(typeof o.limits!="object"||typeof o.features!="object"||typeof o.requestDevice!="function")throw new Error("Invalid GPU adapter set in `env.webgpu.adapter`. It must be a GPUAdapter object.")}else{let i=e.webgpu.powerPreference;if(i!==void 0&&i!=="low-power"&&i!=="high-performance")throw new Error(`Invalid powerPreference setting: "${i}"`);let u=e.webgpu.forceFallbackAdapter;if(u!==void 0&&typeof u!="boolean")throw new Error(`Invalid forceFallbackAdapter setting: "${u}"`);if(o=await navigator.gpu.requestAdapter({powerPreference:i,forceFallbackAdapter:u}),!o)throw new Error(\'Failed to get GPU adapter. You may need to enable flag "--enable-unsafe-webgpu" if you are using Chrome.\')}if(!e.wasm.simd)throw new Error("Not supported for WebGPU=ON and SIMD=OFF. Please set `env.wasm.simd` to true when using `webgpu` EP");await r("webgpu",Le(),e,o)}if(t==="webnn"){if(typeof navigator>"u"||!navigator.ml)throw new Error("WebNN is not supported in current environment");await r("webnn",Le(),e)}}},ar=new Map,am=e=>{let t=Le(),r=t.stackSave();try{let o=t.stackAlloc(8);return t._OrtGetInputOutputCount(e,o,o+4)!==0&&Ue("Can\'t get session input/output count."),[t.HEAP32[o/4],t.HEAP32[o/4+1]]}finally{t.stackRestore(r)}},Xo=e=>{let t=Le(),r=t._malloc(e.byteLength);if(r===0)throw new Error(`Can\'t create a session. failed to allocate a buffer of size ${e.byteLength}.`);return t.HEAPU8.set(e,r),[r,e.byteLength]},qd=async(e,t)=>{let r,o,i=Le();Array.isArray(e)?[r,o]=e:e.buffer===i.HEAPU8.buffer?[r,o]=[e.byteOffset,e.byteLength]:[r,o]=Xo(e);let u=0,a=0,c=0,p=[],h=[],d=[];try{if([a,p]=Gi(t),t?.externalData&&i.mountExternalData){let x=[];for(let E of t.externalData){let P=typeof E=="string"?E:E.path;x.push(Li(typeof E=="string"?E:E.data).then(O=>{i.mountExternalData(P,O)}))}await Promise.all(x)}u=await i._OrtCreateSession(r,o,a),u===0&&Ue("Can\'t create a session.");let[y,w]=am(u),_=!!t?.enableGraphCapture,v=[],S=[],A=[];for(let x=0;x<y;x++){let E=i._OrtGetInputName(u,x);E===0&&Ue("Can\'t get an input name."),h.push(E),v.push(i.UTF8ToString(E))}for(let x=0;x<w;x++){let E=i._OrtGetOutputName(u,x);E===0&&Ue("Can\'t get an output name."),d.push(E);let P=i.UTF8ToString(E);S.push(P);{if(_&&t?.preferredOutputLocation===void 0){A.push("gpu-buffer");continue}let O=typeof t?.preferredOutputLocation=="string"?t.preferredOutputLocation:t?.preferredOutputLocation?.[P]??"cpu";if(O!=="cpu"&&O!=="cpu-pinned"&&O!=="gpu-buffer")throw new Error(`Not supported preferred output location: ${O}.`);if(_&&O!=="gpu-buffer")throw new Error(`Not supported preferred output location: ${O}. Only \'gpu-buffer\' location is supported when enableGraphCapture is true.`);A.push(O)}}let I=null;return A.some(x=>x==="gpu-buffer")&&(c=i._OrtCreateBinding(u),c===0&&Ue("Can\'t create IO binding."),I={handle:c,outputPreferredLocations:A,outputPreferredLocationsEncoded:A.map(x=>yo(x))}),ar.set(u,[u,h,d,I,_,!1]),[u,v,S]}catch(y){throw h.forEach(w=>i._OrtFree(w)),d.forEach(w=>i._OrtFree(w)),c!==0&&i._OrtReleaseBinding(c),u!==0&&i._OrtReleaseSession(u),y}finally{i._free(r),a!==0&&i._OrtReleaseSessionOptions(a),p.forEach(y=>i._free(y)),i.unmountExternalData?.()}},jd=e=>{let t=Le(),r=ar.get(e);if(!r)throw new Error(`cannot release session. invalid session id: ${e}`);let[o,i,u,a,c]=r;a&&(c&&t._OrtClearBoundOutputs(a.handle),t._OrtReleaseBinding(a.handle)),t.jsepOnReleaseSession?.(e),i.forEach(p=>t._OrtFree(p)),u.forEach(p=>t._OrtFree(p)),t._OrtReleaseSession(o),ar.delete(e)},Hd=(e,t,r,o,i,u=!1)=>{if(!e){t.push(0);return}let a=Le(),c=e[0],p=e[1],h=e[3],d,y;if(c==="string"&&h==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");if(u&&h!=="gpu-buffer")throw new Error(`External buffer must be provided for input/output index ${i} when enableGraphCapture is true.`);if(h==="gpu-buffer"){let v=e[2].gpuBuffer,S=tr(go(c));y=p.reduce((I,x)=>I*x,1)*S;let A=a.jsepRegisterBuffer;if(!A)throw new Error(\'Tensor location "gpu-buffer" is not supported without using WebGPU.\');d=A(o,i,v,y)}else{let v=e[2];if(Array.isArray(v)){y=4*v.length,d=a._malloc(y),r.push(d);let S=d/4;for(let A=0;A<v.length;A++){if(typeof v[A]!="string")throw new TypeError(`tensor data at index ${A} is not a string`);a.HEAPU32[S++]=Fe(v[A],r)}}else y=v.byteLength,d=a._malloc(y),r.push(d),a.HEAPU8.set(new Uint8Array(v.buffer,v.byteOffset,y),d)}let w=a.stackSave(),_=a.stackAlloc(4*p.length);try{let v=_/4;p.forEach(A=>a.HEAP32[v++]=A);let S=a._OrtCreateTensor(go(c),d,y,_,p.length,yo(h));S===0&&Ue(`Can\'t create tensor for input/output. session=${o}, index=${i}.`),t.push(S)}finally{a.stackRestore(w)}},Kd=async(e,t,r,o,i,u)=>{let a=Le(),c=ar.get(e);if(!c)throw new Error(`cannot run inference. invalid session id: ${e}`);let p=c[0],h=c[1],d=c[2],y=c[3],w=c[4],_=c[5],v=t.length,S=o.length,A=0,I=[],x=[],E=[],P=[],O=a.stackSave(),R=a.stackAlloc(v*4),L=a.stackAlloc(v*4),N=a.stackAlloc(S*4),K=a.stackAlloc(S*4);try{[A,I]=Ni(u);for(let ee=0;ee<v;ee++)Hd(r[ee],x,P,e,t[ee],w);for(let ee=0;ee<S;ee++)Hd(i[ee],E,P,e,v+o[ee],w);let Q=R/4,he=L/4,W=N/4,se=K/4;for(let ee=0;ee<v;ee++)a.HEAPU32[Q++]=x[ee],a.HEAPU32[he++]=h[t[ee]];for(let ee=0;ee<S;ee++)a.HEAPU32[W++]=E[ee],a.HEAPU32[se++]=d[o[ee]];if(y&&!_){let{handle:ee,outputPreferredLocations:ae,outputPreferredLocationsEncoded:Ae}=y;if(h.length!==v)throw new Error(`input count from feeds (${v}) is expected to be always equal to model\'s input count (${h.length}).`);for(let me=0;me<v;me++){let ie=t[me];await a._OrtBindInput(ee,h[ie],x[me])!==0&&Ue(`Can\'t bind input[${me}] for session=${e}.`)}for(let me=0;me<S;me++){let ie=o[me];i[me]?.[3]?a._OrtBindOutput(ee,d[ie],E[me],0)!==0&&Ue(`Can\'t bind pre-allocated output[${me}] for session=${e}.`):a._OrtBindOutput(ee,d[ie],0,Ae[ie])!==0&&Ue(`Can\'t bind output[${me}] to ${ae[me]} for session=${e}.`)}ar.set(e,[p,h,d,y,w,!0])}a.jsepOnRunStart?.(p);let Ce;y?Ce=await a._OrtRunWithBinding(p,y.handle,S,N,A):Ce=await a._OrtRun(p,L,R,v,K,S,N,A),Ce!==0&&Ue("failed to call OrtRun().");let We=[];for(let ee=0;ee<S;ee++){let ae=a.HEAPU32[N/4+ee];if(ae===E[ee]){We.push(i[ee]);continue}let Ae=a.stackSave(),me=a.stackAlloc(4*4),ie=!1,ue,le=0;try{a._OrtGetTensorData(ae,me,me+4,me+8,me+12)!==0&&Ue(`Can\'t access output tensor data on index ${ee}.`);let G=me/4,ne=a.HEAPU32[G++];le=a.HEAPU32[G++];let xe=a.HEAPU32[G++],Ke=a.HEAPU32[G++],Be=[];for(let Ne=0;Ne<Ke;Ne++)Be.push(a.HEAPU32[xe/4+Ne]);a._OrtFree(xe);let Ge=Be.reduce((Ne,Ye)=>Ne*Ye,1);ue=Gt(ne);let Ut=y?.outputPreferredLocations[o[ee]];if(ue==="string"){if(Ut==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");let Ne=[],Ye=le/4;for(let mt=0;mt<Ge;mt++){let Rt=a.HEAPU32[Ye++],qt=mt===Ge-1?void 0:a.HEAPU32[Ye]-Rt;Ne.push(a.UTF8ToString(Rt,qt))}We.push([ue,Be,Ne,"cpu"])}else if(Ut==="gpu-buffer"&&Ge>0){let Ne=a.jsepGetBuffer;if(!Ne)throw new Error(\'preferredLocation "gpu-buffer" is not supported without using WebGPU.\');let Ye=Ne(le),mt=tr(ne);if(mt===void 0||!Hi(ue))throw new Error(`Unsupported data type: ${ue}`);ie=!0,We.push([ue,Be,{gpuBuffer:Ye,download:a.jsepCreateDownloader(Ye,Ge*mt,ue),dispose:()=>{a._OrtReleaseTensor(ae)}},"gpu-buffer"])}else{let Ne=bn(ue),Ye=new Ne(Ge);new Uint8Array(Ye.buffer,Ye.byteOffset,Ye.byteLength).set(a.HEAPU8.subarray(le,le+Ye.byteLength)),We.push([ue,Be,Ye,"cpu"])}}finally{a.stackRestore(Ae),ue==="string"&&le&&a._free(le),ie||a._OrtReleaseTensor(ae)}}return y&&!w&&(a._OrtClearBoundOutputs(y.handle),ar.set(e,[p,h,d,y,w,!1])),We}finally{a.stackRestore(O),x.forEach(Q=>a._OrtReleaseTensor(Q)),E.forEach(Q=>a._OrtReleaseTensor(Q)),P.forEach(Q=>a._free(Q)),A!==0&&a._OrtReleaseRunOptions(A),I.forEach(Q=>a._free(Q))}},Yd=e=>{let t=Le(),r=ar.get(e);if(!r)throw new Error("invalid session id");let o=r[0],i=t._OrtEndProfiling(o);i===0&&Ue("Can\'t get an profile file name."),t._OrtFree(i)},Zd=e=>{let t=[];for(let r of e){let o=r[2];!Array.isArray(o)&&"buffer"in o&&t.push(o.buffer)}return t};self.onmessage=e=>{let{type:t,in:r}=e.data;try{switch(t){case"init-wasm":Wi(r.wasm).then(()=>{Ld(r).then(()=>{postMessage({type:t})},o=>{postMessage({type:t,err:o})})},o=>{postMessage({type:t,err:o})});break;case"init-ep":{let{epName:o,env:i}=r;Fd(i,o).then(()=>{postMessage({type:t})},u=>{postMessage({type:t,err:u})});break}case"copy-from":{let{buffer:o}=r,i=Xo(o);postMessage({type:t,out:i});break}case"create":{let{model:o,options:i}=r;qd(o,i).then(u=>{postMessage({type:t,out:u})},u=>{postMessage({type:t,err:u})});break}case"release":jd(r),postMessage({type:t});break;case"run":{let{sessionId:o,inputIndices:i,inputs:u,outputIndices:a,options:c}=r;Kd(o,i,u,a,new Array(a.length).fill(null),c).then(p=>{p.some(h=>h[3]!=="cpu")?postMessage({type:t,err:"Proxy does not support non-cpu tensor location."}):postMessage({type:t,out:p},Zd([...u,...p]))},p=>{postMessage({type:t,err:p})});break}case"end-profiling":Yd(r),postMessage({type:t});break;default:}}catch(o){postMessage({type:t,err:o})}};})();\n/**\n * @license\n * Copyright 2021 Google LLC. All Rights Reserved.\n * Licensed under the Apache License, Version 2.0 (the "License");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n * http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an "AS IS" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n * =============================================================================\n */\n/**\n * @license\n * Copyright 2020 Google LLC. All Rights Reserved.\n * Licensed under the Apache License, Version 2.0 (the "License");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n * http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an "AS IS" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n * =============================================================================\n */\n/**\n * @license\n * Copyright 2019 Google LLC. All Rights Reserved.\n * Licensed under the Apache License, Version 2.0 (the "License");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n * http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an "AS IS" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n * =============================================================================\n */\n'}),Tr,Dt,bn,ra,na,eo,wi,Lr,Ur,Jc,aa,qm,Km,Ym,Xm,Qm,Jm,Zm,eg=Z(()=>{ar(),Dy(),va(),Tr=()=>!!Ue.wasm.proxy&&typeof document<"u",bn=!1,ra=!1,na=!1,wi=new Map,Lr=(t,e)=>{let r=wi.get(t);r?r.push(e):wi.set(t,[e])},Ur=()=>{if(bn||!ra||na||!Dt)throw new Error("worker not ready")},Jc=t=>{switch(t.data.type){case"init-wasm":bn=!1,t.data.err?(na=!0,eo[1](t.data.err)):(ra=!0,eo[0]());break;case"init-ep":case"copy-from":case"create":case"release":case"run":case"end-profiling":{let e=wi.get(t.data.type);t.data.err?e.shift()[1](t.data.err):e.shift()[0](t.data.out);break}}},aa=typeof document<"u"?document?.currentScript?.src:void 0,qm=async()=>{if(!ra){if(bn)throw new Error("multiple calls to 'initWasm()' detected.");if(na)throw new Error("previous call to 'initWasm()' failed.");if(bn=!0,Tr())return Ue.wasm.wasmPaths===void 0&&aa&&aa.indexOf("blob:")!==0&&(Ue.wasm.wasmPaths=aa.substr(0,+aa.lastIndexOf("/")+1)),new Promise((t,e)=>{Dt?.terminate();let r=URL.createObjectURL(new Blob([Ny()],{type:"text/javascript"}));Dt=new Worker(r,{name:"ort-wasm-proxy-worker"}),Dt.onerror=a=>e(a),Dt.onmessage=Jc,URL.revokeObjectURL(r),eo=[t,e];let n={type:"init-wasm",in:Ue};Dt.postMessage(n)});try{await ih(Ue.wasm),await Lm(Ue),ra=!0}catch(t){throw na=!0,t}finally{bn=!1}}},Km=async t=>{if(Tr())return Ur(),new Promise((e,r)=>{Lr("init-ep",[e,r]);let n={type:"init-ep",in:{epName:t,env:Ue}};Dt.postMessage(n)});await Um(Ue,t)},Ym=async t=>Tr()?(Ur(),new Promise((e,r)=>{Lr("copy-from",[e,r]);let n={type:"copy-from",in:{buffer:t}};Dt.postMessage(n,[t.buffer])})):Co(t),Xm=async(t,e)=>{if(Tr()){if(e?.preferredOutputLocation)throw new Error('session option "preferredOutputLocation" is not supported for proxy.');return Ur(),new Promise((r,n)=>{Lr("create",[r,n]);let a={type:"create",in:{model:t,options:{...e}}},i=[];t instanceof Uint8Array&&i.push(t.buffer),Dt.postMessage(a,i)})}else return Wm(t,e)},Qm=async t=>{if(Tr())return Ur(),new Promise((e,r)=>{Lr("release",[e,r]);let n={type:"release",in:t};Dt.postMessage(n)});Vm(t)},Jm=async(t,e,r,n,a,i)=>{if(Tr()){if(r.some(s=>s[3]!=="cpu"))throw new Error("input tensor on GPU is not supported for proxy.");if(a.some(s=>s))throw new Error("pre-allocated output tensor is not supported for proxy.");return Ur(),new Promise((s,o)=>{Lr("run",[s,o]);let u=r,l={type:"run",in:{sessionId:t,inputIndices:e,inputs:u,outputIndices:n,options:i}};Dt.postMessage(l,jm(u))})}else return Gm(t,e,r,n,a,i)},Zm=async t=>{if(Tr())return Ur(),new Promise((e,r)=>{Lr("end-profiling",[e,r]);let n={type:"end-profiling",in:t};Dt.postMessage(n)});Hm(t)}}),to,Zc,tg,Fy=Z(()=>{ar(),eg(),$e(),uh(),to=(t,e)=>{switch(t.location){case"cpu":return[t.type,t.dims,t.data,"cpu"];case"gpu-buffer":return[t.type,t.dims,{gpuBuffer:t.gpuBuffer},"gpu-buffer"];default:throw new Error(`invalid data location: ${t.location} for ${e()}`)}},Zc=t=>{switch(t[3]){case"cpu":return new vt(t[0],t[2],t[1]);case"gpu-buffer":{let e=t[0];if(!Ho(e))throw new Error(`not supported data type: ${e} for deserializing GPU tensor`);let{gpuBuffer:r,download:n,dispose:a}=t[2];return vt.fromGpuBuffer(r,{dataType:e,dims:t[1],download:n,dispose:a})}default:throw new Error(`invalid data location: ${t[3]}`)}},tg=class{async fetchModelAndCopyToWasmMemory(t){return Ym(await Ni(t))}async loadModel(t,e){nr();let r;typeof t=="string"?typeof process<"u"&&process.versions&&process.versions.node?r=await Ni(t):r=await this.fetchModelAndCopyToWasmMemory(t):r=t,[this.sessionId,this.inputNames,this.outputNames]=await Xm(r,e),Lt()}async dispose(){return Qm(this.sessionId)}async run(t,e,r){nr();let n=[],a=[];Object.entries(t).forEach(f=>{let m=f[0],c=f[1],y=this.inputNames.indexOf(m);if(y===-1)throw new Error(`invalid input '${m}'`);n.push(c),a.push(y)});let i=[],s=[];Object.entries(e).forEach(f=>{let m=f[0],c=f[1],y=this.outputNames.indexOf(m);if(y===-1)throw new Error(`invalid output '${m}'`);i.push(c),s.push(y)});let o=n.map((f,m)=>to(f,()=>`input "${this.inputNames[a[m]]}"`)),u=i.map((f,m)=>f?to(f,()=>`output "${this.outputNames[s[m]]}"`):null),l=await Jm(this.sessionId,a,o,s,u,r),p={};for(let f=0;f<l.length;f++)p[this.outputNames[s[f]]]=i[f]??Zc(l[f]);return Lt(),p}startProfiling(){}endProfiling(){Zm(this.sessionId)}}}),ep,rg,Ly=Z(()=>{ar(),eg(),Fy(),ep=()=>{if((typeof Ue.wasm.initTimeout!="number"||Ue.wasm.initTimeout<0)&&(Ue.wasm.initTimeout=0),typeof Ue.wasm.simd!="boolean"&&(Ue.wasm.simd=!0),typeof Ue.wasm.proxy!="boolean"&&(Ue.wasm.proxy=!1),typeof Ue.wasm.trace!="boolean"&&(Ue.wasm.trace=!1),typeof Ue.wasm.numThreads!="number"||!Number.isInteger(Ue.wasm.numThreads)||Ue.wasm.numThreads<=0){(typeof self<"u"&&!self.crossOriginIsolated||typeof process<"u"&&process.versions&&process.versions.node)&&(Ue.wasm.numThreads=1);let t=typeof navigator>"u"?(void 0)().length:navigator.hardwareConcurrency;Ue.wasm.numThreads=Math.min(4,Math.ceil((t||1)/2))}},rg=class{async init(t){ep(),await qm(),await Km(t)}async createInferenceSessionHandler(t,e){let r=new tg;return await r.loadModel(t,e),Promise.resolve(r)}}}),ng={};An(ng,{wasmBackend:()=>ag});var ag,Uy=Z(()=>{Ly(),ag=new rg});ar();ar();ar();var Wy="1.18.0",Vy=Xp;{let t=(Uy(),Qr(ng)).wasmBackend;Yr("webgpu",t,5),Yr("webnn",t,5),Yr("cpu",t,10),Yr("wasm",t,10)}Object.defineProperty(Ue.versions,"web",{value:Wy,enumerable:!0});/**
* @license
* Copyright 2021 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*//**
 * @license
 * Copyright 2020 Google LLC. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 * =============================================================================
 *//**
 * @license
 * Copyright 2019 Google LLC. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 * =============================================================================
 */const Gy=Object.freeze(Object.defineProperty({__proto__:null,get InferenceSession(){return Fo},get TRACE(){return ha},get TRACE_FUNC_BEGIN(){return nr},get TRACE_FUNC_END(){return Lt},get Tensor(){return vt},get TrainingSession(){return Lo},default:Vy,get env(){return Ue},get registerBackend(){return Yr}},Symbol.toStringTag,{value:"Module"})),Hy=(t,e)=>{const r=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);r.width=t.dims[3],r.height=t.dims[2];const n=r.getContext("2d");if(n!=null){let a,i;e?.tensorLayout!==void 0&&e.tensorLayout==="NHWC"?(a=t.dims[2],i=t.dims[3]):(a=t.dims[3],i=t.dims[2]);const s=e?.format!==void 0?e.format:"RGB",o=e?.norm;let u,l;o===void 0||o.mean===void 0?u=[255,255,255,255]:typeof o.mean=="number"?u=[o.mean,o.mean,o.mean,o.mean]:(u=[o.mean[0],o.mean[1],o.mean[2],0],o.mean[3]!==void 0&&(u[3]=o.mean[3])),o===void 0||o.bias===void 0?l=[0,0,0,0]:typeof o.bias=="number"?l=[o.bias,o.bias,o.bias,o.bias]:(l=[o.bias[0],o.bias[1],o.bias[2],0],o.bias[3]!==void 0&&(l[3]=o.bias[3]));const p=i*a;let f=0,m=p,c=p*2,y=-1;s==="RGBA"?(f=0,m=p,c=p*2,y=p*3):s==="RGB"?(f=0,m=p,c=p*2):s==="RBG"&&(f=0,c=p,m=p*2);for(let w=0;w<i;w++)for(let v=0;v<a;v++){const k=(t.data[f++]-l[0])*u[0],$=(t.data[m++]-l[1])*u[1],C=(t.data[c++]-l[2])*u[2],T=y===-1?255:(t.data[y++]-l[3])*u[3];n.fillStyle="rgba("+k+","+$+","+C+","+T+")",n.fillRect(v,w,1,1)}if("toDataURL"in r)return r.toDataURL();throw new Error("toDataURL is not supported")}else throw new Error("Can not access image data")},jy=(t,e)=>{const r=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d");let n;if(r!=null){let a,i,s;e?.tensorLayout!==void 0&&e.tensorLayout==="NHWC"?(a=t.dims[2],i=t.dims[1],s=t.dims[3]):(a=t.dims[3],i=t.dims[2],s=t.dims[1]);const o=e!==void 0&&e.format!==void 0?e.format:"RGB",u=e?.norm;let l,p;u===void 0||u.mean===void 0?l=[255,255,255,255]:typeof u.mean=="number"?l=[u.mean,u.mean,u.mean,u.mean]:(l=[u.mean[0],u.mean[1],u.mean[2],255],u.mean[3]!==void 0&&(l[3]=u.mean[3])),u===void 0||u.bias===void 0?p=[0,0,0,0]:typeof u.bias=="number"?p=[u.bias,u.bias,u.bias,u.bias]:(p=[u.bias[0],u.bias[1],u.bias[2],0],u.bias[3]!==void 0&&(p[3]=u.bias[3]));const f=i*a;if(e!==void 0&&(e.format!==void 0&&s===4&&e.format!=="RGBA"||s===3&&e.format!=="RGB"&&e.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");const m=4;let c=0,y=1,w=2,v=3,k=0,$=f,C=f*2,T=-1;o==="RGBA"?(k=0,$=f,C=f*2,T=f*3):o==="RGB"?(k=0,$=f,C=f*2):o==="RBG"&&(k=0,C=f,$=f*2),n=r.createImageData(a,i);for(let A=0;A<i*a;c+=m,y+=m,w+=m,v+=m,A++)n.data[c]=(t.data[k++]-p[0])*l[0],n.data[y]=(t.data[$++]-p[1])*l[1],n.data[w]=(t.data[C++]-p[2])*l[2],n.data[v]=T===-1?255:(t.data[T++]-p[3])*l[3]}else throw new Error("Can not access image data");return n},ro=(t,e)=>{if(t===void 0)throw new Error("Image buffer must be defined");if(e.height===void 0||e.width===void 0)throw new Error("Image height and width must be defined");if(e.tensorLayout==="NHWC")throw new Error("NHWC Tensor layout is not supported yet");const{height:r,width:n}=e,a=e.norm??{mean:255,bias:0};let i,s;typeof a.mean=="number"?i=[a.mean,a.mean,a.mean,a.mean]:i=[a.mean[0],a.mean[1],a.mean[2],a.mean[3]??255],typeof a.bias=="number"?s=[a.bias,a.bias,a.bias,a.bias]:s=[a.bias[0],a.bias[1],a.bias[2],a.bias[3]??0];const o=e.format!==void 0?e.format:"RGBA",u=e.tensorFormat!==void 0&&e.tensorFormat!==void 0?e.tensorFormat:"RGB",l=r*n,p=u==="RGBA"?new Float32Array(l*4):new Float32Array(l*3);let f=4,m=0,c=1,y=2,w=3,v=0,k=l,$=l*2,C=-1;o==="RGB"&&(f=3,m=0,c=1,y=2,w=-1),u==="RGBA"?C=l*3:u==="RBG"?(v=0,$=l,k=l*2):u==="BGR"&&($=0,k=l,v=l*2);for(let A=0;A<l;A++,m+=f,y+=f,c+=f,w+=f)p[v++]=(t[m]+s[0])/i[0],p[k++]=(t[c]+s[1])/i[1],p[$++]=(t[y]+s[2])/i[2],C!==-1&&w!==-1&&(p[C++]=(t[w]+s[3])/i[3]);return u==="RGBA"?new lr("float32",p,[1,4,r,n]):new lr("float32",p,[1,3,r,n])},qy=async(t,e)=>{const r=typeof HTMLImageElement<"u"&&t instanceof HTMLImageElement,n=typeof ImageData<"u"&&t instanceof ImageData,a=typeof ImageBitmap<"u"&&t instanceof ImageBitmap,i=typeof t=="string";let s,o=e??{};const u=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},l=p=>p instanceof HTMLCanvasElement||p instanceof OffscreenCanvas?p.getContext("2d"):null;if(r){const p=u();p.width=t.width,p.height=t.height;const f=l(p);if(f!=null){let m=t.height,c=t.width;if(e!==void 0&&e.resizedHeight!==void 0&&e.resizedWidth!==void 0&&(m=e.resizedHeight,c=e.resizedWidth),e!==void 0){if(o=e,e.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");o.tensorFormat="RGBA",o.height=m,o.width=c}else o.tensorFormat="RGBA",o.height=m,o.width=c;f.drawImage(t,0,0),s=f.getImageData(0,0,c,m).data}else throw new Error("Can not access image data")}else if(n){let p,f;if(e!==void 0&&e.resizedWidth!==void 0&&e.resizedHeight!==void 0?(p=e.resizedHeight,f=e.resizedWidth):(p=t.height,f=t.width),e!==void 0&&(o=e),o.format="RGBA",o.height=p,o.width=f,e!==void 0){const m=u();m.width=f,m.height=p;const c=l(m);if(c!=null)c.putImageData(t,0,0),s=c.getImageData(0,0,f,p).data;else throw new Error("Can not access image data")}else s=t.data}else if(a){if(e===void 0)throw new Error("Please provide image config with format for Imagebitmap");const p=u();p.width=t.width,p.height=t.height;const f=l(p);if(f!=null){const m=t.height,c=t.width;return f.drawImage(t,0,0,c,m),s=f.getImageData(0,0,c,m).data,o.height=m,o.width=c,ro(s,o)}else throw new Error("Can not access image data")}else{if(i)return new Promise((p,f)=>{const m=u(),c=l(m);if(!t||!c)return f();const y=new Image;y.crossOrigin="Anonymous",y.src=t,y.onload=()=>{m.width=y.width,m.height=y.height,c.drawImage(y,0,0,m.width,m.height);const w=c.getImageData(0,0,m.width,m.height);o.height=m.height,o.width=m.width,p(ro(w.data,o))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(s!==void 0)return ro(s,o);throw new Error("Input data provided is not supported - aborted tensor creation")},Ky=(t,e)=>{const{width:r,height:n,download:a,dispose:i}=e,s=[1,n,r,4];return new lr({location:"texture",type:"float32",texture:t,dims:s,download:a,dispose:i})},Yy=(t,e)=>{const{dataType:r,dims:n,download:a,dispose:i}=e;return new lr({location:"gpu-buffer",type:r??"float32",gpuBuffer:t,dims:n,download:a,dispose:i})},Xy=(t,e,r)=>new lr({location:"cpu-pinned",type:t,data:e,dims:r??[e.length]}),Sn=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array]]),Ai=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]);let tp=!1;const Qy=()=>{if(!tp){tp=!0;const t=typeof BigInt64Array<"u"&&BigInt64Array.from,e=typeof BigUint64Array<"u"&&BigUint64Array.from,r=typeof Float16Array<"u"&&Float16Array.from;t&&(Sn.set("int64",BigInt64Array),Ai.set(BigInt64Array,"int64")),e&&(Sn.set("uint64",BigUint64Array),Ai.set(BigUint64Array,"uint64")),r?(Sn.set("float16",Float16Array),Ai.set(Float16Array,"float16")):Sn.set("float16",Uint16Array)}},Jy=t=>{let e=1;for(let r=0;r<t.length;r++){const n=t[r];if(typeof n!="number"||!Number.isSafeInteger(n))throw new TypeError(`dims[${r}] must be an integer, got: ${n}`);if(n<0)throw new RangeError(`dims[${r}] must be a non-negative integer, got: ${n}`);e*=n}return e},Zy=(t,e)=>{switch(t.location){case"cpu":return new lr(t.type,t.data,e);case"cpu-pinned":return new lr({location:"cpu-pinned",data:t.data,type:t.type,dims:e});case"texture":return new lr({location:"texture",texture:t.texture,type:t.type,dims:e});case"gpu-buffer":return new lr({location:"gpu-buffer",gpuBuffer:t.gpuBuffer,type:t.type,dims:e});default:throw new Error(`tensorReshape: tensor location ${t.location} is not supported`)}};let lr=class{constructor(e,r,n){Qy();let a,i;if(typeof e=="object"&&"location"in e)switch(this.dataLocation=e.location,a=e.type,i=e.dims,e.location){case"cpu-pinned":{const o=Sn.get(a);if(!o)throw new TypeError(`unsupported type "${a}" to create tensor from pinned buffer`);if(!(e.data instanceof o))throw new TypeError(`buffer should be of type ${o.name}`);this.cpuData=e.data;break}case"texture":{if(a!=="float32")throw new TypeError(`unsupported type "${a}" to create tensor from texture`);this.gpuTextureData=e.texture,this.downloader=e.download,this.disposer=e.dispose;break}case"gpu-buffer":{if(a!=="float32"&&a!=="float16"&&a!=="int32"&&a!=="int64"&&a!=="uint32"&&a!=="uint8"&&a!=="bool")throw new TypeError(`unsupported type "${a}" to create tensor from gpu buffer`);this.gpuBufferData=e.gpuBuffer,this.downloader=e.download,this.disposer=e.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let o,u;if(typeof e=="string")if(a=e,u=n,e==="string"){if(!Array.isArray(r))throw new TypeError("A string tensor's data must be a string array.");o=r}else{const l=Sn.get(e);if(l===void 0)throw new TypeError(`Unsupported tensor type: ${e}.`);if(Array.isArray(r)){if(e==="float16"&&l===Uint16Array)throw new TypeError("Creating a float16 tensor from number array is not supported. Please use Uint16Array as data.");e==="uint64"||e==="int64"?o=l.from(r,BigInt):o=l.from(r)}else if(r instanceof l)o=r;else throw new TypeError(`A ${a} tensor's data must be type of ${l}`)}else if(u=r,Array.isArray(e)){if(e.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");const l=typeof e[0];if(l==="string")a="string",o=e;else if(l==="boolean")a="bool",o=Uint8Array.from(e);else throw new TypeError(`Invalid element type of data array: ${l}.`)}else{const l=Ai.get(e.constructor);if(l===void 0)throw new TypeError(`Unsupported type for tensor data: ${e.constructor}.`);a=l,o=e}if(u===void 0)u=[o.length];else if(!Array.isArray(u))throw new TypeError("A tensor's dims must be a number array");i=u,this.cpuData=o,this.dataLocation="cpu"}const s=Jy(i);if(this.cpuData&&s!==this.cpuData.length)throw new Error(`Tensor's size(${s}) does not match data length(${this.cpuData.length}).`);this.type=a,this.dims=i,this.size=s}static async fromImage(e,r){return qy(e,r)}static fromTexture(e,r){return Ky(e,r)}static fromGpuBuffer(e,r){return Yy(e,r)}static fromPinnedBuffer(e,r,n){return Xy(e,r,n)}toDataURL(e){return Hy(this,e)}toImageData(e){return jy(this,e)}get data(){if(this.ensureValid(),!this.cpuData)throw new Error("The data is not on CPU. Use `getData()` to download GPU data to CPU, or use `texture` or `gpuBuffer` property to access the GPU data directly.");return this.cpuData}get location(){return this.dataLocation}get texture(){if(this.ensureValid(),!this.gpuTextureData)throw new Error("The data is not stored as a WebGL texture.");return this.gpuTextureData}get gpuBuffer(){if(this.ensureValid(),!this.gpuBufferData)throw new Error("The data is not stored as a WebGPU buffer.");return this.gpuBufferData}async getData(e){switch(this.ensureValid(),this.dataLocation){case"cpu":case"cpu-pinned":return this.data;case"texture":case"gpu-buffer":{if(!this.downloader)throw new Error("The current tensor is not created with a specified data downloader.");if(this.isDownloading)throw new Error("The current tensor is being downloaded.");try{this.isDownloading=!0;const r=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=r,e&&this.disposer&&(this.disposer(),this.disposer=void 0),r}finally{this.isDownloading=!1}}default:throw new Error(`cannot get data from location: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(e){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return Zy(this,e)}};const ew=lr,la=[];let To,ga;Jr.IS_NODE_ENV?(ga=Ze??n_,la.push("cpu"),To=["cpu"]):(ga=Gy,Jr.IS_WEBGPU_AVAILABLE&&la.push("webgpu"),la.push("wasm"),To=["wasm"]);const tw=ga.InferenceSession;function rw(t){let e=To;if(t){if(!la.includes(t))throw new Error(`Unsupported device: "${t}". Should be one of: ${la.join(", ")}.`);e=[t]}return e}async function ig(t,e){return await tw.create(t,e)}function sg(t){return t instanceof ga.Tensor}const Kr=ga?.env;Kr?.wasm&&(Kr.wasm.wasmPaths="https://cdn.jsdelivr.net/npm/onnxruntime-web@1.18.0/dist/",Kr.wasm.proxy=!Jr.IS_WEBWORKER_ENV,(typeof crossOriginIsolated>"u"||!crossOriginIsolated)&&(Kr.wasm.numThreads=1),typeof navigator<"u"&&/iP(hone|od|ad).+16_4.+AppleWebKit/.test(navigator.userAgent)&&(Kr.wasm.simd=!1));function nw(){return Kr?.wasm?.proxy}wt.backends.onnx=Kr;const no=async(t,e,r)=>{const n=await ig(t,e);return async a=>{const i=Object.fromEntries(Object.entries(a).map(([o,u])=>[o,u.ort_tensor])),s=await n.run(i);return new pe(s[r])}};class rp{static session_options={};static get bilinear_interpolate_4d(){return this._bilinear_interpolate_4d||(this._bilinear_interpolate_4d=no(new Uint8Array([8,9,18,0,58,128,1,10,40,10,1,120,10,0,10,0,10,1,115,18,1,121,34,6,82,101,115,105,122,101,42,17,10,4,109,111,100,101,34,6,108,105,110,101,97,114,160,1,3,18,1,114,90,31,10,1,120,18,26,10,24,8,1,18,20,10,3,18,1,98,10,3,18,1,99,10,3,18,1,104,10,3,18,1,119,90,15,10,1,115,18,10,10,8,8,7,18,4,10,2,8,4,98,31,10,1,121,18,26,10,24,8,1,18,20,10,3,18,1,98,10,3,18,1,99,10,3,18,1,104,10,3,18,1,119,66,2,16,20]),this.session_options,"y")),this._bilinear_interpolate_4d}static get bicubic_interpolate_4d(){return this._bicubic_interpolate_4d||(this._bicubic_interpolate_4d=no(new Uint8Array([8,9,18,0,58,127,10,39,10,1,120,10,0,10,0,10,1,115,18,1,121,34,6,82,101,115,105,122,101,42,16,10,4,109,111,100,101,34,5,99,117,98,105,99,160,1,3,18,1,114,90,31,10,1,120,18,26,10,24,8,1,18,20,10,3,18,1,98,10,3,18,1,99,10,3,18,1,104,10,3,18,1,119,90,15,10,1,115,18,10,10,8,8,7,18,4,10,2,8,4,98,31,10,1,121,18,26,10,24,8,1,18,20,10,3,18,1,98,10,3,18,1,99,10,3,18,1,104,10,3,18,1,119,66,2,16,20]),this.session_options,"y")),this._bicubic_interpolate_4d}static get matmul(){return this._matmul||(this._matmul=no(new Uint8Array([8,9,18,0,58,55,10,17,10,1,97,10,1,98,18,1,99,34,6,77,97,116,77,117,108,18,1,114,90,9,10,1,97,18,4,10,2,8,1,90,9,10,1,98,18,4,10,2,8,1,98,9,10,1,99,18,4,10,2,8,1,66,2,16,20]),this.session_options,"c")),this._matmul}}const np=Object.freeze({float32:Float32Array,float16:Uint16Array,float64:Float64Array,string:Array,int8:Int8Array,uint8:Uint8Array,int16:Int16Array,uint16:Uint16Array,int32:Int32Array,uint32:Uint32Array,int64:BigInt64Array,uint64:BigUint64Array,bool:Uint8Array});class pe{get dims(){return this.ort_tensor.dims}set dims(e){this.ort_tensor.dims=e}get type(){return this.ort_tensor.type}get data(){return this.ort_tensor.data}get size(){return this.ort_tensor.size}get location(){return this.ort_tensor.location}ort_tensor;constructor(...e){return sg(e[0])?this.ort_tensor=e[0]:this.ort_tensor=new ew(e[0],e[1],e[2]),new Proxy(this,{get:(r,n)=>{if(typeof n=="string"){let a=Number(n);if(Number.isInteger(a))return r._getitem(a)}return r[n]},set:(r,n,a)=>r[n]=a})}dispose(){this.ort_tensor.dispose()}*[Symbol.iterator](){const[e,...r]=this.dims;if(r.length>0){const n=r.reduce((a,i)=>a*i);for(let a=0;a<e;++a)yield this._subarray(a,n,r)}else yield*this.data}_getitem(e){const[r,...n]=this.dims;if(e=ur(e,r),n.length>0){const a=n.reduce((i,s)=>i*s);return this._subarray(e,a,n)}else return new pe(this.type,[this.data[e]],n)}indexOf(e){const r=this.data;for(let n=0;n<r.length;++n)if(r[n]==e)return n;return-1}_subarray(e,r,n){const a=e*r,i=(e+1)*r,s="subarray"in this.data?this.data.subarray(a,i):this.data.slice(a,i);return new pe(this.type,s,n)}item(){const e=this.data;if(e.length!==1)throw new Error(`a Tensor with ${e.length} elements cannot be converted to Scalar`);return e[0]}tolist(){return aw(this.data,this.dims)}sigmoid(){return this.clone().sigmoid_()}sigmoid_(){const e=this.data;for(let r=0;r<e.length;++r)e[r]=1/(1+Math.exp(-e[r]));return this}mul(e){return this.clone().mul_(e)}mul_(e){const r=this.data;for(let n=0;n<r.length;++n)r[n]*=e;return this}div(e){return this.clone().div_(e)}div_(e){const r=this.data;for(let n=0;n<r.length;++n)r[n]/=e;return this}add(e){return this.clone().add_(e)}add_(e){const r=this.data;for(let n=0;n<r.length;++n)r[n]+=e;return this}clone(){return new pe(this.type,this.data.slice(),this.dims.slice())}slice(...e){let r=[],n=[];for(let l=0;l<this.dims.length;++l){let p=e[l];if(p==null)n.push([0,this.dims[l]]),r.push(this.dims[l]);else if(typeof p=="number")p=ur(p,this.dims[l],l),n.push([p,p+1]);else if(Array.isArray(p)&&p.length===2){let[f,m]=p;if(f=f===null?0:ur(f,this.dims[l],l,!1),m=m===null?this.dims[l]:ur(m,this.dims[l],l,!1),f>m)throw new Error(`Invalid slice: ${p}`);let c=[Math.max(f,0),Math.min(m,this.dims[l])];n.push(c),r.push(c[1]-c[0])}else throw new Error(`Invalid slice: ${p}`)}let a=n.map(([l,p])=>p-l),i=a.reduce((l,p)=>l*p);const s=this.data;let o=new s.constructor(i);const u=this.stride();for(let l=0;l<i;++l){let p=0;for(let f=a.length-1,m=l;f>=0;--f){const c=a[f];p+=(m%c+n[f][0])*u[f],m=Math.floor(m/c)}o[l]=s[p]}return new pe(this.type,o,r)}permute(...e){return iw(this,e)}transpose(...e){return this.permute(...e)}sum(e=null,r=!1){return this.norm(1,e,r)}norm(e="fro",r=null,n=!1){if(e==="fro")e=2;else if(typeof e=="string")throw Error(`Unsupported norm: ${e}`);const a=this.data;if(r===null){let o=a.reduce((u,l)=>u+l**e,0)**(1/e);return new pe(this.type,[o],[])}r=ur(r,this.dims.length);const i=this.dims.slice();i[r]=1;const s=new a.constructor(a.length/this.dims[r]);for(let o=0;o<a.length;++o){let u=0;for(let l=this.dims.length-1,p=o,f=1;l>=0;--l){const m=this.dims[l];if(l!==r){const c=p%m;u+=c*f,f*=i[l]}p=Math.floor(p/m)}s[u]+=a[o]**e}if(e!==1)for(let o=0;o<s.length;++o)s[o]=s[o]**(1/e);return n||i.splice(r,1),new pe(this.type,s,i)}normalize_(e=2,r=1){r=ur(r,this.dims.length);const n=this.norm(e,r,!0),a=this.data;for(let i=0;i<a.length;++i){let s=0;for(let o=this.dims.length-1,u=i,l=1;o>=0;--o){const p=this.dims[o];if(o!==r){const f=u%p;s+=f*l,l*=this.dims[o]}u=Math.floor(u/p)}a[i]/=n.data[s]}return this}normalize(e=2,r=1){return this.clone().normalize_(e,r)}stride(){return uw(this.dims)}squeeze(e=null){return new pe(this.type,this.data,ip(this.dims,e))}squeeze_(e=null){return this.dims=ip(this.dims,e),this}unsqueeze(e=null){return new pe(this.type,this.data,sp(this.dims,e))}unsqueeze_(e=null){return this.dims=sp(this.dims,e),this}flatten_(e=0,r=-1){r=(r+this.dims.length)%this.dims.length;let n=this.dims.slice(0,e),a=this.dims.slice(e,r+1),i=this.dims.slice(r+1);return this.dims=[...n,a.reduce((s,o)=>s*o,1),...i],this}flatten(e=0,r=-1){return this.clone().flatten_(e,r)}view(...e){let r=-1;for(let n=0;n<e.length;++n)if(e[n]===-1){if(r!==-1)throw new Error("Only one dimension can be inferred");r=n}if(r!==-1){const n=e.reduce((a,i,s)=>s!==r?a*i:a,1);e[r]=this.data.length/n}return new pe(this.type,this.data,e)}neg_(){const e=this.data;for(let r=0;r<e.length;++r)e[r]=-e[r];return this}neg(){return this.clone().neg_()}clamp_(e,r){const n=this.data;for(let a=0;a<n.length;++a)n[a]=Math.min(Math.max(n[a],e),r);return this}clamp(e,r){return this.clone().clamp_(e,r)}round_(){const e=this.data;for(let r=0;r<e.length;++r)e[r]=Math.round(e[r]);return this}round(){return this.clone().round_()}mean(e=null,r=!1){return nu(this,e,r)}to(e){if(this.type===e)return this;if(!np.hasOwnProperty(e))throw new Error(`Unsupported type: ${e}`);return new pe(e,np[e].from(this.data),this.dims)}}function aw(t,e){const r=t.length,n=e.reduce((i,s)=>i*s);if(r!==n)throw Error(`cannot reshape array of size ${r} into shape (${e})`);let a=t;for(let i=e.length-1;i>=0;i--)a=a.reduce((s,o)=>{let u=s[s.length-1];return u.length<e[i]?u.push(o):s.push([o]),s},[[]]);return a[0]}function iw(t,e){const[r,n]=__(t.data,t.dims,e);return new pe(t.type,r,n)}function og(t,[e,r],n="bilinear",a=!1){const i=t.dims.at(-3)??1,s=t.dims.at(-2),o=t.dims.at(-1);let u=g_(t.data,[i,s,o],[e,r],n,a);return new pe(t.type,u,[i,e,r])}async function ap(t,{size:e=null,mode:r="bilinear"}={}){if(t.dims.length!==4)throw new Error("`interpolate_4d` currently only supports 4D input.");if(!e)throw new Error("`interpolate_4d` requires a `size` argument.");let n;if(e.length===2)n=[...t.dims.slice(0,2),...e];else if(e.length===3)n=[t.dims[0],...e];else if(e.length===4)n=e;else throw new Error("`size` must be of length 2, 3, or 4.");let a;if(r==="bilinear")a=await rp.bilinear_interpolate_4d;else if(r==="bicubic")a=await rp.bicubic_interpolate_4d;else throw new Error(`Unsupported mode: ${r}`);const i=new pe("int64",new BigInt64Array(n.map(BigInt)),[n.length]);return await a({x:t,s:i})}function ip(t,e){return t=t.slice(),e===null?t=t.filter(r=>r!==1):typeof e=="number"?t[e]===1&&t.splice(e,1):Array.isArray(e)&&(t=t.filter((r,n)=>r!==1||!e.includes(n))),t}function sp(t,e){return e=ur(e,t.length+1),t=t.slice(),t.splice(e,0,1),t}function ur(t,e,r=null,n=!0){if(n&&(t<-e||t>=e))throw new Error(`IndexError: index ${t} is out of bounds for dimension${r===null?"":" "+r} with size ${e}`);return t<0&&(t=(t%e+e)%e),t}function dr(t,e=0){e=ur(e,t[0].dims.length);const r=t[0].dims.slice();r[e]=t.reduce((s,o)=>s+o.dims[e],0);const n=r.reduce((s,o)=>s*o,1),a=new t[0].data.constructor(n),i=t[0].type;if(e===0){let s=0;for(let o of t)a.set(o.data,s),s+=o.data.length}else{let s=0;for(let o=0;o<t.length;++o){let u=t[o];for(let l=0;l<u.data.length;++l){let p=0;for(let f=u.dims.length-1,m=l,c=1;f>=0;--f){const y=u.dims[f];let w=m%y;f===e&&(w+=s),p+=w*c,c*=r[f],m=Math.floor(m/y)}a[p]=u.data[l]}s+=u.dims[e]}}return new pe(i,a,r)}function _a(t,e=0){return dr(t.map(r=>r.unsqueeze(e)),e)}function sw(t,e=null,r=1,n=!1){if(e===null){const l=t.data.reduce((c,y)=>c+y,0)/t.data.length,p=Math.sqrt(t.data.reduce((c,y)=>c+(y-l)**2,0)/(t.data.length-r)),f=new pe(t.type,[l],[]);return[new pe(t.type,[p],[]),f]}e=ur(e,t.dims.length);const a=nu(t,e,n),i=t.dims.slice();i[e]=1;const s=new t.data.constructor(t.data.length/t.dims[e]);for(let u=0;u<t.data.length;++u){let l=0;for(let p=t.dims.length-1,f=u,m=1;p>=0;--p){const c=t.dims[p];if(p!==e){const y=f%c;l+=y*m,m*=i[p]}f=Math.floor(f/c)}s[l]+=(t.data[u]-a.data[l])**2}for(let u=0;u<s.length;++u)s[u]=Math.sqrt(s[u]/(t.dims[e]-r));return n||i.splice(e,1),[new pe(t.type,s,i),a]}function nu(t,e=null,r=!1){if(e===null){let i=t.data.reduce((s,o)=>s+o,0);return new pe(t.type,[i/t.data.length],[])}e=ur(e,t.dims.length);const n=t.dims.slice();n[e]=1;const a=new t.data.constructor(t.data.length/t.dims[e]);for(let i=0;i<t.data.length;++i){let s=0;for(let o=t.dims.length-1,u=i,l=1;o>=0;--o){const p=t.dims[o];if(o!==e){const f=u%p;s+=f*l,l*=n[o]}u=Math.floor(u/p)}a[s]+=t.data[i]}if(t.dims[e]!==1)for(let i=0;i<a.length;++i)a[i]=a[i]/t.dims[e];return r||n.splice(e,1),new pe(t.type,a,n)}function ow(t){const[e,r]=t.dims,n=[e+1,r+1],a=new pe("float32",new Float32Array(n[0]*n[1]).fill(1/0),n),i=new pe("float32",new Float32Array(n[0]*n[1]).fill(-1),n);a[0].data[0]=0;for(let p=1;p<r+1;++p)for(let f=1;f<e+1;++f){const m=a[f-1][p-1].item(),c=a[f-1][p].item(),y=a[f][p-1].item();let w,v;m<c&&m<y?(w=m,v=0):c<m&&c<y?(w=c,v=1):(w=y,v=2),a[f].data[p]=t[f-1][p-1].item()+w,i[f].data[p]=v}let s=e,o=r;i.data.fill(2,0,n[1]);for(let p=0;p<n[0];++p)i[p].data[0]=1;let u=[],l=[];for(;s>0||o>0;)switch(u.push(s-1),l.push(o-1),i[s][o].item()){case 0:--s,--o;break;case 1:--s;break;case 2:--o;break;default:throw new Error(`Internal error in dynamic time warping. Unexpected trace[${s}, ${o}]. Please file a bug report.`)}return u.reverse(),l.reverse(),[u,l]}function uw(t){const e=new Array(t.length);for(let r=t.length-1,n=1;r>=0;--r)e[r]=n,n*=t[r];return e}function au(t,e,r,n){const a=t.reduce((i,s)=>i*s,1);return new pe(r,new n(a).fill(e),t)}function lw(t,e){let r,n;return r="float32",n=Float32Array,au(t,e,r,n)}function dw(t,e){return lw(t.dims,e)}function xa(t){return au(t,1n,"int64",BigInt64Array)}function cw(t){return xa(t.dims)}function pw(t){return au(t,0n,"int64",BigInt64Array)}function hw(t){return pw(t.dims)}class fw{constructor(e=(r,n)=>r>n){this._heap=[],this._comparator=e}get size(){return this._heap.length}isEmpty(){return this.size===0}peek(){return this._heap[0]}push(...e){return this.extend(e)}extend(e){for(const r of e)this._heap.push(r),this._siftUp();return this.size}pop(){const e=this.peek(),r=this.size-1;return r>0&&this._swap(0,r),this._heap.pop(),this._siftDown(),e}replace(e){const r=this.peek();return this._heap[0]=e,this._siftDown(),r}_parent(e){return(e+1>>>1)-1}_left(e){return(e<<1)+1}_right(e){return e+1<<1}_greater(e,r){return this._comparator(this._heap[e],this._heap[r])}_swap(e,r){const n=this._heap[e];this._heap[e]=this._heap[r],this._heap[r]=n}_siftUp(){let e=this.size-1;for(;e>0&&this._greater(e,this._parent(e));)this._swap(e,this._parent(e)),e=this._parent(e)}_siftDown(){let e=0;for(;this._left(e)<this.size&&this._greater(this._left(e),e)||this._right(e)<this.size&&this._greater(this._right(e),e);){const r=this._right(e)<this.size&&this._greater(this._right(e),this._left(e))?this._right(e):this._left(e);this._swap(e,r),e=r}}}class mw{constructor(){this.root=Hi.default()}extend(e){for(let r of e)this.push(r)}push(e){let r=this.root;for(let n of e){let a=r.children.get(n);a===void 0&&(a=Hi.default(),r.children.set(n,a)),r=a}r.isLeaf=!0}*commonPrefixSearch(e){let r=this.root,n="";for(let a=0;a<e.length&&r!==void 0;++a){const i=e[a];n+=i,r=r.children.get(i),r!==void 0&&r.isLeaf&&(yield n)}}}class Hi{constructor(e,r){this.isLeaf=e,this.children=r}static default(){return new Hi(!1,new Map)}}class gw{constructor(e,r,n){this.sentence=e,this.len=e.length,this.bosTokenId=r,this.eosTokenId=n,this.nodes=[],this.beginNodes=Array.from({length:this.len+1},()=>[]),this.endNodes=Array.from({length:this.len+1},()=>[]);const a=new da(this.bosTokenId,0,0,0,0),i=new da(this.eosTokenId,1,this.len,0,0);this.nodes.push(a.clone()),this.nodes.push(i.clone()),this.beginNodes[this.len].push(i),this.endNodes[0].push(a)}insert(e,r,n,a){const i=this.nodes.length,s=new da(a,i,e,r,n);this.beginNodes[e].push(s),this.endNodes[e+r].push(s),this.nodes.push(s)}viterbi(){const e=this.len;let r=0;for(;r<=e;){if(this.beginNodes[r].length==0)return[];for(let o of this.beginNodes[r]){o.prev=null;let u=0,l=null;for(let p of this.endNodes[r]){const f=p.backtraceScore+o.score;(l===null||f>u)&&(l=p.clone(),u=f)}if(l!==null)o.prev=l,o.backtraceScore=u;else return[]}++r}const n=[],i=this.beginNodes[e][0].prev;if(i===null)return[];let s=i.clone();for(;s.prev!==null;)n.push(s.clone()),s=s.clone().prev.clone();return n.reverse(),n}piece(e){return this.sentence.slice(e.pos,e.pos+e.length)}tokens(){return this.viterbi().map(r=>this.piece(r))}tokenIds(){return this.viterbi().map(r=>r.tokenId)}}class da{constructor(e,r,n,a,i){this.tokenId=e,this.nodeId=r,this.pos=n,this.length=a,this.score=i,this.prev=null,this.backtraceScore=0}clone(){const e=new da(this.tokenId,this.nodeId,this.pos,this.length,this.score);return e.prev=this.prev,e.backtraceScore=this.backtraceScore,e}}var W=Object.freeze({Text:"Text",NumericLiteral:"NumericLiteral",BooleanLiteral:"BooleanLiteral",StringLiteral:"StringLiteral",Identifier:"Identifier",Equals:"Equals",OpenParen:"OpenParen",CloseParen:"CloseParen",OpenStatement:"OpenStatement",CloseStatement:"CloseStatement",OpenExpression:"OpenExpression",CloseExpression:"CloseExpression",OpenSquareBracket:"OpenSquareBracket",CloseSquareBracket:"CloseSquareBracket",OpenCurlyBracket:"OpenCurlyBracket",CloseCurlyBracket:"CloseCurlyBracket",Comma:"Comma",Dot:"Dot",Colon:"Colon",Pipe:"Pipe",CallOperator:"CallOperator",AdditiveBinaryOperator:"AdditiveBinaryOperator",MultiplicativeBinaryOperator:"MultiplicativeBinaryOperator",ComparisonBinaryOperator:"ComparisonBinaryOperator",UnaryOperator:"UnaryOperator",Set:"Set",If:"If",For:"For",In:"In",Is:"Is",NotIn:"NotIn",Else:"Else",EndIf:"EndIf",ElseIf:"ElseIf",EndFor:"EndFor",And:"And",Or:"Or",Not:"UnaryOperator"}),op=Object.freeze({set:W.Set,for:W.For,in:W.In,is:W.Is,if:W.If,else:W.Else,endif:W.EndIf,elif:W.ElseIf,endfor:W.EndFor,and:W.And,or:W.Or,not:W.Not,"not in":W.NotIn,true:W.BooleanLiteral,false:W.BooleanLiteral}),Wr=class{constructor(t,e){this.value=t,this.type=e}};function up(t){return/\w/.test(t)}function ao(t){return/[0-9]/.test(t)}var _w=[["{%",W.OpenStatement],["%}",W.CloseStatement],["{{",W.OpenExpression],["}}",W.CloseExpression],["(",W.OpenParen],[")",W.CloseParen],["{",W.OpenCurlyBracket],["}",W.CloseCurlyBracket],["[",W.OpenSquareBracket],["]",W.CloseSquareBracket],[",",W.Comma],[".",W.Dot],[":",W.Colon],["|",W.Pipe],["<=",W.ComparisonBinaryOperator],[">=",W.ComparisonBinaryOperator],["==",W.ComparisonBinaryOperator],["!=",W.ComparisonBinaryOperator],["<",W.ComparisonBinaryOperator],[">",W.ComparisonBinaryOperator],["+",W.AdditiveBinaryOperator],["-",W.AdditiveBinaryOperator],["*",W.MultiplicativeBinaryOperator],["/",W.MultiplicativeBinaryOperator],["%",W.MultiplicativeBinaryOperator],["=",W.Equals]],yw=new Map([["n",`
`],["t","	"],["r","\r"],["b","\b"],["f","\f"],["v","\v"],["'","'"],['"','"'],["\\","\\"]]);function ww(t,e={}){return t.endsWith(`
`)&&(t=t.slice(0,-1)),t=t.replace(/{#.*?#}/gs,"{##}"),e.lstrip_blocks&&(t=t.replace(/^[ \t]*({[#%])/gm,"$1")),e.trim_blocks&&(t=t.replace(/([#%]})\n/g,"$1")),t.replace(/{##}/g,"").replace(/-%}\s*/g,"%}").replace(/\s*{%-/g,"{%").replace(/-}}\s*/g,"}}").replace(/\s*{{-/g,"{{")}function bw(t,e={}){const r=[],n=ww(t,e);let a=0;const i=s=>{let o="";for(;s(n[a]);){if(n[a]==="\\"){if(++a,a>=n.length)throw new SyntaxError("Unexpected end of input");const u=n[a++],l=yw.get(u);if(l===void 0)throw new SyntaxError(`Unexpected escaped character: ${u}`);o+=l;continue}if(o+=n[a++],a>=n.length)throw new SyntaxError("Unexpected end of input")}return o};e:for(;a<n.length;){const s=r.at(-1)?.type;if(s===void 0||s===W.CloseStatement||s===W.CloseExpression){let u="";for(;a<n.length&&!(n[a]==="{"&&(n[a+1]==="%"||n[a+1]==="{"));)u+=n[a++];if(u.length>0){r.push(new Wr(u,W.Text));continue}}i(u=>/\s/.test(u));const o=n[a];if(o==="-"||o==="+"){const u=r.at(-1)?.type;if(u===W.Text||u===void 0)throw new SyntaxError(`Unexpected character: ${o}`);switch(u){case W.Identifier:case W.NumericLiteral:case W.BooleanLiteral:case W.StringLiteral:case W.CloseParen:case W.CloseSquareBracket:break;default:{++a;const l=i(ao);r.push(new Wr(`${o}${l}`,l.length>0?W.NumericLiteral:W.UnaryOperator));continue}}}for(const[u,l]of _w)if(n.slice(a,a+u.length)===u){r.push(new Wr(u,l)),a+=u.length;continue e}if(o==="'"||o==='"'){++a;const u=i(l=>l!==o);r.push(new Wr(u,W.StringLiteral)),++a;continue}if(ao(o)){const u=i(ao);r.push(new Wr(u,W.NumericLiteral));continue}if(up(o)){const u=i(up),l=Object.hasOwn(op,u)?op[u]:W.Identifier;l===W.In&&r.at(-1)?.type===W.Not?(r.pop(),r.push(new Wr("not in",W.NotIn))):r.push(new Wr(u,l));continue}throw new SyntaxError(`Unexpected character: ${o}`)}return r}var Sa=class{type="Statement"},vw=class extends Sa{constructor(t){super(),this.body=t}type="Program"},lp=class extends Sa{constructor(t,e,r){super(),this.test=t,this.body=e,this.alternate=r}type="If"},$w=class extends Sa{constructor(t,e,r){super(),this.loopvar=t,this.iterable=e,this.body=r}type="For"},xw=class extends Sa{constructor(t,e){super(),this.assignee=t,this.value=e}type="Set"},cr=class extends Sa{type="Expression"},Sw=class extends cr{constructor(t,e,r){super(),this.object=t,this.property=e,this.computed=r}type="MemberExpression"},kw=class extends cr{constructor(t,e){super(),this.callee=t,this.args=e}type="CallExpression"},vn=class extends cr{constructor(t){super(),this.value=t}type="Identifier"},Mn=class extends cr{constructor(t){super(),this.value=t}type="Literal"},Ew=class extends Mn{type="NumericLiteral"},dp=class extends Mn{type="StringLiteral"},cp=class extends Mn{type="BooleanLiteral"},Cw=class extends Mn{type="ArrayLiteral"},pp=class extends Mn{type="TupleLiteral"},Tw=class extends Mn{type="ObjectLiteral"},ia=class extends cr{constructor(t,e,r){super(),this.operator=t,this.left=e,this.right=r}type="BinaryExpression"},Iw=class extends cr{constructor(t,e){super(),this.operand=t,this.filter=e}type="FilterExpression"},Aw=class extends cr{constructor(t,e,r){super(),this.operand=t,this.negate=e,this.test=r}type="TestExpression"},Mw=class extends cr{constructor(t,e){super(),this.operator=t,this.argument=e}type="UnaryExpression"},Ow=class extends cr{constructor(t=void 0,e=void 0,r=void 0){super(),this.start=t,this.stop=e,this.step=r}type="SliceExpression"},zw=class extends cr{constructor(t,e){super(),this.key=t,this.value=e}type="KeywordArgumentExpression"};function Pw(t){const e=new vw([]);let r=0;function n(M,P){const H=t[r++];if(!H||H.type!==M)throw new Error(`Parser Error: ${P}. ${H.type} !== ${M}.`);return H}function a(){switch(t[r].type){case W.Text:return o();case W.OpenStatement:return u();case W.OpenExpression:return l();default:throw new SyntaxError(`Unexpected token type: ${t[r].type}`)}}function i(...M){return r+M.length<=t.length&&M.some((P,H)=>P!==t[r+H].type)}function s(...M){return r+M.length<=t.length&&M.every((P,H)=>P===t[r+H].type)}function o(){return new dp(n(W.Text,"Expected text token").value)}function u(){n(W.OpenStatement,"Expected opening statement token");let M;switch(t[r].type){case W.Set:++r,M=p(),n(W.CloseStatement,"Expected closing statement token");break;case W.If:++r,M=f(),n(W.OpenStatement,"Expected {% token"),n(W.EndIf,"Expected endif token"),n(W.CloseStatement,"Expected %} token");break;case W.For:++r,M=c(),n(W.OpenStatement,"Expected {% token"),n(W.EndFor,"Expected endfor token"),n(W.CloseStatement,"Expected %} token");break;default:throw new SyntaxError(`Unknown statement type: ${t[r].type}`)}return M}function l(){n(W.OpenExpression,"Expected opening expression token");const M=y();return n(W.CloseExpression,"Expected closing expression token"),M}function p(){const M=y();if(s(W.Equals)){++r;const P=p();return new xw(M,P)}return M}function f(){const M=y();n(W.CloseStatement,"Expected closing statement token");const P=[],H=[];for(;!(t[r]?.type===W.OpenStatement&&(t[r+1]?.type===W.ElseIf||t[r+1]?.type===W.Else||t[r+1]?.type===W.EndIf));)P.push(a());if(t[r]?.type===W.OpenStatement&&t[r+1]?.type!==W.EndIf)if(++r,s(W.ElseIf))n(W.ElseIf,"Expected elseif token"),H.push(f());else for(n(W.Else,"Expected else token"),n(W.CloseStatement,"Expected closing statement token");!(t[r]?.type===W.OpenStatement&&t[r+1]?.type===W.EndIf);)H.push(a());return new lp(M,P,H)}function m(M=!1){const P=M?re:y,H=[P()],le=s(W.Comma);for(;le&&(++r,H.push(P()),!!s(W.Comma)););return le?new pp(H):H[0]}function c(){const M=m(!0);if(!(M instanceof vn||M instanceof pp))throw new SyntaxError(`Expected identifier/tuple for the loop variable, got ${M.type} instead`);n(W.In,"Expected `in` keyword following loop variable");const P=y();n(W.CloseStatement,"Expected closing statement token");const H=[];for(;i(W.OpenStatement,W.EndFor);)H.push(a());return new $w(M,P,H)}function y(){return w()}function w(){const M=v();if(s(W.If)){++r;const P=v();n(W.Else,"Expected else token");const H=v();return new lp(P,[M],[H])}return M}function v(){let M=k();for(;s(W.Or);){const P=t[r];++r;const H=k();M=new ia(P,M,H)}return M}function k(){let M=$();for(;s(W.And);){const P=t[r];++r;const H=$();M=new ia(P,M,H)}return M}function $(){let M;for(;s(W.Not);){const P=t[r];++r;const H=$();M=new Mw(P,H)}return M??C()}function C(){let M=T();for(;s(W.ComparisonBinaryOperator)||s(W.In)||s(W.NotIn);){const P=t[r];++r;const H=T();M=new ia(P,M,H)}return M}function T(){let M=ie();for(;s(W.AdditiveBinaryOperator);){const P=t[r];++r;const H=ie();M=new ia(P,M,H)}return M}function A(){const M=j();return s(W.OpenParen)?B(M):M}function B(M){let P=new kw(M,R());return s(W.OpenParen)&&(P=B(P)),P}function R(){n(W.OpenParen,"Expected opening parenthesis for arguments list");const M=D();return n(W.CloseParen,"Expected closing parenthesis for arguments list"),M}function D(){const M=[];for(;!s(W.CloseParen);){let P=y();if(s(W.Equals)){if(++r,!(P instanceof vn))throw new SyntaxError("Expected identifier for keyword argument");const H=y();P=new zw(P,H)}M.push(P),s(W.Comma)&&++r}return M}function K(){const M=[];let P=!1;for(;!s(W.CloseSquareBracket);)s(W.Colon)?(M.push(void 0),++r,P=!0):(M.push(y()),s(W.Colon)&&(++r,P=!0));if(M.length===0)throw new SyntaxError("Expected at least one argument for member/slice expression");if(P){if(M.length>3)throw new SyntaxError("Expected 0-3 arguments for slice expression");return new Ow(...M)}return M[0]}function j(){let M=re();for(;s(W.Dot)||s(W.OpenSquareBracket);){const P=t[r];++r;let H;const le=P.type!==W.Dot;if(le)H=K(),n(W.CloseSquareBracket,"Expected closing square bracket");else if(H=re(),H.type!=="Identifier")throw new SyntaxError("Expected identifier following dot operator");M=new Sw(M,H,le)}return M}function ie(){let M=te();for(;s(W.MultiplicativeBinaryOperator);){const P=t[r];++r;const H=te();M=new ia(P,M,H)}return M}function te(){let M=oe();for(;s(W.Is);){++r;const P=s(W.Not);P&&++r;let H=re();if(H instanceof cp&&(H=new vn(H.value.toString())),!(H instanceof vn))throw new SyntaxError("Expected identifier for the test");M=new Aw(M,P,H)}return M}function oe(){let M=A();for(;s(W.Pipe);){++r;let P=re();if(!(P instanceof vn))throw new SyntaxError("Expected identifier for the filter");s(W.OpenParen)&&(P=B(P)),M=new Iw(M,P)}return M}function re(){const M=t[r];switch(M.type){case W.NumericLiteral:return++r,new Ew(Number(M.value));case W.StringLiteral:return++r,new dp(M.value);case W.BooleanLiteral:return++r,new cp(M.value==="true");case W.Identifier:return++r,new vn(M.value);case W.OpenParen:{++r;const P=m();if(t[r].type!==W.CloseParen)throw new SyntaxError(`Expected closing parenthesis, got ${t[r].type} instead`);return++r,P}case W.OpenSquareBracket:{++r;const P=[];for(;!s(W.CloseSquareBracket);)P.push(y()),s(W.Comma)&&++r;return++r,new Cw(P)}case W.OpenCurlyBracket:{++r;const P=new Map;for(;!s(W.CloseCurlyBracket);){const H=y();n(W.Colon,"Expected colon between key and value in object literal");const le=y();P.set(H,le),s(W.Comma)&&++r}return++r,new Tw(P)}default:throw new SyntaxError(`Unexpected token: ${M.type}`)}}for(;r<t.length;)e.body.push(a());return e}function Rw(t,e,r=1){e===void 0&&(e=t,t=0);const n=[];for(let a=t;a<e;a+=r)n.push(a);return n}function hp(t,e,r,n=1){const a=Math.sign(n);a>=0?(e=(e??=0)<0?Math.max(t.length+e,0):Math.min(e,t.length),r=(r??=t.length)<0?Math.max(t.length+r,0):Math.min(r,t.length)):(e=(e??=t.length-1)<0?Math.max(t.length+e,-1):Math.min(e,t.length-1),r=(r??=-1)<-1?Math.max(t.length+r,-1):Math.min(r,t.length-1));const i=[];for(let s=e;a*s<a*r;s+=n)i.push(t[s]);return i}function ug(t){return t.replace(/\b\w/g,e=>e.toUpperCase())}var $r=class{type="RuntimeValue";value;builtins=new Map;constructor(t=void 0){this.value=t}__bool__(){return new et(!!this.value)}},He=class extends $r{type="NumericValue"},Pe=class extends $r{type="StringValue";builtins=new Map([["upper",new wr(()=>new Pe(this.value.toUpperCase()))],["lower",new wr(()=>new Pe(this.value.toLowerCase()))],["strip",new wr(()=>new Pe(this.value.trim()))],["title",new wr(()=>new Pe(ug(this.value)))],["length",new He(this.value.length)]])},et=class extends $r{type="BooleanValue"},er=class extends $r{type="ObjectValue";__bool__(){return new et(this.value.size>0)}builtins=new Map([["get",new wr(([t,e])=>{if(!(t instanceof Pe))throw new Error(`Object key must be a string: got ${t.type}`);return this.value.get(t.value)??e??new ca})],["items",new wr(()=>new Je(Array.from(this.value.entries()).map(([t,e])=>new Je([new Pe(t),e]))))]])},Je=class extends $r{type="ArrayValue";builtins=new Map([["length",new He(this.value.length)]]);__bool__(){return new et(this.value.length>0)}},Bw=class extends Je{type="TupleValue"},wr=class extends $r{type="FunctionValue"},ca=class extends $r{type="NullValue"},Zt=class extends $r{type="UndefinedValue"},Io=class{constructor(t){this.parent=t}variables=new Map([["namespace",new wr(t=>{if(t.length===0)return new er(new Map);if(t.length!==1||!(t[0]instanceof er))throw new Error("`namespace` expects either zero arguments or a single object argument");return t[0]})]]);tests=new Map([["boolean",t=>t.type==="BooleanValue"],["callable",t=>t instanceof wr],["odd",t=>{if(t.type!=="NumericValue")throw new Error(`Cannot apply test "odd" to type: ${t.type}`);return t.value%2!==0}],["even",t=>{if(t.type!=="NumericValue")throw new Error(`Cannot apply test "even" to type: ${t.type}`);return t.value%2===0}],["false",t=>t.type==="BooleanValue"&&!t.value],["true",t=>t.type==="BooleanValue"&&t.value],["number",t=>t.type==="NumericValue"],["integer",t=>t.type==="NumericValue"&&Number.isInteger(t.value)],["iterable",t=>t instanceof Je||t instanceof Pe],["lower",t=>{const e=t.value;return t.type==="StringValue"&&e===e.toLowerCase()}],["upper",t=>{const e=t.value;return t.type==="StringValue"&&e===e.toUpperCase()}],["none",t=>t.type==="NullValue"],["defined",t=>t.type!=="UndefinedValue"],["undefined",t=>t.type==="UndefinedValue"],["equalto",(t,e)=>t.value===e.value]]);set(t,e){return this.declareVariable(t,Mi(e))}declareVariable(t,e){if(this.variables.has(t))throw new SyntaxError(`Variable already declared: ${t}`);return this.variables.set(t,e),e}setVariable(t,e){return this.variables.set(t,e),e}resolve(t){if(this.variables.has(t))return this;if(this.parent)return this.parent.resolve(t);throw new Error(`Unknown variable: ${t}`)}lookupVariable(t){try{return this.resolve(t).variables.get(t)??new Zt}catch{return new Zt}}},Dw=class{global;constructor(t){this.global=t??new Io}run(t){return this.evaluate(t,this.global)}evaluateBinaryExpression(t,e){const r=this.evaluate(t.left,e);switch(t.operator.value){case"and":return r.__bool__().value?this.evaluate(t.right,e):r;case"or":return r.__bool__().value?r:this.evaluate(t.right,e)}const n=this.evaluate(t.right,e);switch(t.operator.value){case"==":return new et(r.value==n.value);case"!=":return new et(r.value!=n.value)}if(r instanceof Zt||n instanceof Zt)throw new Error("Cannot perform operation on undefined values");if(r instanceof ca||n instanceof ca)throw new Error("Cannot perform operation on null values");if(r instanceof He&&n instanceof He)switch(t.operator.value){case"+":return new He(r.value+n.value);case"-":return new He(r.value-n.value);case"*":return new He(r.value*n.value);case"/":return new He(r.value/n.value);case"%":return new He(r.value%n.value);case"<":return new et(r.value<n.value);case">":return new et(r.value>n.value);case">=":return new et(r.value>=n.value);case"<=":return new et(r.value<=n.value)}else if(r instanceof Je&&n instanceof Je)switch(t.operator.value){case"+":return new Je(r.value.concat(n.value))}else if(n instanceof Je){const a=n.value.find(i=>i.value===r.value)!==void 0;switch(t.operator.value){case"in":return new et(a);case"not in":return new et(!a)}}if(r instanceof Pe||n instanceof Pe)switch(t.operator.value){case"+":return new Pe(r.value.toString()+n.value.toString())}if(r instanceof Pe&&n instanceof Pe)switch(t.operator.value){case"in":return new et(n.value.includes(r.value));case"not in":return new et(!n.value.includes(r.value))}if(r instanceof Pe&&n instanceof er)switch(t.operator.value){case"in":return new et(n.value.has(r.value));case"not in":return new et(!n.value.has(r.value))}throw new SyntaxError(`Unknown operator "${t.operator.value}" between ${r.type} and ${n.type}`)}evaluateFilterExpression(t,e){const r=this.evaluate(t.operand,e);if(t.filter.type==="Identifier"){const n=t.filter;if(r instanceof Je)switch(n.value){case"list":return r;case"first":return r.value[0];case"last":return r.value[r.value.length-1];case"length":return new He(r.value.length);case"reverse":return new Je(r.value.reverse());case"sort":return new Je(r.value.sort((a,i)=>{if(a.type!==i.type)throw new Error(`Cannot compare different types: ${a.type} and ${i.type}`);switch(a.type){case"NumericValue":return a.value-i.value;case"StringValue":return a.value.localeCompare(i.value);default:throw new Error(`Cannot compare type: ${a.type}`)}}));default:throw new Error(`Unknown ArrayValue filter: ${n.value}`)}else if(r instanceof Pe)switch(n.value){case"length":return new He(r.value.length);case"upper":return new Pe(r.value.toUpperCase());case"lower":return new Pe(r.value.toLowerCase());case"title":return new Pe(ug(r.value));case"capitalize":return new Pe(r.value.charAt(0).toUpperCase()+r.value.slice(1));case"trim":return new Pe(r.value.trim());default:throw new Error(`Unknown StringValue filter: ${n.value}`)}else if(r instanceof He)switch(n.value){case"abs":return new He(Math.abs(r.value));default:throw new Error(`Unknown NumericValue filter: ${n.value}`)}else if(r instanceof er)switch(n.value){case"items":return new Je(Array.from(r.value.entries()).map(([a,i])=>new Je([new Pe(a),i])));case"length":return new He(r.value.size);default:throw new Error(`Unknown ObjectValue filter: ${n.value}`)}throw new Error(`Cannot apply filter "${n.value}" to type: ${r.type}`)}else if(t.filter.type==="CallExpression"){const n=t.filter;if(n.callee.type!=="Identifier")throw new Error(`Unknown filter: ${n.callee.type}`);const a=n.callee.value;if(r instanceof Je){switch(a){case"selectattr":{if(r.value.some(p=>!(p instanceof er)))throw new Error("`selectattr` can only be applied to array of objects");if(n.args.some(p=>p.type!=="StringLiteral"))throw new Error("arguments of `selectattr` must be strings");const[i,s,o]=n.args.map(p=>this.evaluate(p,e));let u;if(s){const p=e.tests.get(s.value);if(!p)throw new Error(`Unknown test: ${s.value}`);u=p}else u=(...p)=>p[0].__bool__().value;const l=r.value.filter(p=>{const f=p.value.get(i.value);return f?u(f,o):!1});return new Je(l)}}throw new Error(`Unknown ArrayValue filter: ${a}`)}else throw new Error(`Cannot apply filter "${a}" to type: ${r.type}`)}throw new Error(`Unknown filter: ${t.filter.type}`)}evaluateTestExpression(t,e){const r=this.evaluate(t.operand,e),n=e.tests.get(t.test.value);if(!n)throw new Error(`Unknown test: ${t.test.value}`);const a=n(r);return new et(t.negate?!a:a)}evaluateUnaryExpression(t,e){const r=this.evaluate(t.argument,e);switch(t.operator.value){case"not":return new et(!r.value);default:throw new SyntaxError(`Unknown operator: ${t.operator.value}`)}}evalProgram(t,e){return this.evaluateBlock(t.body,e)}evaluateBlock(t,e){let r="";for(const n of t){const a=this.evaluate(n,e);a.type!=="NullValue"&&a.type!=="UndefinedValue"&&(r+=a.value)}return new Pe(r)}evaluateIdentifier(t,e){return e.lookupVariable(t.value)}evaluateCallExpression(t,e){const r=[],n=new Map;for(const i of t.args)if(i.type==="KeywordArgumentExpression"){const s=i;n.set(s.key.value,this.evaluate(s.value,e))}else r.push(this.evaluate(i,e));n.size>0&&r.push(new er(n));const a=this.evaluate(t.callee,e);if(a.type!=="FunctionValue")throw new Error(`Cannot call something that is not a function: got ${a.type}`);return a.value(r,e)}evaluateSliceExpression(t,e,r){if(!(t instanceof Je||t instanceof Pe))throw new Error("Slice object must be an array or string");const n=this.evaluate(e.start,r),a=this.evaluate(e.stop,r),i=this.evaluate(e.step,r);if(!(n instanceof He||n instanceof Zt))throw new Error("Slice start must be numeric or undefined");if(!(a instanceof He||a instanceof Zt))throw new Error("Slice stop must be numeric or undefined");if(!(i instanceof He||i instanceof Zt))throw new Error("Slice step must be numeric or undefined");return t instanceof Je?new Je(hp(t.value,n.value,a.value,i.value)):new Pe(hp(Array.from(t.value),n.value,a.value,i.value).join(""))}evaluateMemberExpression(t,e){const r=this.evaluate(t.object,e);let n;if(t.computed){if(t.property.type==="SliceExpression")return this.evaluateSliceExpression(r,t.property,e);n=this.evaluate(t.property,e)}else n=new Pe(t.property.value);let a;if(r instanceof er){if(!(n instanceof Pe))throw new Error(`Cannot access property with non-string: got ${n.type}`);a=r.value.get(n.value)??r.builtins.get(n.value)}else if(r instanceof Je||r instanceof Pe)if(n instanceof He)a=r.value.at(n.value),r instanceof Pe&&(a=new Pe(r.value.at(n.value)));else if(n instanceof Pe)a=r.builtins.get(n.value);else throw new Error(`Cannot access property with non-string/non-number: got ${n.type}`);else{if(!(n instanceof Pe))throw new Error(`Cannot access property with non-string: got ${n.type}`);a=r.builtins.get(n.value)}return a instanceof $r?a:new Zt}evaluateSet(t,e){const r=this.evaluate(t.value,e);if(t.assignee.type==="Identifier"){const n=t.assignee.value;e.setVariable(n,r)}else if(t.assignee.type==="MemberExpression"){const n=t.assignee,a=this.evaluate(n.object,e);if(!(a instanceof er))throw new Error("Cannot assign to member of non-object");if(n.property.type!=="Identifier")throw new Error("Cannot assign to member with non-identifier property");a.value.set(n.property.value,r)}else throw new Error(`Invalid LHS inside assignment expression: ${JSON.stringify(t.assignee)}`);return new ca}evaluateIf(t,e){const r=this.evaluate(t.test,e);return this.evaluateBlock(r.__bool__().value?t.body:t.alternate,e)}evaluateFor(t,e){const r=new Io(e),n=this.evaluate(t.iterable,r);if(!(n instanceof Je))throw new Error(`Expected iterable type in for loop: got ${n.type}`);let a="";for(let i=0;i<n.value.length;++i){const s=new Map([["index",new He(i+1)],["index0",new He(i)],["revindex",new He(n.value.length-i)],["revindex0",new He(n.value.length-i-1)],["first",new et(i===0)],["last",new et(i===n.value.length-1)],["length",new He(n.value.length)],["previtem",i>0?n.value[i-1]:new Zt],["nextitem",i<n.value.length-1?n.value[i+1]:new Zt]]);r.setVariable("loop",new er(s));const o=n.value[i];if(t.loopvar.type==="Identifier")r.setVariable(t.loopvar.value,o);else if(t.loopvar.type==="TupleLiteral"){const l=t.loopvar;if(o.type!=="ArrayValue")throw new Error(`Cannot unpack non-iterable type: ${o.type}`);const p=o;if(l.value.length!==p.value.length)throw new Error(`Too ${l.value.length>p.value.length?"few":"many"} items to unpack`);for(let f=0;f<l.value.length;++f){if(l.value[f].type!=="Identifier")throw new Error(`Cannot unpack non-identifier type: ${l.value[f].type}`);r.setVariable(l.value[f].value,p.value[f])}}const u=this.evaluateBlock(t.body,r);a+=u.value}return new Pe(a)}evaluate(t,e){if(t===void 0)return new Zt;switch(t.type){case"Program":return this.evalProgram(t,e);case"Set":return this.evaluateSet(t,e);case"If":return this.evaluateIf(t,e);case"For":return this.evaluateFor(t,e);case"NumericLiteral":return new He(Number(t.value));case"StringLiteral":return new Pe(t.value);case"BooleanLiteral":return new et(t.value);case"ArrayLiteral":return new Je(t.value.map(r=>this.evaluate(r,e)));case"TupleLiteral":return new Bw(t.value.map(r=>this.evaluate(r,e)));case"ObjectLiteral":{const r=new Map;for(const[n,a]of t.value){const i=this.evaluate(n,e);if(!(i instanceof Pe))throw new Error(`Object keys must be strings: got ${i.type}`);r.set(i.value,this.evaluate(a,e))}return new er(r)}case"Identifier":return this.evaluateIdentifier(t,e);case"CallExpression":return this.evaluateCallExpression(t,e);case"MemberExpression":return this.evaluateMemberExpression(t,e);case"UnaryExpression":return this.evaluateUnaryExpression(t,e);case"BinaryExpression":return this.evaluateBinaryExpression(t,e);case"FilterExpression":return this.evaluateFilterExpression(t,e);case"TestExpression":return this.evaluateTestExpression(t,e);default:throw new SyntaxError(`Unknown node type: ${t.type}`)}}};function Mi(t){switch(typeof t){case"number":return new He(t);case"string":return new Pe(t);case"boolean":return new et(t);case"object":return t===null?new ca:Array.isArray(t)?new Je(t.map(Mi)):new er(new Map(Object.entries(t).map(([e,r])=>[e,Mi(r)])));case"function":return new wr((e,r)=>{const n=t(...e.map(a=>a.value))??null;return Mi(n)});default:throw new Error(`Cannot convert to runtime value: ${t}`)}}var Nw=class{parsed;constructor(t){const e=bw(t,{lstrip_blocks:!0,trim_blocks:!0});this.parsed=Pw(e)}render(t){const e=new Io;e.set("false",!1),e.set("true",!0),e.set("raise_exception",a=>{throw new Error(a)}),e.set("range",Rw);for(const[a,i]of Object.entries(t))e.set(a,i);return new Dw(e).run(this.parsed).value}};const lg=[["en","english"],["zh","chinese"],["de","german"],["es","spanish"],["ru","russian"],["ko","korean"],["fr","french"],["ja","japanese"],["pt","portuguese"],["tr","turkish"],["pl","polish"],["ca","catalan"],["nl","dutch"],["ar","arabic"],["sv","swedish"],["it","italian"],["id","indonesian"],["hi","hindi"],["fi","finnish"],["vi","vietnamese"],["he","hebrew"],["uk","ukrainian"],["el","greek"],["ms","malay"],["cs","czech"],["ro","romanian"],["da","danish"],["hu","hungarian"],["ta","tamil"],["no","norwegian"],["th","thai"],["ur","urdu"],["hr","croatian"],["bg","bulgarian"],["lt","lithuanian"],["la","latin"],["mi","maori"],["ml","malayalam"],["cy","welsh"],["sk","slovak"],["te","telugu"],["fa","persian"],["lv","latvian"],["bn","bengali"],["sr","serbian"],["az","azerbaijani"],["sl","slovenian"],["kn","kannada"],["et","estonian"],["mk","macedonian"],["br","breton"],["eu","basque"],["is","icelandic"],["hy","armenian"],["ne","nepali"],["mn","mongolian"],["bs","bosnian"],["kk","kazakh"],["sq","albanian"],["sw","swahili"],["gl","galician"],["mr","marathi"],["pa","punjabi"],["si","sinhala"],["km","khmer"],["sn","shona"],["yo","yoruba"],["so","somali"],["af","afrikaans"],["oc","occitan"],["ka","georgian"],["be","belarusian"],["tg","tajik"],["sd","sindhi"],["gu","gujarati"],["am","amharic"],["yi","yiddish"],["lo","lao"],["uz","uzbek"],["fo","faroese"],["ht","haitian creole"],["ps","pashto"],["tk","turkmen"],["nn","nynorsk"],["mt","maltese"],["sa","sanskrit"],["lb","luxembourgish"],["my","myanmar"],["bo","tibetan"],["tl","tagalog"],["mg","malagasy"],["as","assamese"],["tt","tatar"],["haw","hawaiian"],["ln","lingala"],["ha","hausa"],["ba","bashkir"],["jw","javanese"],["su","sundanese"]],Oi=new Map(lg),Fw=new Map([...lg.map(([t,e])=>[e,t]),["burmese","my"],["valencian","ca"],["flemish","nl"],["haitian","ht"],["letzeburgesch","lb"],["pushto","ps"],["panjabi","pa"],["moldavian","ro"],["moldovan","ro"],["sinhalese","si"],["castilian","es"]]);function dg(t){t=t.toLowerCase();let e=Fw.get(t);if(e===void 0)if(Oi.has(t))e=t;else{const n=t.length===2?Oi.keys():Oi.values();throw new Error(`Language "${t}" is not supported. Must be one of: ${JSON.stringify(n)}`)}return e}const io="https://github.com/xenova/transformers.js/issues/new/choose";async function cg(t,e){const r=await Promise.all([Ar(t,"tokenizer.json",!0,e),Ar(t,"tokenizer_config.json",!0,e)]);return e.legacy!==null&&(r[1].legacy=e.legacy),r}function Lw(t,e){const r=[];let n=0;for(const a of t.matchAll(e)){const i=a[0];n<a.index&&r.push(t.slice(n,a.index)),i.length>0&&r.push(i),n=a.index+i.length}return n<t.length&&r.push(t.slice(n)),r}function Qi(t,e=!0){if(t.Regex!==void 0){let r=t.Regex.replace(/\\([#&~])/g,"$1");for(const[n,a]of Hw)r=r.replaceAll(n,a);return new RegExp(r,"gu")}else if(t.String!==void 0){const r=Ip(t.String);return new RegExp(e?r:`(${r})`,"gu")}else return console.warn("Unknown pattern type:",t),null}function iu(t){return new Map(Object.entries(t))}function pg(t){const e=t.dims;switch(e.length){case 1:return t.tolist();case 2:if(e[0]!==1)throw new Error("Unable to decode tensor with `batch size !== 1`. Use `tokenizer.batch_decode(...)` for batched inputs.");return t.tolist()[0];default:throw new Error(`Expected tensor to have 1-2 dimensions, got ${e.length}.`)}}function su(t){return t.replace(/ \./g,".").replace(/ \?/g,"?").replace(/ \!/g,"!").replace(/ ,/g,",").replace(/ \' /g,"'").replace(/ n\'t/g,"n't").replace(/ \'m/g,"'m").replace(/ \'s/g,"'s").replace(/ \'ve/g,"'ve").replace(/ \'re/g,"'re")}function hg(t){return t.replace(/[\u0300-\u036f]/g,"")}function Uw(t){return hg(t.toLowerCase())}function Ww(t){return t>=19968&&t<=40959||t>=13312&&t<=19903||t>=131072&&t<=173791||t>=173824&&t<=177983||t>=177984&&t<=178207||t>=178208&&t<=183983||t>=63744&&t<=64255||t>=194560&&t<=195103}function Vw(t,e,r){const n=[];let a=0;for(;a<t.length;){if(n.push(t[a]),(r.get(t[a])??e)!==e){++a;continue}for(;a<t.length&&(r.get(t[a])??e)===e;)++a}return n}function Gw(t){return t.match(/\S+/g)||[]}const ya="\\p{P}\\u0021-\\u002F\\u003A-\\u0040\\u005B-\\u0060\\u007B-\\u007E",Hw=new Map([["(?i:'s|'t|'re|'ve|'m|'ll|'d)","(?:'([sS]|[tT]|[rR][eE]|[vV][eE]|[mM]|[lL][lL]|[dD]))"]]);class jw{constructor(e){this.content=e.content,this.id=e.id,this.single_word=e.single_word??!1,this.lstrip=e.lstrip??!1,this.rstrip=e.rstrip??!1,this.special=e.special??!1,this.normalized=e.normalized??null}}class ka extends $t{constructor(e){super(),this.config=e,this.vocab=[],this.tokens_to_ids=new Map,this.unk_token_id=void 0,this.unk_token=void 0,this.end_of_word_suffix=void 0,this.fuse_unk=this.config.fuse_unk??!1}static fromConfig(e,...r){switch(e.type){case"WordPiece":return new qw(e);case"Unigram":return new Kw(e,...r);case"BPE":return new Xw(e);default:if(e.vocab)return new Qw(e,...r);throw new Error(`Unknown TokenizerModel type: ${e.type}`)}}_call(e){let r=this.encode(e);return this.fuse_unk&&(r=Vw(r,this.unk_token_id,this.tokens_to_ids)),r}encode(e){throw Error("encode should be implemented in subclass.")}convert_tokens_to_ids(e){return e.map(r=>this.tokens_to_ids.get(r)??this.unk_token_id)}convert_ids_to_tokens(e){return e.map(r=>this.vocab[r]??this.unk_token)}}class qw extends ka{constructor(e){super(e),this.tokens_to_ids=iu(e.vocab),this.unk_token_id=this.tokens_to_ids.get(e.unk_token),this.unk_token=e.unk_token,this.max_input_chars_per_word=e.max_input_chars_per_word??100,this.vocab=new Array(this.tokens_to_ids.size);for(const[r,n]of this.tokens_to_ids)this.vocab[n]=r}encode(e){const r=[];for(const n of e){const a=[...n];if(a.length>this.max_input_chars_per_word){r.push(this.unk_token);continue}let i=!1,s=0;const o=[];for(;s<a.length;){let u=a.length,l=null;for(;s<u;){let p=a.slice(s,u).join("");if(s>0&&(p=this.config.continuing_subword_prefix+p),this.tokens_to_ids.has(p)){l=p;break}--u}if(l===null){i=!0;break}o.push(l),s=u}i?r.push(this.unk_token):r.push(...o)}return r}}class Kw extends ka{constructor(e,r){super(e);const n=e.vocab.length;this.vocab=new Array(n),this.scores=new Array(n);for(let a=0;a<n;++a){const i=e.vocab[a];this.vocab[a]=i[0],this.scores[a]=i[1]}this.unk_token_id=e.unk_id,this.unk_token=this.vocab[e.unk_id],this.tokens_to_ids=new Map(this.vocab.map((a,i)=>[a,i])),this.bosToken=" ",this.bosTokenId=this.tokens_to_ids.get(this.bosToken),this.eosToken=r.eos_token,this.eosTokenId=this.tokens_to_ids.get(this.eosToken),this.unkToken=this.vocab[this.unk_token_id],this.minScore=Mp(this.scores)[0],this.unkScore=this.minScore-10,this.scores[this.unk_token_id]=this.unkScore,this.trie=new mw,this.trie.extend(this.vocab),this.fuse_unk=!0}populateNodes(e){const r=e.sentence,n=r.length;let a=0;for(;a<n;){let s=!1;for(let o of this.trie.commonPrefixSearch(r.slice(a))){const u=this.tokens_to_ids.get(o),l=this.scores[u],p=o.length;e.insert(a,p,l,u),!s&&p===1&&(s=!0)}s||e.insert(a,1,this.unkScore,this.unk_token_id),a+=1}}tokenize(e){const r=new gw(e,this.bosTokenId,this.eosTokenId);return this.populateNodes(r),r.tokens()}encode(e){const r=[];for(const n of e){const a=this.tokenize(n);r.push(...a)}return r}}const fg=(()=>{const t=[...Array.from({length:94},(a,i)=>i+33),...Array.from({length:12},(a,i)=>i+161),...Array.from({length:82},(a,i)=>i+174)],e=t.slice();let r=0;for(let a=0;a<256;++a)t.includes(a)||(t.push(a),e.push(256+r),r+=1);const n=e.map(a=>String.fromCharCode(a));return Object.fromEntries(t.map((a,i)=>[a,n[i]]))})(),Yw=d_(fg);class Xw extends ka{constructor(e){super(e),this.BPE_SPLIT_TOKEN=" ",this.tokens_to_ids=iu(e.vocab),this.unk_token_id=this.tokens_to_ids.get(e.unk_token),this.unk_token=e.unk_token,this.vocab=new Array(this.tokens_to_ids.size);for(const[r,n]of this.tokens_to_ids)this.vocab[n]=r;this.bpe_ranks=new Map(e.merges.map((r,n)=>[r,n])),this.merges=e.merges.map(r=>r.split(this.BPE_SPLIT_TOKEN)),this.end_of_word_suffix=e.end_of_word_suffix,this.continuing_subword_suffix=e.continuing_subword_suffix??null,this.byte_fallback=this.config.byte_fallback??!1,this.byte_fallback&&(this.text_encoder=new TextEncoder),this.ignore_merges=this.config.ignore_merges??!1,this.cache=new Map}bpe(e){if(e.length===0)return[];const r=this.cache.get(e);if(r!==void 0)return r;const n=Array.from(e);this.end_of_word_suffix&&(n[n.length-1]+=this.end_of_word_suffix);let a=[];if(n.length>1){const i=new fw((u,l)=>u.score<l.score);let s={token:n[0],bias:0,prev:null,next:null},o=s;for(let u=1;u<n.length;++u){const l={bias:u/n.length,token:n[u],prev:o,next:null};o.next=l,this._add_node(i,o),o=l}for(;!i.isEmpty();){const u=i.pop();if(u.deleted||!u.next||u.next.deleted)continue;if(u.deleted=!0,u.next.deleted=!0,u.prev){const p={...u.prev};u.prev.deleted=!0,u.prev=p,p.prev?p.prev.next=p:s=p}const l={token:u.token+u.next.token,bias:u.bias,prev:u.prev,next:u.next.next};l.prev?(l.prev.next=l,this._add_node(i,l.prev)):s=l,l.next&&(l.next.prev=l,this._add_node(i,l))}for(let u=s;u!==null;u=u.next)a.push(u.token)}else a=n;if(this.continuing_subword_suffix)for(let i=0;i<a.length-1;++i)a[i]+=this.continuing_subword_suffix;return this.cache.set(e,a),a}_add_node(e,r){const n=this.bpe_ranks.get(r.token+this.BPE_SPLIT_TOKEN+r.next.token);n!==void 0&&(r.score=n+r.bias,e.push(r))}encode(e){const r=[];for(const n of e){if(this.ignore_merges&&this.tokens_to_ids.has(n)){r.push(n);continue}const a=this.bpe(n);for(const i of a)this.tokens_to_ids.has(i)?r.push(i):this.byte_fallback?r.push(...Array.from(this.text_encoder.encode(i)).map(s=>`<0x${s.toString(16).toUpperCase().padStart(2,"0")}>`)):r.push(this.unk_token)}return r}}class Qw extends ka{constructor(e,r){super(e),this.tokens_to_ids=iu(r.target_lang?e.vocab[r.target_lang]:e.vocab),this.bos_token=r.bos_token,this.bos_token_id=this.tokens_to_ids.get(this.bos_token),this.eos_token=r.eos_token,this.eos_token_id=this.tokens_to_ids.get(this.eos_token),this.pad_token=r.pad_token,this.pad_token_id=this.tokens_to_ids.get(this.pad_token),this.unk_token=r.unk_token,this.unk_token_id=this.tokens_to_ids.get(this.unk_token),this.vocab=new Array(this.tokens_to_ids.size);for(const[n,a]of this.tokens_to_ids)this.vocab[a]=n}encode(e){return e}}class Pt extends $t{constructor(e){super(),this.config=e}static fromConfig(e){if(e===null)return null;switch(e.type){case"BertNormalizer":return new ob(e);case"Precompiled":return new Eb(e);case"Sequence":return new sb(e);case"Replace":return new Jw(e);case"NFC":return new Zw(e);case"NFKC":return new eb(e);case"NFKD":return new tb(e);case"Strip":return new rb(e);case"StripAccents":return new nb(e);case"Lowercase":return new ab(e);case"Prepend":return new ib(e);default:throw new Error(`Unknown Normalizer type: ${e.type}`)}}normalize(e){throw Error("normalize should be implemented in subclass.")}_call(e){return this.normalize(e)}}class Jw extends Pt{normalize(e){const r=Qi(this.config.pattern);return r===null?e:e.replaceAll(r,this.config.content)}}class Zw extends Pt{normalize(e){return e=e.normalize("NFC"),e}}class eb extends Pt{normalize(e){return e=e.normalize("NFKC"),e}}class tb extends Pt{normalize(e){return e=e.normalize("NFKD"),e}}class rb extends Pt{normalize(e){return this.config.strip_left&&this.config.strip_right?e=e.trim():(this.config.strip_left&&(e=e.trimStart()),this.config.strip_right&&(e=e.trimEnd())),e}}class nb extends Pt{normalize(e){return e=hg(e),e}}class ab extends Pt{normalize(e){return e=e.toLowerCase(),e}}class ib extends Pt{normalize(e){return e=this.config.prepend+e,e}}class sb extends Pt{constructor(e){super(e),this.normalizers=e.normalizers.map(r=>Pt.fromConfig(r))}normalize(e){return this.normalizers.reduce((r,n)=>n.normalize(r),e)}}class ob extends Pt{_tokenize_chinese_chars(e){const r=[];for(let n=0;n<e.length;++n){const a=e[n],i=a.charCodeAt(0);Ww(i)?(r.push(" "),r.push(a),r.push(" ")):r.push(a)}return r.join("")}stripAccents(e){return e.normalize("NFD").replace(/[\u0300-\u036f]/g,"")}_is_control(e){switch(e){case"	":case`
`:case"\r":return!1;default:return/^\p{Cc}|\p{Cf}|\p{Co}|\p{Cs}$/u.test(e)}}_clean_text(e){const r=[];for(const n of e){const a=n.charCodeAt(0);a===0||a===65533||this._is_control(n)||(/^\s$/.test(n)?r.push(" "):r.push(n))}return r.join("")}normalize(e){return this.config.clean_text&&(e=this._clean_text(e)),this.config.handle_chinese_chars&&(e=this._tokenize_chinese_chars(e)),this.config.lowercase?(e=e.toLowerCase(),this.config.strip_accents!==!1&&(e=this.stripAccents(e))):this.config.strip_accents&&(e=this.stripAccents(e)),e}}class Ut extends $t{static fromConfig(e){if(e===null)return null;switch(e.type){case"BertPreTokenizer":return new ub(e);case"Sequence":return new Cb(e);case"Whitespace":return new Tb(e);case"WhitespaceSplit":return new Ib(e);case"Metaspace":return new yg(e);case"ByteLevel":return new lb(e);case"Split":return new db(e);case"Punctuation":return new cb(e);case"Digits":return new pb(e);case"Replace":return new Ab(e);default:throw new Error(`Unknown PreTokenizer type: ${e.type}`)}}pre_tokenize_text(e,r){throw Error("pre_tokenize_text should be implemented in subclass.")}pre_tokenize(e,r){return(Array.isArray(e)?e.map(n=>this.pre_tokenize_text(n,r)):this.pre_tokenize_text(e,r)).flat()}_call(e,r){return this.pre_tokenize(e,r)}}class ub extends Ut{constructor(e){super(),this.pattern=new RegExp(`[^\\s${ya}]+|[${ya}]`,"gu")}pre_tokenize_text(e,r){return e.trim().match(this.pattern)||[]}}class lb extends Ut{constructor(e){super(),this.config=e,this.add_prefix_space=this.config.add_prefix_space,this.trim_offsets=this.config.trim_offsets,this.use_regex=this.config.use_regex??!0,this.pattern=/'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+/gu,this.byte_encoder=fg,this.text_encoder=new TextEncoder}pre_tokenize_text(e,r){return this.add_prefix_space&&!e.startsWith(" ")&&(e=" "+e),(this.use_regex?e.match(this.pattern)||[]:[e]).map(a=>Array.from(this.text_encoder.encode(a),i=>this.byte_encoder[i]).join(""))}}class db extends Ut{constructor(e){super(),this.config=e,this.pattern=Qi(this.config.pattern,this.config.invert)}pre_tokenize_text(e,r){return this.pattern===null?[]:this.config.invert?e.match(this.pattern)||[]:Lw(e,this.pattern)}}class cb extends Ut{constructor(e){super(),this.config=e,this.pattern=new RegExp(`[^${ya}]+|[${ya}]+`,"gu")}pre_tokenize_text(e,r){return e.match(this.pattern)||[]}}class pb extends Ut{constructor(e){super(),this.config=e;const r=`[^\\d]+|\\d${this.config.individual_digits?"":"+"}`;this.pattern=new RegExp(r,"gu")}pre_tokenize_text(e,r){return e.match(this.pattern)||[]}}class Cn extends $t{constructor(e){super(),this.config=e}static fromConfig(e){if(e===null)return null;switch(e.type){case"TemplateProcessing":return new hb(e);case"ByteLevel":return new _g(e);case"RobertaProcessing":return new gg(e);case"BertProcessing":return new mg(e);case"Sequence":return new fb(e);default:throw new Error(`Unknown PostProcessor type: ${e.type}`)}}post_process(e,...r){throw Error("post_process should be implemented in subclass.")}_call(e,...r){return this.post_process(e,...r)}}class mg extends Cn{constructor(e){super(e),this.cls=e.cls[0],this.sep=e.sep[0]}post_process(e,r=null,{add_special_tokens:n=!0}={}){n&&(e=ct([this.cls],e,[this.sep]));let a=new Array(e.length).fill(0);if(r!==null){const i=n&&this instanceof gg?[this.sep]:[],s=n?[this.sep]:[];e=ct(e,i,r,s),a=ct(a,new Array(r.length+i.length+s.length).fill(1))}return{tokens:e,token_type_ids:a}}}class gg extends mg{}class hb extends Cn{constructor(e){super(e),this.single=e.single,this.pair=e.pair}post_process(e,r=null,{add_special_tokens:n=!0}={}){const a=r===null?this.single:this.pair;let i=[],s=[];for(const o of a)"SpecialToken"in o?n&&(i.push(o.SpecialToken.id),s.push(o.SpecialToken.type_id)):"Sequence"in o&&(o.Sequence.id==="A"?(i=ct(i,e),s=ct(s,new Array(e.length).fill(o.Sequence.type_id))):o.Sequence.id==="B"&&(i=ct(i,r),s=ct(s,new Array(r.length).fill(o.Sequence.type_id))));return{tokens:i,token_type_ids:s}}}class _g extends Cn{post_process(e,r=null){return r&&(e=ct(e,r)),{tokens:e}}}class fb extends Cn{constructor(e){super(e),this.processors=e.processors.map(r=>Cn.fromConfig(r))}post_process(e,r=null,n={}){let a;for(const i of this.processors)if(i instanceof _g)e=i.post_process(e).tokens,r&&(r=i.post_process(r).tokens);else{const s=i.post_process(e,r,n);e=s.tokens,a=s.token_type_ids}return{tokens:e,token_type_ids:a}}}class Rt extends $t{constructor(e){super(),this.config=e,this.added_tokens=[],this.end_of_word_suffix=null,this.trim_offsets=e.trim_offsets}static fromConfig(e){if(e===null)return null;switch(e.type){case"WordPiece":return new wb(e);case"Metaspace":return new kb(e);case"ByteLevel":return new bb(e);case"Replace":return new mb(e);case"ByteFallback":return new gb(e);case"Fuse":return new _b(e);case"Strip":return new yb(e);case"Sequence":return new $b(e);case"CTC":return new vb(e);case"BPEDecoder":return new xb(e);default:throw new Error(`Unknown Decoder type: ${e.type}`)}}_call(e){return this.decode(e)}decode(e){return this.decode_chain(e).join("")}decode_chain(e){throw Error("`decode_chain` should be implemented in subclass.")}}class mb extends Rt{decode_chain(e){const r=Qi(this.config.pattern);return r===null?e:e.map(n=>n.replaceAll(r,this.config.content))}}class gb extends Rt{constructor(e){super(e),this.text_decoder=new TextDecoder}decode_chain(e){const r=[];let n=[];for(const a of e){let i=null;if(a.length===6&&a.startsWith("<0x")&&a.endsWith(">")){const s=parseInt(a.slice(3,5),16);isNaN(s)||(i=s)}if(i!==null)n.push(i);else{if(n.length>0){const s=this.text_decoder.decode(Uint8Array.from(n));r.push(s),n=[]}r.push(a)}}if(n.length>0){const a=this.text_decoder.decode(Uint8Array.from(n));r.push(a),n=[]}return r}}class _b extends Rt{decode_chain(e){return[e.join("")]}}class yb extends Rt{constructor(e){super(e),this.content=this.config.content,this.start=this.config.start,this.stop=this.config.stop}decode_chain(e){return e.map(r=>{let n=0;for(let i=0;i<this.start&&r[i]===this.content;++i){n=i+1;continue}let a=r.length;for(let i=0;i<this.stop;++i){const s=r.length-i-1;if(r[s]===this.content){a=s;continue}else break}return r.slice(n,a)})}}class wb extends Rt{constructor(e){super(e),this.cleanup=e.cleanup}decode_chain(e){return e.map((r,n)=>(n!==0&&(r.startsWith(this.config.prefix)?r=r.replace(this.config.prefix,""):r=" "+r),this.cleanup&&(r=su(r)),r))}}class bb extends Rt{constructor(e){super(e),this.byte_decoder=Yw,this.text_decoder=new TextDecoder("utf-8",{fatal:!1,ignoreBOM:!0}),this.end_of_word_suffix=null}convert_tokens_to_string(e){const r=e.join(""),n=new Uint8Array([...r].map(i=>this.byte_decoder[i]));return this.text_decoder.decode(n)}decode_chain(e){const r=[];let n=[];for(const a of e)this.added_tokens.find(i=>i.content===a)!==void 0?(n.length>0&&(r.push(this.convert_tokens_to_string(n)),n=[]),r.push(a)):n.push(a);return n.length>0&&r.push(this.convert_tokens_to_string(n)),r}}class vb extends Rt{constructor(e){super(e),this.pad_token=this.config.pad_token,this.word_delimiter_token=this.config.word_delimiter_token,this.cleanup=this.config.cleanup}convert_tokens_to_string(e){if(e.length===0)return"";const r=[e[0]];for(let i=1;i<e.length;++i)e[i]!==r.at(-1)&&r.push(e[i]);let a=r.filter(i=>i!==this.pad_token).join("");return this.cleanup&&(a=su(a).replaceAll(this.word_delimiter_token," ").trim()),a}decode_chain(e){return[this.convert_tokens_to_string(e)]}}class $b extends Rt{constructor(e){super(e),this.decoders=e.decoders.map(r=>Rt.fromConfig(r))}decode_chain(e){return this.decoders.reduce((r,n)=>n.decode_chain(r),e)}}class xb extends Rt{constructor(e){super(e),this.suffix=this.config.suffix}decode_chain(e){return e.map((r,n)=>r.replaceAll(this.suffix,n===e.length-1?"":" "))}}class Sb extends Rt{decode_chain(e){let r="";for(let n=1;n<e.length;n+=2)r+=e[n];return[r]}}class yg extends Ut{constructor(e){super(),this.addPrefixSpace=e.add_prefix_space,this.replacement=e.replacement,this.strRep=e.str_rep||this.replacement,this.prepend_scheme=e.prepend_scheme??"always"}pre_tokenize_text(e,{section_index:r=void 0}={}){let n=e.replaceAll(" ",this.strRep);return this.addPrefixSpace&&!n.startsWith(this.replacement)&&(this.prepend_scheme==="always"||this.prepend_scheme==="first"&&r===0)&&(n=this.strRep+n),[n]}}class kb extends Rt{constructor(e){super(e),this.addPrefixSpace=e.add_prefix_space,this.replacement=e.replacement}decode_chain(e){const r=[];for(let n=0;n<e.length;++n){let a=e[n].replaceAll(this.replacement," ");this.addPrefixSpace&&n==0&&a.startsWith(" ")&&(a=a.substring(1)),r.push(a)}return r}}class Eb extends Pt{constructor(e){super(e),this.charsmap=e.precompiled_charsmap}normalize(e){return e=e.replace(/[\u0001-\u0008\u000B\u000E-\u001F\u007F\u008F\u009F]/gm,""),e=e.replace(/[\u0009\u000A\u000C\u000D\u1680\u200B\u200C\u200E\u200F\u2028\u2029\u2581\uFEFF\uFFFD]/gm," "),e.includes("～")?e=e.split("～").map(n=>n.normalize("NFKC")).join("～"):e=e.normalize("NFKC"),e}}class Cb extends Ut{constructor(e){super(),this.tokenizers=e.pretokenizers.map(r=>Ut.fromConfig(r))}pre_tokenize_text(e,r){return this.tokenizers.reduce((n,a)=>a.pre_tokenize(n,r),[e])}}class Tb extends Ut{constructor(e){super()}pre_tokenize_text(e,r){return e.match(/\w+|[^\w\s]+/g)||[]}}class Ib extends Ut{constructor(e){super()}pre_tokenize_text(e,r){return Gw(e)}}class Ab extends Ut{constructor(e){super(),this.config=e,this.pattern=Qi(this.config.pattern),this.content=this.config.content}pre_tokenize_text(e,r){return this.pattern===null?[e]:[e.replaceAll(this.pattern,this.config.content)]}}const Mb=["bos_token","eos_token","unk_token","sep_token","pad_token","cls_token","mask_token"];function Ob(t,e,r,n){for(const a of Object.keys(t)){const i=e-t[a].length,s=r(a),o=new Array(i).fill(s);t[a]=n==="right"?ct(t[a],o):ct(o,t[a])}}function zb(t,e){for(const r of Object.keys(t))t[r].length=e}class Ee extends $t{return_token_type_ids=!1;_default_chat_template=`{% for message in messages %}{{'<|im_start|>' + message['role'] + '
' + message['content'] + '<|im_end|>' + '
'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant
' }}{% endif %}`;padding_side="right";constructor(e,r){super(),this._tokenizer_config=r,this.normalizer=Pt.fromConfig(e.normalizer),this.pre_tokenizer=Ut.fromConfig(e.pre_tokenizer),this.model=ka.fromConfig(e.model,r),this.post_processor=Cn.fromConfig(e.post_processor),this.decoder=Rt.fromConfig(e.decoder),this.special_tokens=[],this.all_special_ids=[],this.added_tokens=[];for(const n of e.added_tokens){const a=new jw(n);this.added_tokens.push(a),this.model.tokens_to_ids.set(a.content,a.id),this.model.vocab[a.id]=a.content,a.special&&(this.special_tokens.push(a.content),this.all_special_ids.push(a.id))}if(this.additional_special_tokens=r.additional_special_tokens??[],this.special_tokens.push(...this.additional_special_tokens),this.special_tokens=[...new Set(this.special_tokens)],this.decoder&&(this.decoder.added_tokens=this.added_tokens,this.decoder.end_of_word_suffix=this.model.end_of_word_suffix),this.added_tokens_regex=this.added_tokens.length>0?new RegExp(this.added_tokens.map(n=>`${n.lstrip?"\\s*":""}(${Ip(n.content)})${n.rstrip?"\\s*":""}`).join("|")):null,this.mask_token=this.getToken("mask_token"),this.mask_token_id=this.model.tokens_to_ids.get(this.mask_token),this.pad_token=this.getToken("pad_token","eos_token"),this.pad_token_id=this.model.tokens_to_ids.get(this.pad_token),this.sep_token=this.getToken("sep_token"),this.sep_token_id=this.model.tokens_to_ids.get(this.sep_token),this.unk_token=this.getToken("unk_token"),this.unk_token_id=this.model.tokens_to_ids.get(this.unk_token),this.model_max_length=r.model_max_length,this.remove_space=r.remove_space,this.clean_up_tokenization_spaces=r.clean_up_tokenization_spaces??!0,this.do_lowercase_and_remove_accent=r.do_lowercase_and_remove_accent??!1,r.padding_side&&(this.padding_side=r.padding_side),this.legacy=!1,this.chat_template=r.chat_template??null,Array.isArray(this.chat_template)){const n=Object.create(null);for(const{name:a,template:i}of this.chat_template){if(typeof a!="string"||typeof i!="string")throw new Error('Chat template must be a list of objects with "name" and "template" properties');n[a]=i}this.chat_template=n}this._compiled_template_cache=new Map}getToken(...e){for(const r of e){const n=this._tokenizer_config[r];if(n)if(typeof n=="object"){if(n.__type==="AddedToken")return n.content;throw Error(`Unknown token: ${n}`)}else return n}return null}static async from_pretrained(e,{progress_callback:r=null,config:n=null,cache_dir:a=null,local_files_only:i=!1,revision:s="main",legacy:o=null}={}){const u=await cg(e,{progress_callback:r,config:n,cache_dir:a,local_files_only:i,revision:s,legacy:o});return new this(...u)}_call(e,{text_pair:r=null,add_special_tokens:n=!0,padding:a=!1,truncation:i=null,max_length:s=null,return_tensor:o=!0,return_token_type_ids:u=null}={}){const l=Array.isArray(e);let p;if(l){if(e.length===0)throw Error("text array must be non-empty");if(r!==null){if(Array.isArray(r)){if(e.length!==r.length)throw Error("text and text_pair must have the same length")}else throw Error("text_pair must also be an array");p=e.map((m,c)=>this._encode_plus(m,{text_pair:r[c],add_special_tokens:n,return_token_type_ids:u}))}else p=e.map(m=>this._encode_plus(m,{add_special_tokens:n,return_token_type_ids:u}))}else{if(e==null)throw Error("text may not be null or undefined");if(Array.isArray(r))throw Error("When specifying `text_pair`, since `text` is a string, `text_pair` must also be a string (i.e., not an array).");p=[this._encode_plus(e,{text_pair:r,add_special_tokens:n,return_token_type_ids:u})]}if(s===null?a==="max_length"?s=this.model_max_length:s=vr(p.map(m=>m.input_ids.length))[0]:i||console.warn("Truncation was not explicitly activated but `max_length` is provided a specific value, please use `truncation=true` to explicitly truncate examples to max length."),s=Math.min(s,this.model_max_length),a||i)for(let m=0;m<p.length;++m)p[m].input_ids.length!==s&&(p[m].input_ids.length>s?i&&zb(p[m],s):a&&Ob(p[m],s,c=>c==="input_ids"?this.pad_token_id:0,this.padding_side));const f={};if(o){if(!(a&&i)&&p.some(c=>{for(const y of Object.keys(c))if(c[y].length!==p[0][y]?.length)return!0;return!1}))throw Error("Unable to create tensor, you should probably activate truncation and/or padding with 'padding=true' and 'truncation=true' to have batched tensors with the same length.");const m=[p.length,p[0].input_ids.length];for(const c of Object.keys(p[0]))f[c]=new pe("int64",BigInt64Array.from(p.flatMap(y=>y[c]).map(BigInt)),m)}else{for(const m of Object.keys(p[0]))f[m]=p.map(c=>c[m]);if(!l)for(const m of Object.keys(f))f[m]=f[m][0]}return f}_encode_text(e){return e===null?null:(this.added_tokens_regex?e.split(this.added_tokens_regex).filter(a=>a):[e]).map((a,i)=>{if(this.added_tokens.find(o=>o.content===a)!==void 0)return a;{if(this.remove_space===!0&&(a=a.trim().split(/\s+/).join(" ")),this.do_lowercase_and_remove_accent&&(a=Uw(a)),this.normalizer!==null&&(a=this.normalizer(a)),a.length===0)return[];const o=this.pre_tokenizer!==null?this.pre_tokenizer(a,{section_index:i}):[a];return this.model(o)}}).flat()}_encode_plus(e,{text_pair:r=null,add_special_tokens:n=!0,return_token_type_ids:a=null}={}){const{tokens:i,token_type_ids:s}=this._tokenize_helper(e,{pair:r,add_special_tokens:n}),o=this.model.convert_tokens_to_ids(i),u={input_ids:o,attention_mask:new Array(o.length).fill(1)};return(a??this.return_token_type_ids)&&s&&(u.token_type_ids=s),u}_tokenize_helper(e,{pair:r=null,add_special_tokens:n=!1}={}){const a=this._encode_text(e),i=this._encode_text(r);return this.post_processor?this.post_processor(a,i,{add_special_tokens:n}):{tokens:ct(a??[],i??[])}}tokenize(e,{pair:r=null,add_special_tokens:n=!1}={}){return this._tokenize_helper(e,{pair:r,add_special_tokens:n}).tokens}encode(e,{text_pair:r=null,add_special_tokens:n=!0,return_token_type_ids:a=null}={}){return this._encode_plus(e,{text_pair:r,add_special_tokens:n,return_token_type_ids:a}).input_ids}batch_decode(e,r={}){return e instanceof pe&&(e=e.tolist()),e.map(n=>this.decode(n,r))}decode(e,r={}){if(e instanceof pe&&(e=pg(e)),!Array.isArray(e)||e.length===0||!c_(e[0]))throw Error("token_ids must be a non-empty array of integers.");return this.decode_single(e,r)}decode_single(e,{skip_special_tokens:r=!1,clean_up_tokenization_spaces:n=null}){let a=this.model.convert_ids_to_tokens(e);r&&(a=a.filter(s=>!this.special_tokens.includes(s)));let i=this.decoder?this.decoder(a):a.join(" ");return this.decoder&&this.decoder.end_of_word_suffix&&(i=i.replaceAll(this.decoder.end_of_word_suffix," "),r&&(i=i.trim())),(n??this.clean_up_tokenization_spaces)&&(i=su(i)),i}get default_chat_template(){return this._warned_about_chat_template||(console.warn("No chat template is defined for this tokenizer - using a default chat template that implements the ChatML format. If the default is not appropriate for your model, please set `tokenizer.chat_template` to an appropriate template. See https://huggingface.co/docs/transformers/main/chat_templating for more information."),this._warned_about_chat_template=!0),this._default_chat_template}apply_chat_template(e,{chat_template:r=null,add_generation_prompt:n=!1,tokenize:a=!0,padding:i=!1,truncation:s=!1,max_length:o=null,return_tensor:u=!0,return_dict:l=!1,tokenizer_kwargs:p={},...f}={}){if(this.chat_template&&typeof this.chat_template=="object"||this.chat_template===null&&this.default_chat_template&&typeof this.default_chat_template=="object"){const w=this.chat_template??this.default_chat_template;if(r!==null&&Object.hasOwn(w,r))r=w[r];else if(r===null&&"default"in w)r=w.default;else if(r===null)throw Error(`This model has multiple chat templates with no default specified! Please either pass a chat template or the name of the template you wish to use to the 'chat_template' argument. Available template names are ${Object.keys(w).sort()}.`)}else r??=this.chat_template??this.default_chat_template;if(typeof r!="string")throw Error(`chat_template must be a string, but got ${typeof r}`);let m=this._compiled_template_cache.get(r);m===void 0&&(m=new Nw(r),this._compiled_template_cache.set(r,m));const c=Object.create(null);for(const w of Mb){const v=this.getToken(w);v&&(c[w]=v)}const y=m.render({messages:e,add_generation_prompt:n,...c,...f});if(a){const w=this._call(y,{add_special_tokens:!1,padding:i,truncation:s,max_length:o,return_tensor:u,...p});return l?w:w.input_ids}return y}}class Pb extends Ee{return_token_type_ids=!0}class Rb extends Ee{return_token_type_ids=!0}class Bb extends Ee{return_token_type_ids=!0}class Db extends Ee{return_token_type_ids=!0}class Nb extends Ee{return_token_type_ids=!0}class Fb extends Ee{return_token_type_ids=!0}class Lb extends Ee{return_token_type_ids=!0}class Ub extends Ee{return_token_type_ids=!0}class Wb extends Ee{return_token_type_ids=!0}class Vb extends Ee{}class Gb extends Ee{}class Hb extends Ee{return_token_type_ids=!0;constructor(e,r){super(e,r),console.warn('WARNING: `XLMTokenizer` is not yet supported by Hugging Face\'s "fast" tokenizers library. Therefore, you may experience slightly inaccurate results.')}}class jb extends Ee{return_token_type_ids=!0}class qb extends Ee{}class wg extends Ee{_default_chat_template='{% for message in messages %}" "{{ message.content }}{{ eos_token }}" "{% endfor %}'}class Kb extends Ee{}class bg extends Ee{constructor(e,r){super(e,r),this.languageRegex=/^[a-z]{2}_[A-Z]{2}$/,this.language_codes=this.special_tokens.filter(n=>this.languageRegex.test(n)),this.lang_to_token=n=>n}_build_translation_inputs(e,r,n){return ou(this,e,r,n)}}class Yb extends bg{}class Xb extends Ee{}class Qb extends wg{constructor(e,r){const n=".,!?…。，、।۔،",a=e.pre_tokenizer?.pretokenizers[0]?.pattern;a&&a.Regex===` ?[^(\\s|[${n}])]+`&&(a.Regex=` ?[^\\s${n}]+`),super(e,r)}}const bi="▁";class vg extends Ee{_default_chat_template=`{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif USE_DEFAULT_PROMPT == true and not '<<SYS>>' in messages[0]['content'] %}{% set loop_messages = messages %}{% set system_message = 'DEFAULT_SYSTEM_MESSAGE' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if loop.index0 == 0 and system_message != false %}{% set content = '<<SYS>>
' + system_message + '
<</SYS>>

' + message['content'] %}{% else %}{% set content = message['content'] %}{% endif %}{% if message['role'] == 'user' %}{{ bos_token + '[INST] ' + content.strip() + ' [/INST]' }}{% elif message['role'] == 'system' %}{{ '<<SYS>>
' + content.strip() + '
<</SYS>>

' }}{% elif message['role'] == 'assistant' %}{{ ' '  + content.strip() + ' ' + eos_token }}{% endif %}{% endfor %}`;DEFAULT_SYSTEM_PROMPT=`You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.

If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.`;padding_side="left";constructor(e,r){super(e,r),this.use_default_system_prompt=r.use_default_system_prompt??!1,this.legacy=r.legacy??!0,this.legacy||(this.normalizer=null,this.pre_tokenizer=new yg({replacement:bi,add_prefix_space:!0,prepend_scheme:"first"}))}_encode_text(e){if(e===null)return null;if(this.legacy||e.length===0)return super._encode_text(e);let r=super._encode_text(bi+e.replaceAll(bi," "));return r.length>1&&r[0]===bi&&this.special_tokens.includes(r[1])&&(r=r.slice(1)),r}get default_chat_template(){return super.default_chat_template.replaceAll("USE_DEFAULT_PROMPT",this.use_default_system_prompt?"true":"false").replaceAll("DEFAULT_SYSTEM_MESSAGE",this.DEFAULT_SYSTEM_PROMPT.replaceAll(`
`,"\\n").replaceAll("'","\\'"))}}class Jb extends vg{}class Zb extends Ee{}class ev extends Ee{}class tv extends Ee{}class rv extends Ee{}class nv extends Ee{}class av extends Ee{}class iv extends Ee{_default_chat_template=`{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '<start_of_turn>' + role + '
' + message['content'] | trim + '<end_of_turn>
' }}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model
'}}{% endif %}`}class sv extends Ee{}function ou(t,e,r,n){if(!("language_codes"in t)||!Array.isArray(t.language_codes))throw new Error("Tokenizer must have `language_codes` attribute set and it should be an array of language ids.");if(!("languageRegex"in t)||!(t.languageRegex instanceof RegExp))throw new Error("Tokenizer must have `languageRegex` attribute set and it should be a regular expression.");if(!("lang_to_token"in t)||typeof t.lang_to_token!="function")throw new Error("Tokenizer must have `lang_to_token` attribute set and it should be a function.");const a=n.src_lang,i=n.tgt_lang;if(!t.language_codes.includes(i))throw new Error(`Target language code "${i}" is not valid. Must be one of: {${t.language_codes.join(", ")}}`);if(a!==void 0){if(!t.language_codes.includes(a))throw new Error(`Source language code "${a}" is not valid. Must be one of: {${t.language_codes.join(", ")}}`);for(const s of t.post_processor.config.single)if("SpecialToken"in s&&t.languageRegex.test(s.SpecialToken.id)){s.SpecialToken.id=t.lang_to_token(a);break}}return n.forced_bos_token_id=t.model.convert_tokens_to_ids([t.lang_to_token(i)])[0],t._call(e,r)}class ov extends Ee{constructor(e,r){super(e,r),this.languageRegex=/^[a-z]{3}_[A-Z][a-z]{3}$/,this.language_codes=this.special_tokens.filter(n=>this.languageRegex.test(n)),this.lang_to_token=n=>n}_build_translation_inputs(e,r,n){return ou(this,e,r,n)}}class uv extends Ee{constructor(e,r){super(e,r),this.languageRegex=/^__[a-z]{2,3}__$/,this.language_codes=this.special_tokens.filter(n=>this.languageRegex.test(n)).map(n=>n.slice(2,-2)),this.lang_to_token=n=>`__${n}__`}_build_translation_inputs(e,r,n){return ou(this,e,r,n)}}class lv extends Ee{_default_chat_template='{% for message in messages %}" "{{ message.content }}{{ eos_token }}" "{% endfor %}';get timestamp_begin(){return this.model.convert_tokens_to_ids(["<|notimestamps|>"])[0]+1}_decode_asr(e,{return_timestamps:r=!1,return_language:n=!1,time_precision:a=null,force_full_sequences:i=!0}={}){if(a===null)throw Error("Must specify time_precision");let s=null;const o=r==="word";function u(){return{language:s,timestamp:[null,null],text:""}}const l=[];let p=u(),f=0;const m=this.timestamp_begin;let c=[],y=[],w=!1,v=null;const k=new Set(this.all_special_ids);for(const T of e){const A=T.tokens,B=o?T.token_timestamps:null;let R=null,D=m;if("stride"in T){const[ie,te,oe]=T.stride;if(f-=te,v=ie-oe,te&&(D=te/a+m),oe)for(let re=A.length-1;re>=0;--re){const M=Number(A[re]);if(M>=m){if(R!==null&&(M-m)*a<v)break;R=M}}}let K=[],j=[];for(let ie=0;ie<A.length;++ie){const te=Number(A[ie]);if(k.has(te)){const oe=this.decode([te]),re=Oi.get(oe.slice(2,-2));if(re!==void 0){if(s!==null&&re!==s&&!r){c.push(K);const M=this.findLongestCommonSequence(c)[0],P=this.decode(M);p.text=P,l.push(p),c=[],K=[],p=u()}s=p.language=re}}else if(te>=m){const oe=(te-m)*a+f,re=Ss(oe,2);if(R!==null&&te>=R)w=!0;else if(w||c.length>0&&te<D)w=!1;else if(p.timestamp[0]===null)p.timestamp[0]=re;else if(re!==p.timestamp[0]){p.timestamp[1]=re,c.push(K),o&&y.push(j);const[M,P]=this.findLongestCommonSequence(c,y),H=this.decode(M);p.text=H,o&&(p.words=this.collateWordTimestamps(M,P,s)),l.push(p),c=[],K=[],y=[],j=[],p=u()}}else if(K.push(te),o){let oe=Ss(B[ie]+f,2),re;ie+1<B.length?re=Ss(B[ie+1]+f,2):re=null,j.push([oe,re])}}if("stride"in T){const[ie,te,oe]=T.stride;f+=ie-oe}K.length>0?(c.push(K),o&&y.push(j)):c.every(ie=>ie.length===0)&&(p=u(),c=[],K=[],y=[],j=[])}if(c.length>0){if(i&&r)throw new Error("Whisper did not predict an ending timestamp, which can happen if audio is cut off in the middle of a word. Also make sure WhisperTimeStampLogitsProcessor was used during generation.");const[T,A]=this.findLongestCommonSequence(c,y),B=this.decode(T);p.text=B,o&&(p.words=this.collateWordTimestamps(T,A,s)),l.push(p)}let $=Object.create(null);const C=l.map(T=>T.text).join("");if(r||n){for(let T=0;T<l.length;++T){const A=l[T];r||delete A.timestamp,n||delete A.language}if(o){const T=[];for(const A of l)for(const B of A.words)T.push(B);$={chunks:T}}else $={chunks:l}}return[C,$]}findLongestCommonSequence(e,r=null){let n=e[0],a=n.length,i=[];const s=Array.isArray(r)&&r.length>0;let o=s?[]:null,u=s?r[0]:null;for(let l=1;l<e.length;++l){const p=e[l];let f=0,m=[a,a,0,0];const c=p.length;for(let T=1;T<a+c;++T){const A=T/1e4,B=Math.max(0,a-T),R=Math.min(a,a+c-T),D=n.slice(B,R),K=Math.max(0,T-a),j=Math.min(c,T),ie=p.slice(K,j);if(D.length!==ie.length)throw new Error("There is a bug within whisper `decode_asr` function, please report it. Dropping to prevent bad inference.");const te=D.filter((re,M)=>re===ie[M]).length,oe=te/T+A;te>1&&oe>f&&(f=oe,m=[B,R,K,j])}const[y,w,v,k]=m,$=Math.floor((w+y)/2),C=Math.floor((k+v)/2);i.push(...n.slice(0,$)),n=p.slice(C),a=n.length,s&&(o.push(...u.slice(0,$)),u=r[l].slice(C))}return i.push(...n),s?(o.push(...u),[i,o]):[i,[]]}collateWordTimestamps(e,r,n){const[a,i,s]=this.combineTokensIntoWords(e,n),o=[];for(let u=0;u<a.length;++u){const l=s[u];o.push({text:a[u],timestamp:[r[l.at(0)][0],r[l.at(-1)][1]]})}return o}combineTokensIntoWords(e,r,n=`"'“¡¿([{-`,a=`"'.。,，!！?？:：”)]}、`){r=r??"english";let i,s,o;return["chinese","japanese","thai","lao","myanmar"].includes(r)?[i,s,o]=this.splitTokensOnUnicode(e):[i,s,o]=this.splitTokensOnSpaces(e),this.mergePunctuations(i,s,o,n,a)}decode(e,r){let n;return r?.decode_with_timestamps?(e instanceof pe&&(e=pg(e)),n=this.decodeWithTimestamps(e,r)):n=super.decode(e,r),n}decodeWithTimestamps(e,r){const n=r?.time_precision??.02,a=Array.from(this.all_special_ids).at(-1)+1;let i=[[]];for(let s of e)if(s=Number(s),s>=a){const o=((s-a)*n).toFixed(2);i.push(`<|${o}|>`),i.push([])}else i[i.length-1].push(s);return i=i.map(s=>typeof s=="string"?s:super.decode(s,r)),i.join("")}splitTokensOnUnicode(e){const r=this.decode(e,{decode_with_timestamps:!0}),n="�",a=[],i=[],s=[];let o=[],u=[],l=0;for(let p=0;p<e.length;++p){const f=e[p];o.push(f),u.push(p);const m=this.decode(o,{decode_with_timestamps:!0});(!m.includes(n)||r[l+m.indexOf(n)]===n)&&(a.push(m),i.push(o),s.push(u),o=[],u=[],l+=m.length)}return[a,i,s]}splitTokensOnSpaces(e){const[r,n,a]=this.splitTokensOnUnicode(e),i=[],s=[],o=[],u=new RegExp(`^[${ya}]$`,"gu");for(let l=0;l<r.length;++l){const p=r[l],f=n[l],m=a[l],c=f[0]>=this.model.tokens_to_ids.get("<|endoftext|>"),y=p.startsWith(" "),w=p.trim(),v=u.test(w);if(c||y||v||i.length===0)i.push(p),s.push(f),o.push(m);else{const k=i.length-1;i[k]+=p,s[k].push(...f),o[k].push(...m)}}return[i,s,o]}mergePunctuations(e,r,n,a,i){const s=structuredClone(e),o=structuredClone(r),u=structuredClone(n);let l=s.length-2,p=s.length-1;for(;l>=0;)s[l].startsWith(" ")&&a.includes(s[l].trim())?(s[p]=s[l]+s[p],o[p]=ct(o[l],o[p]),u[p]=ct(u[l],u[p]),s[l]="",o[l]=[],u[l]=[]):p=l,--l;for(l=0,p=1;p<s.length;)!s[l].endsWith(" ")&&i.includes(s[p])?(s[l]+=s[p],o[l]=ct(o[l],o[p]),u[l]=ct(u[l],u[p]),s[p]="",o[p]=[],u[p]=[]):l=p,++p;return[s.filter(f=>f),o.filter(f=>f.length>0),u.filter(f=>f.length>0)]}get_decoder_prompt_ids({language:e=null,task:r=null,no_timestamps:n=!0}={}){const a=[];if(e){const i=dg(e),s=this.model.tokens_to_ids.get(`<|${i}|>`);if(s===void 0)throw new Error(`Unable to find language "${i}" in model vocabulary. Please report this issue at ${io}.`);a.push(s)}else a.push(null);if(r){if(r=r.toLowerCase(),r!=="transcribe"&&r!=="translate")throw new Error(`Task "${r}" is not supported. Must be one of: ["transcribe", "translate"]`);const i=this.model.tokens_to_ids.get(`<|${r}|>`);if(i===void 0)throw new Error(`Unable to find task "${r}" in model vocabulary. Please report this issue at ${io}.`);a.push(i)}else a.push(null);if(n){const i=this.model.tokens_to_ids.get("<|notimestamps|>");if(i===void 0)throw new Error(`Unable to find "<|notimestamps|>" in model vocabulary. Please report this issue at ${io}.`);a.push(i)}return a.map((i,s)=>[s+1,i]).filter(i=>i[1]!==null)}}class dv extends Ee{}class cv extends Ee{}class pv extends Ee{}class hv extends Ee{constructor(e,r){super(e,r),this.languageRegex=/^(>>\w+<<)\s*/g,this.supported_language_codes=this.model.vocab.filter(n=>this.languageRegex.test(n)),console.warn('WARNING: `MarianTokenizer` is not yet supported by Hugging Face\'s "fast" tokenizers library. Therefore, you may experience slightly inaccurate results.')}_encode_text(e){if(e===null)return null;const[r,...n]=e.trim().split(this.languageRegex);if(n.length===0)return super._encode_text(r);if(n.length===2){const[a,i]=n;return this.supported_language_codes.includes(a)||console.warn(`Unsupported language code "${a}" detected, which may lead to unexpected behavior. Should be one of: ${JSON.stringify(this.supported_language_codes)}`),ct([a],super._encode_text(i))}}}class fv extends Ee{}class $g extends Ee{_default_chat_template="{% for message in messages %}{% if message['role'] == 'user' %}{{ ' ' }}{% endif %}{{ message['content'] }}{% if not loop.last %}{{ '  ' }}{% endif %}{% endfor %}{{ eos_token }}"}class mv extends $g{}class gv extends Ee{}class _v extends Ee{}class yv extends Ee{constructor(e,r){super(e,r),this.decoder=new Sb({})}}class wv extends Ee{}class bv{static TOKENIZER_CLASS_MAPPING={T5Tokenizer:qb,DistilBertTokenizer:Vb,CamembertTokenizer:Gb,DebertaTokenizer:Nb,DebertaV2Tokenizer:Fb,BertTokenizer:Pb,HerbertTokenizer:Lb,ConvBertTokenizer:Ub,RoFormerTokenizer:Wb,XLMTokenizer:Hb,ElectraTokenizer:jb,MobileBertTokenizer:Bb,SqueezeBertTokenizer:Db,AlbertTokenizer:Rb,GPT2Tokenizer:wg,BartTokenizer:Kb,MBartTokenizer:bg,MBart50Tokenizer:Yb,RobertaTokenizer:Xb,WhisperTokenizer:lv,CodeGenTokenizer:dv,CLIPTokenizer:cv,SiglipTokenizer:pv,MarianTokenizer:hv,BloomTokenizer:Qb,NllbTokenizer:ov,M2M100Tokenizer:uv,LlamaTokenizer:vg,CodeLlamaTokenizer:Jb,XLMRobertaTokenizer:Zb,MPNetTokenizer:ev,FalconTokenizer:tv,GPTNeoXTokenizer:rv,EsmTokenizer:nv,Wav2Vec2CTCTokenizer:fv,BlenderbotTokenizer:$g,BlenderbotSmallTokenizer:mv,SpeechT5Tokenizer:gv,NougatTokenizer:_v,VitsTokenizer:yv,Qwen2Tokenizer:av,GemmaTokenizer:iv,Grok1Tokenizer:sv,CohereTokenizer:wv,PreTrainedTokenizer:Ee};static async from_pretrained(e,{progress_callback:r=null,config:n=null,cache_dir:a=null,local_files_only:i=!1,revision:s="main",legacy:o=null}={}){const[u,l]=await cg(e,{progress_callback:r,config:n,cache_dir:a,local_files_only:i,revision:s,legacy:o}),p=l.tokenizer_class?.replace(/Fast$/,"")??"PreTrainedTokenizer";let f=this.TOKENIZER_CLASS_MAPPING[p];return f||(console.warn(`Unknown tokenizer class "${p}", attempting to construct from base class.`),f=Ee),new f(u,l)}}async function vv(t,e){return await Ar(t,"config.json",!0,e)}function $n(t){const e={};let r={};switch(t.model_type){case"llava":case"paligemma":r=$n(t.text_config);break;case"moondream1":r=$n(t.phi_config);break;case"musicgen":r=$n(t.decoder);break;case"gpt2":case"gptj":case"codegen":case"gpt_bigcode":e.num_heads="n_head",e.num_layers="n_layer",e.hidden_size="n_embd";break;case"gpt_neox":case"stablelm":case"opt":case"phi":case"phi3":case"falcon":e.num_heads="num_attention_heads",e.num_layers="num_hidden_layers",e.hidden_size="hidden_size";break;case"llama":case"mistral":case"starcoder2":case"qwen2":e.num_heads="num_key_value_heads",e.num_layers="num_hidden_layers",e.hidden_size="hidden_size",e.num_attention_heads="num_attention_heads";break;case"gemma":e.num_heads="num_key_value_heads",e.num_layers="num_hidden_layers",e.dim_kv="head_dim";break;case"openelm":e.num_heads="num_kv_heads",e.num_layers="num_transformer_layers",e.dim_kv="head_dim";break;case"gpt_neo":e.num_heads="num_heads",e.num_layers="num_layers",e.hidden_size="hidden_size";break;case"bloom":e.num_heads="n_head",e.num_layers="n_layer",e.hidden_size="hidden_size";break;case"mpt":e.num_heads="n_heads",e.num_layers="n_layers",e.hidden_size="d_model";break;case"t5":case"mt5":case"longt5":e.num_decoder_layers="num_decoder_layers",e.num_decoder_heads="num_heads",e.decoder_dim_kv="d_kv",e.num_encoder_layers="num_layers",e.num_encoder_heads="num_heads",e.encoder_dim_kv="d_kv";break;case"bart":case"mbart":case"marian":case"whisper":case"m2m_100":case"blenderbot":case"blenderbot-small":e.num_decoder_layers="decoder_layers",e.num_decoder_heads="decoder_attention_heads",e.decoder_hidden_size="d_model",e.num_encoder_layers="encoder_layers",e.num_encoder_heads="encoder_attention_heads",e.encoder_hidden_size="d_model";break;case"speecht5":e.num_decoder_layers="decoder_layers",e.num_decoder_heads="decoder_attention_heads",e.decoder_hidden_size="hidden_size",e.num_encoder_layers="encoder_layers",e.num_encoder_heads="encoder_attention_heads",e.encoder_hidden_size="hidden_size";break;case"trocr":e.num_encoder_layers=e.num_decoder_layers="decoder_layers",e.num_encoder_heads=e.num_decoder_heads="decoder_attention_heads",e.encoder_hidden_size=e.decoder_hidden_size="d_model";break;case"musicgen_decoder":e.num_encoder_layers=e.num_decoder_layers="num_hidden_layers",e.num_encoder_heads=e.num_decoder_heads="num_attention_heads",e.encoder_hidden_size=e.decoder_hidden_size="hidden_size";break;case"vision-encoder-decoder":const a=$n(t.encoder),i=$n(t.decoder),s="num_decoder_layers"in i,o={};return s?(o.num_decoder_layers=i.num_layers,o.num_decoder_heads=i.num_heads,o.decoder_hidden_size=i.hidden_size,o.num_encoder_layers=a.num_layers,o.num_encoder_heads=a.num_heads,o.encoder_hidden_size=a.hidden_size):(o.num_layers=i.num_layers,o.num_heads=i.num_heads,o.hidden_size=i.hidden_size),o}const n={...r,...Xr(t,["model_type","multi_query","is_encoder_decoder"])};for(const a in e)n[a]=t[e[a]];return n}function xg(t,{prefix:e="past_key_values",encoder_add_pkv:r=!0}={}){const n={},a=t.normalized_config,i=1;if(a.is_encoder_decoder&&r){const s=a.encoder_dim_kv??a.encoder_hidden_size/a.num_encoder_heads,o=a.decoder_dim_kv??a.decoder_hidden_size/a.num_decoder_heads,u=[i,a.num_encoder_heads,0,s],l=[i,a.num_decoder_heads,0,o];for(let p=0;p<a.num_decoder_layers;++p)n[`${e}.${p}.encoder.key`]=u,n[`${e}.${p}.encoder.value`]=u,n[`${e}.${p}.decoder.key`]=l,n[`${e}.${p}.decoder.value`]=l}else{const s=a.num_heads,o=a.num_layers,u=a.dim_kv??a.hidden_size/(a.num_attention_heads??s);if(a.model_type==="falcon"){const l=[i*s,0,u];for(let p=0;p<o;++p)n[`${e}.${p}.key`]=l,n[`${e}.${p}.value`]=l}else if(a.multi_query){const l=[i*s,0,2*u];for(let p=0;p<o;++p)n[`${e}.${p}.key_value`]=l}else if(a.model_type==="bloom"){const l=[i*s,u,0],p=[i*s,0,u];for(let f=0;f<o;++f)n[`${e}.${f}.key`]=l,n[`${e}.${f}.value`]=p}else if(a.model_type==="openelm")for(let l=0;l<o;++l){const p=[i,s[l],0,u];n[`${e}.${l}.key`]=p,n[`${e}.${l}.value`]=p}else{const l=[i,s,0,u];for(let p=0;p<o;++p)n[`${e}.${p}.key`]=l,n[`${e}.${p}.value`]=l}}return n}class ji{max_position_embeddings;constructor(e){this.model_type=null,this.is_encoder_decoder=!1,Object.assign(this,e),this.normalized_config=$n(this)}static async from_pretrained(e,{progress_callback:r=null,config:n=null,cache_dir:a=null,local_files_only:i=!1,revision:s="main"}={}){n&&!(n instanceof ji)&&(n=new ji(n));const o=n??await vv(e,{progress_callback:r,config:n,cache_dir:a,local_files_only:i,revision:s});return new this(o)}}class Sg{static async from_pretrained(...e){return ji.from_pretrained(...e)}}const vi=Object.freeze({cpu:"cpu",gpu:"gpu",wasm:"wasm",webgpu:"webgpu"}),$v=function(){let t;return async function(){if(t===void 0)if(Jr.IS_NODE_ENV)t=!0;else if(!Jr.IS_WEBGPU_AVAILABLE)t=!1;else try{t=(await navigator.gpu.requestAdapter()).features.has("shader-f16")}catch{t=!1}return t}}(),It=Object.freeze({fp32:"fp32",fp16:"fp16",q8:"q8",int8:"int8",uint8:"uint8",q4:"q4",bnb4:"bnb4",q4f16:"q4f16"}),xv=Object.freeze({[vi.cpu]:It.q8,[vi.gpu]:It.fp32,[vi.wasm]:It.q8,[vi.webgpu]:It.fp32}),fp=Object.freeze({[It.fp32]:"",[It.fp16]:"_fp16",[It.int8]:"_int8",[It.uint8]:"_uint8",[It.q8]:"_quantized",[It.q4]:"_q4",[It.q4f16]:"_q4f16",[It.bnb4]:"_bnb4"});class pr extends $t{_call(e,r){throw Error("`_call` should be implemented in a subclass")}}class uu extends $t{_call(e,r){throw Error("`_call` should be implemented in a subclass")}}class Ao extends $t{constructor(){super(),this.processors=[]}push(e){this.processors.push(e)}extend(e){this.processors.push(...e)}_call(e,r){let n=r;for(const a of this.processors)n=a(e,n);return n}[Symbol.iterator](){return this.processors.values()}}class Sv extends pr{constructor(e){super(),this.bos_token_id=e}_call(e,r){for(let n=0;n<e.length;++n)if(e[n].length===1){const a=r[n];a.data.fill(-1/0),a.data[this.bos_token_id]=0}return r}}class kv extends pr{constructor(e,r){super(),this.max_length=e,this.forced_eos_token_id=r}_call(e,r){}}class Ev extends pr{constructor(e,r){super(),this.begin_suppress_tokens=e,this.begin_index=r}_call(e,r){for(let n=0;n<e.length;++n)if(e[n].length===this.begin_index){const a=r[n];for(const i of this.begin_suppress_tokens)a.data[i]=-1/0}return r}}class Cv extends pr{constructor(e,r){super(),this.eos_token_id=Array.isArray(e.eos_token_id)?e.eos_token_id[0]:e.eos_token_id,this.no_timestamps_token_id=e.no_timestamps_token_id,this.timestamp_begin=this.no_timestamps_token_id+1,this.begin_index=r.length,r.at(-1)===this.no_timestamps_token_id&&(this.begin_index-=1),this.max_initial_timestamp_index=e.max_initial_timestamp_index}_call(e,r){for(let n=0;n<e.length;++n){const i=r[n].data;if(i[this.no_timestamps_token_id]=-1/0,e[n].length===this.begin_index-1){i.fill(-1/0),i[this.timestamp_begin]=0;continue}const s=e[n].slice(this.begin_index),o=s.length>=1&&s[s.length-1]>=this.timestamp_begin,u=s.length<2||s[s.length-2]>=this.timestamp_begin;if(o&&(u?i.subarray(this.timestamp_begin).fill(-1/0):i.subarray(0,this.eos_token_id).fill(-1/0)),e[n].length===this.begin_index&&this.max_initial_timestamp_index!==null){const m=this.timestamp_begin+this.max_initial_timestamp_index;i.subarray(m+1).fill(-1/0)}const l=y_(i),p=Math.log(l.subarray(this.timestamp_begin).map(Math.exp).reduce((m,c)=>m+c)),f=vr(l.subarray(0,this.timestamp_begin))[0];p>f&&i.subarray(0,this.timestamp_begin).fill(-1/0)}return r}}class Tv extends pr{constructor(e){super(),this.no_repeat_ngram_size=e}getNgrams(e){const r=e.length,n=[];for(let i=0;i<r+1-this.no_repeat_ngram_size;++i){const s=[];for(let o=0;o<this.no_repeat_ngram_size;++o)s.push(e[i+o]);n.push(s)}const a=new Map;for(const i of n){const s=i.slice(0,i.length-1),o=JSON.stringify(s),u=a.get(o)??[];u.push(i[i.length-1]),a.set(o,u)}return a}getGeneratedNgrams(e,r){const n=r.slice(r.length+1-this.no_repeat_ngram_size,r.length);return e.get(JSON.stringify(n))??[]}calcBannedNgramTokens(e){const r=[];if(e.length+1<this.no_repeat_ngram_size)return r;{const n=this.getNgrams(e);return this.getGeneratedNgrams(n,e)}}_call(e,r){for(let n=0;n<e.length;++n){const a=r[n],i=this.calcBannedNgramTokens(e[n]);for(const s of i)a.data[s]=-1/0}return r}}class Iv extends pr{constructor(e){super(),this.penalty=e}_call(e,r){for(let n=0;n<e.length;++n){const a=r[n];for(const i of e[n])a.data[i]<0?a.data[i]*=this.penalty:a.data[i]/=this.penalty}return r}}class Av extends pr{constructor(e,r){super(),this.min_length=e,this.eos_token_id=Array.isArray(r)?r:[r]}_call(e,r){for(let n=0;n<e.length;++n)if(e[n].length<this.min_length){const a=r[n];for(const i of this.eos_token_id)a.data[i]=-1/0}return r}}class Mv extends pr{constructor(e,r,n){super(),this.prompt_length_to_skip=e,this.min_new_tokens=r,this.eos_token_id=Array.isArray(n)?n:[n]}_call(e,r){for(let n=0;n<e.length;++n)if(e[n].length-this.prompt_length_to_skip<this.min_new_tokens){const i=r[n];for(const s of this.eos_token_id)i[s]=-1/0}return r}}class Ov extends pr{constructor(e,r){super(),this.bad_words_ids=e,this.eos_token_id=Array.isArray(r)?r:[r]}_call(e,r){for(let n=0;n<e.length;++n){const a=r[n];for(const i of this.bad_words_ids){let s=!0;for(let o=1;o<=i.length-1&&i.length<e[o].length;++o)if(i.at(-o-1)!==e[o].at(-o)){s=!1;break}s&&(a[i.at(-1)]=-1/0)}}return r}}class zv extends pr{constructor(e){if(super(),e<=1)throw new Error(`Require guidance scale >1 to use the classifier free guidance processor, got guidance scale ${e}.`);this.guidance_scale=e}_call(e,r){if(r.dims[0]!==2*e.length)throw new Error(`Logits should have twice the batch size of the input ids, the first half of batches corresponding to the conditional inputs, and the second half of batches corresponding to the unconditional inputs. Got batch size ${r.dims[0]} for the logits and ${e.length} for the input ids.`);const n=e.length,a=r.slice([0,n],null),i=r.slice([n,r.dims[0]],null);for(let s=0;s<i.data.length;++s)i.data[s]+=(a.data[s]-i.data[s])*this.guidance_scale;return i}}class Pv extends uu{constructor(e){super(),this.temperature=e}_call(e,r){const n=r.data;for(let a=0;a<n.length;++a)n[a]/=this.temperature;return r}}class Rv extends uu{constructor(e,{filter_value:r=-1/0,min_tokens_to_keep:n=1}={}){if(super(),e<0||e>1)throw new Error(`\`top_p\` must be a float > 0 and < 1, but is ${e}`);if(!Number.isInteger(n)||n<1)throw new Error(`\`min_tokens_to_keep\` must be a positive integer, but is ${n}`);this.top_p=e,this.filter_value=r,this.min_tokens_to_keep=n}}class Bv extends uu{constructor(e,{filter_value:r=-1/0,min_tokens_to_keep:n=1}={}){if(super(),!Number.isInteger(e)||e<0)throw new Error(`\`top_k\` must be a positive integer, but is ${e}`);this.top_k=Math.max(e,n),this.filter_value=r}}class kg{max_length=20;max_new_tokens=null;min_length=0;min_new_tokens=null;early_stopping=!1;max_time=null;do_sample=!1;num_beams=1;num_beam_groups=1;penalty_alpha=null;use_cache=!0;temperature=1;top_k=50;top_p=1;typical_p=1;epsilon_cutoff=0;eta_cutoff=0;diversity_penalty=0;repetition_penalty=1;encoder_repetition_penalty=1;length_penalty=1;no_repeat_ngram_size=0;bad_words_ids=null;force_words_ids=null;renormalize_logits=!1;constraints=null;forced_bos_token_id=null;forced_eos_token_id=null;remove_invalid_values=!1;exponential_decay_length_penalty=null;suppress_tokens=null;begin_suppress_tokens=null;forced_decoder_ids=null;guidance_scale=null;num_return_sequences=1;output_attentions=!1;output_hidden_states=!1;output_scores=!1;return_dict_in_generate=!1;pad_token_id=null;bos_token_id=null;eos_token_id=null;encoder_no_repeat_ngram_size=0;decoder_start_token_id=null;generation_kwargs={};constructor(e){Object.assign(this,Xr(e,Object.getOwnPropertyNames(this)))}}class lu extends $t{_call(e,r){throw Error("StoppingCriteria needs to be subclassed")}}class du extends $t{constructor(){super(),this.criteria=[]}push(e){this.criteria.push(e)}extend(e){e instanceof du?e=e.criteria:e instanceof lu&&(e=[e]),this.criteria.push(...e)}_call(e,r){const n=new Array(e.length).fill(!1);for(const a of this.criteria){const i=a(e,r);for(let s=0;s<n.length;++s)n[s]||=i[s]}return n}[Symbol.iterator](){return this.criteria.values()}}class Dv extends lu{constructor(e,r=null){super(),this.max_length=e,this.max_position_embeddings=r}_call(e){return e.map(r=>r.length>=this.max_length)}}class Nv extends lu{constructor(e){super(),Array.isArray(e)||(e=[e]),this.eos_token_id=e}_call(e,r){return e.map(n=>{const a=n.at(-1);return this.eos_token_id.some(i=>a==i)})}}class Ji extends $t{constructor(e){super(),this.generation_config=e}_call(e,r=-1){return this.sample(e,r)}sample(e,r){throw Error("sample should be implemented in subclasses.")}getLogits(e,r){let n=e.dims.at(-1),a=e.data;if(r===-1)a=a.slice(-n);else{let i=r*n;a=a.slice(i,i+n)}return a}randomSelect(e){let r=e.reduce((a,i)=>a+i,0),n=Math.random()*r;for(let a=0;a<e.length;++a)if(n-=e[a],n<=0)return a;return 0}static getSampler(e){if(e.do_sample)return new Lv(e);if(e.num_beams>1)return new Uv(e);if(e.num_return_sequences>1)throw Error(`num_return_sequences has to be 1 when doing greedy search, but is ${e.num_return_sequences}.`);return new Fv(e)}}class Fv extends Ji{sample(e,r=-1){let n=this.getLogits(e,r);return[[vr(n)[1],0]]}}class Lv extends Ji{sample(e,r=-1){let n=e.dims.at(-1);this.generation_config.top_k>0&&(n=Math.min(this.generation_config.top_k,n));const a=this.getLogits(e,r),i=Ap(a,n),s=Tn(i.map(o=>o[1]));return Array.from({length:this.generation_config.num_beams},()=>{const o=this.randomSelect(s);return[i[o][0],Math.log(s[o])]})}}class Uv extends Ji{sample(e,r=-1){let n=e.dims.at(-1);this.generation_config.top_k>0&&(n=Math.min(this.generation_config.top_k,n));const a=this.getLogits(e,r),i=Ap(a,n),s=Tn(i.map(o=>o[1]));return Array.from({length:this.generation_config.num_beams},(o,u)=>[i[u][0],Math.log(s[u])])}}class Wv extends kg{return_timestamps=null;return_token_timestamps=null;num_frames=null;alignment_heads=null;task=null;language=null;no_timestamps_token_id=null;prompt_ids=null;is_multilingual=null;lang_to_id=null;task_to_id=null;max_initial_timestamp_index=1}const ve={EncoderOnly:0,EncoderDecoder:1,Seq2Seq:2,Vision2Seq:3,DecoderOnly:4,MaskGeneration:5,ImageTextToText:6,Musicgen:7},qi=new Map,Eg=new Map,pa=new Map;async function Vv(t,e,r){let n=r.device;n&&typeof n!="string"&&(n.hasOwnProperty(e)?n=n[e]:(console.warn(`Device not specified for ${e}. Using the default device.`),n=null));const a=rw(n);let i=r.dtype;if(typeof i!="string"&&(i&&i.hasOwnProperty(e)?i=i[e]:(i=xv[a[0]],console.warn(`Dtype not specified for ${e}. Using the default dtype: ${i}.`))),fp.hasOwnProperty(i)){if(i===It.fp16&&!await $v())throw new Error("The device does not support fp16.")}else throw new Error(`Invalid dtype: ${i}. Should be one of: ${Object.keys(It).join(", ")}`);const s=fp[i],o=`${r.subfolder??""}/${e}${s}.onnx`,u={...r.session_options};u.executionProviders??=a;const l=ki(t,o,!0,r);let p=[];if(r.use_external_data_format){if(Jr.IS_NODE_ENV)throw new Error("External data format is not yet supported in Node.js");const m=`${e}${s}.onnx_data`,c=`${r.subfolder??""}/${m}`;p.push(new Promise(async(y,w)=>{const v=await ki(t,c,!0,r);y({path:m,data:v})}))}else u.externalData!==void 0&&(p=u.externalData.map(async m=>{if(typeof m.data=="string"){const c=await ki(t,m.data,!0,r);return{...m,data:c}}return m}));if(p.length>0&&(u.externalData=await Promise.all(p)),n==="webgpu"){const m=xg(r.config,{prefix:"present"}),c={};for(const y in m)c[y]="gpu-buffer";u.preferredOutputLocation=c}return{buffer:await l,session_options:u}}async function Vr(t,e,r){const n=Object.keys(e),a=await Promise.all(n.map(async s=>Vv(t,e[s],r))),i={};for(let s=0;s<n.length;++s){const{buffer:o,session_options:u}=a[s],l=await ig(o,u);i[n[s]]=l}return i}function Gv(t,e){const r=Object.create(null),n=[];for(const s of t.inputNames){const o=e[s];if(!(o instanceof pe)){n.push(s);continue}r[s]=nw()?o.clone():o}if(n.length>0)throw new Error(`An error occurred during model execution: "Missing the following inputs: ${n.join(", ")}.`);const a=Object.keys(e).length,i=t.inputNames.length;if(a>i){let s=Object.keys(e).filter(o=>!t.inputNames.includes(o));console.warn(`WARNING: Too many inputs were provided (${a} > ${i}). The following inputs will be ignored: "${s.join(", ")}".`)}return r}async function Or(t,e){const r=Gv(t,e);try{const n=Object.fromEntries(Object.entries(r).map(([i,s])=>[i,s.ort_tensor]));let a=await t.run(n);return a=Cg(a),a}catch(n){throw console.error(`An error occurred during model execution: "${n}".`),console.error("Inputs given to model:",r),n}}function Cg(t){for(let e in t)sg(t[e])?t[e]=new pe(t[e]):typeof t[e]=="object"&&Cg(t[e]);return t}function Tg(t){if(t instanceof pe)return t;if(t.length===0)throw Error("items must be non-empty");if(Array.isArray(t[0])){if(t.some(e=>e.length!==t[0].length))throw Error("Unable to create tensor, you should probably activate truncation and/or padding with 'padding=True' and/or 'truncation=True' to have batched tensors with the same length.");return new pe("int64",BigInt64Array.from(t.flat().map(e=>BigInt(e))),[t.length,t[0].length])}else return new pe("int64",BigInt64Array.from(t.map(e=>BigInt(e))),[1,t.length])}function Ig(t){return new pe("bool",[t],[1])}async function mp(t,e){let{encoder_outputs:r,past_key_values:n}=e;if(!r){const u=Xr(e,t.sessions.model.inputNames);r=(await wa(t,u)).last_hidden_state}const{input_ids:a,decoder_input_ids:i,...s}=e;return s.input_ids=i,s.encoder_hidden_states=r,t.sessions.decoder_model_merged.inputNames.includes("encoder_attention_mask")&&(s.encoder_attention_mask=e.attention_mask),await cu(t,s,!0)}async function wa(t,e){const r=t.sessions.model,n=Object.create(null);for(const a of r.inputNames)n[a]=e[a];return r.inputNames.includes("token_type_ids")&&!n.token_type_ids&&(n.token_type_ids=new pe("int64",new BigInt64Array(n.input_ids.data.length),n.input_ids.dims)),await Or(r,n)}async function cu(t,e,r=!1){const n=t.sessions[r?"decoder_model_merged":"model"],{past_key_values:a,...i}=e;n.inputNames.includes("use_cache_branch")&&(i.use_cache_branch=Ig(!!a)),n.inputNames.includes("position_ids")&&i.attention_mask&&!i.position_ids&&(i.position_ids=jv(i,a)),t.addPastKeyValues(i,a);const s=Xr(i,n.inputNames);return await Or(n,s)}async function Hv(t,{input_ids:e=null,attention_mask:r=null,pixel_values:n=null,position_ids:a=null,inputs_embeds:i=null,past_key_values:s=null,generation_config:o=null,logits_processor:u=null,...l}){if(!i){if(i=await t.encode_text({input_ids:e}),n&&e.dims[1]!==1){const f=await t.encode_image({pixel_values:n});({inputs_embeds:i,attention_mask:r}=t._merge_input_ids_with_image_features({image_features:f,inputs_embeds:i,input_ids:e,attention_mask:r}))}else if(s&&n&&e.dims[1]===1){const f=e.dims[1],m=Object.values(s)[0].dims.at(-2);r=dr([xa([e.dims[0],m]),r.slice(null,[r.dims[1]-f,r.dims[1]])],1)}}return await cu(t,{inputs_embeds:i,past_key_values:s,attention_mask:r,position_ids:a,generation_config:o,logits_processor:u},!0)}function jv(t,e=null){const{input_ids:r,inputs_embeds:n,attention_mask:a}=t,[i,s]=a.dims,o=new BigInt64Array(a.data.length);for(let l=0;l<i;++l){const p=l*s;let f=BigInt(0);for(let m=0;m<s;++m){const c=p+m;a.data[c]===0n?o[c]=BigInt(1):(o[c]=f,f+=a.data[c])}}let u=new pe("int64",o,a.dims);if(e){const l=-(r??n).dims.at(1);u=u.slice(null,[l,null])}return u}function gp(t,e,r,n){if(r.past_key_values){const a=Object.values(r.past_key_values)[0].dims.at(-2),{input_ids:i,attention_mask:s}=r;if(!(s&&s.dims[1]>i.dims[1])){if(a<i.dims[1])r.input_ids=i.slice(null,[a,null]);else if(t.config.image_token_index!=null&&i.data.some(o=>o==t.config.image_token_index)){const o=t.config.num_image_tokens;if(!o)throw new Error("`num_image_tokens` is missing in the model configuration.");const u=i.dims[1]-(a-o);r.input_ids=i.slice(null,[-u,null]),r.attention_mask=xa([1,a+u])}}}return r}function qv(t,e,r,n){const{...a}=r;return r.past_key_values&&(e=e.map(s=>[s.at(-1)])),a.decoder_input_ids=Tg(e),a}class ee extends $t{main_input_name="input_ids";forward_params=["input_ids","attention_mask"];constructor(e,r){super(),this.config=e,this.sessions=r;const n=pa.get(this.constructor),a=qi.get(n);this.can_generate=!1,this._forward=null,this._prepare_inputs_for_generation=null,a===ve.DecoderOnly?(this.can_generate=!0,this._forward=cu,this._prepare_inputs_for_generation=gp):a===ve.Seq2Seq||a===ve.Vision2Seq||a===ve.Musicgen?(this.can_generate=!0,this._forward=mp,this._prepare_inputs_for_generation=qv):a===ve.EncoderDecoder?this._forward=mp:a===ve.ImageTextToText?(this.can_generate=!0,this._forward=Hv,this._prepare_inputs_for_generation=gp):this._forward=wa,this.can_generate&&this.forward_params.push("past_key_values"),this.custom_config=this.config["transformers.js_config"]??{}}async dispose(){const e=[];for(const r of Object.values(this.sessions))r?.handler?.dispose&&e.push(r.handler.dispose());return await Promise.all(e)}static async from_pretrained(e,{progress_callback:r=null,config:n=null,cache_dir:a=null,local_files_only:i=!1,revision:s="main",model_file_name:o=null,subfolder:u="onnx",device:l=null,dtype:p=null,use_external_data_format:f=null,session_options:m={}}={}){let c={progress_callback:r,config:n,cache_dir:a,local_files_only:i,revision:s,model_file_name:o,subfolder:u,device:l,dtype:p,use_external_data_format:f,session_options:m};const y=pa.get(this),w=qi.get(y);c.config=await Sg.from_pretrained(e,c);let v;return w===ve.DecoderOnly?v=await Promise.all([Vr(e,{model:c.model_file_name??"model"},c),Ar(e,"generation_config.json",!1,c)]):w===ve.Seq2Seq||w===ve.Vision2Seq?v=await Promise.all([Vr(e,{model:"encoder_model",decoder_model_merged:"decoder_model_merged"},c),Ar(e,"generation_config.json",!1,c)]):w===ve.MaskGeneration?v=await Promise.all([Vr(e,{model:"vision_encoder",prompt_encoder_mask_decoder:"prompt_encoder_mask_decoder"},c)]):w===ve.EncoderDecoder?v=await Promise.all([Vr(e,{model:"encoder_model",decoder_model_merged:"decoder_model_merged"},c)]):w===ve.ImageTextToText?v=await Promise.all([Vr(e,{embed_tokens:"embed_tokens",vision_encoder:"vision_encoder",decoder_model_merged:"decoder_model_merged"},c),Ar(e,"generation_config.json",!1,c)]):w===ve.Musicgen?v=await Promise.all([Vr(e,{model:"text_encoder",decoder_model_merged:"decoder_model_merged",encodec_decode:"encodec_decode"},c),Ar(e,"generation_config.json",!1,c)]):(w!==ve.EncoderOnly&&console.warn(`Model type for '${y??n?.model_type}' not found, assuming encoder-only architecture. Please report this at https://github.com/xenova/transformers.js/issues/new/choose.`),v=await Promise.all([Vr(e,{model:c.model_file_name??"model"},c)])),new this(c.config,...v)}async _call(e){return await this.forward(e)}async forward(e){return await this._forward(this,e)}_get_logits_warper(e){const r=new Ao;return e.temperature!==null&&e.temperature!==1&&r.push(new Pv(e.temperature)),e.top_k!==null&&e.top_k!==0&&r.push(new Bv(e.top_k)),e.top_p!==null&&e.top_p<1&&r.push(new Rv(e.top_p)),r}_get_logits_processor(e,r,n=null){const a=new Ao;if(e.repetition_penalty!==null&&e.repetition_penalty!==1&&a.push(new Iv(e.repetition_penalty)),e.no_repeat_ngram_size!==null&&e.no_repeat_ngram_size>0&&a.push(new Tv(e.no_repeat_ngram_size)),e.bad_words_ids!==null&&a.push(new Ov(e.bad_words_ids,e.eos_token_id)),e.min_length!==null&&e.eos_token_id!==null&&e.min_length>0&&a.push(new Av(e.min_length,e.eos_token_id)),e.min_new_tokens!==null&&e.eos_token_id!==null&&e.min_new_tokens>0&&a.push(new Mv(r,e.min_new_tokens,e.eos_token_id)),e.forced_bos_token_id!==null&&a.push(new Sv(e.forced_bos_token_id)),e.forced_eos_token_id!==null&&a.push(new kv(e.max_length,e.forced_eos_token_id)),e.begin_suppress_tokens!==null){const i=r>1||e.forced_bos_token_id===null?r:r+1;a.push(new Ev(e.begin_suppress_tokens,i))}return e.guidance_scale!==null&&e.guidance_scale>1&&a.push(new zv(e.guidance_scale)),n!==null&&a.extend(n),a}_prepare_generation_config(e,r,n=kg){const a={...this.config};for(const s of["decoder","generator","text_config"])s in a&&Object.assign(a,a[s]);const i=new n(a);return"generation_config"in this&&Object.assign(i,this.generation_config),e&&Object.assign(i,e),r&&Object.assign(i,Xr(r,Object.getOwnPropertyNames(i))),i}_get_stopping_criteria(e,r=null){const n=new du;return e.max_length!==null&&n.push(new Dv(e.max_length,this.config.max_position_embeddings??null)),e.eos_token_id!==null&&n.push(new Nv(e.eos_token_id)),r&&n.extend(r),n}_validate_model_class(){if(!this.can_generate){const e=[R0,B0,P0,z0],r=pa.get(this.constructor),n=new Set,a=this.config.model_type;for(const s of e){const o=s.get(a);o&&n.add(o[0])}let i=`The current model class (${r}) is not compatible with \`.generate()\`, as it doesn't have a language model head.`;throw n.size>0&&(i+=` Please use the following class instead: ${[...n].join(", ")}`),Error(i)}}prepare_inputs_for_generation(...e){return this._prepare_inputs_for_generation(this,...e)}_update_model_kwargs_for_generation({generated_input_ids:e,outputs:r,model_inputs:n,is_encoder_decoder:a}){return n.past_key_values=this.getPastKeyValues(r,n.past_key_values),n.input_ids=new pe("int64",e.flat(),[e.length,1]),a||(n.attention_mask=dr([n.attention_mask,xa([n.attention_mask.dims[0],1])],1)),n.position_ids=null,n}_prepare_model_inputs({inputs:e,bos_token_id:r,model_kwargs:n}){const a=Xr(n,this.forward_params),i=this.main_input_name;if(i in a){if(e)throw new Error("`inputs`: {inputs}` were passed alongside {input_name} which is not allowed. Make sure to either pass {inputs} or {input_name}=...")}else a[i]=e;return{inputs_tensor:a[i],model_inputs:a,model_input_name:i}}async _prepare_encoder_decoder_kwargs_for_generation({inputs_tensor:e,model_inputs:r,model_input_name:n,generation_config:a}){const i=Xr(r,this.sessions.model.inputNames);let{last_hidden_state:s}=await wa(this,i);return a.guidance_scale!==null&&a.guidance_scale>1&&(s=dr([s,dw(s,0)],0),"attention_mask"in r&&(r.attention_mask=dr([r.attention_mask,hw(r.attention_mask)],0))),r.encoder_outputs=s,r}_prepare_decoder_input_ids_for_generation({batch_size:e,model_input_name:r,model_kwargs:n,decoder_start_token_id:a,bos_token_id:i,generation_config:s}){let{decoder_input_ids:o,...u}=n;if(!o)if(a??=i,this.config.model_type==="musicgen")o=Array.from({length:e*this.config.decoder.num_codebooks},()=>[a]);else if(Array.isArray(a)){if(a.length!==e)throw new Error(`\`decoder_start_token_id\` expcted to have length ${e} but got ${a.length}`);o=a}else o=Array.from({length:e},()=>[a]);return o=Tg(o),n.decoder_attention_mask=cw(o),{input_ids:o,model_inputs:u}}async generate({inputs:e=null,generation_config:r=null,logits_processor:n=null,stopping_criteria:a=null,streamer:i=null,...s}){this._validate_model_class(),r=this._prepare_generation_config(r,s);let{inputs_tensor:o,model_inputs:u,model_input_name:l}=this._prepare_model_inputs({inputs:e,model_kwargs:s});const p=this.config.is_encoder_decoder;p&&("encoder_outputs"in u||(u=await this._prepare_encoder_decoder_kwargs_for_generation({inputs_tensor:o,model_inputs:u,model_input_name:l,generation_config:r})));let f;p?{input_ids:f,model_inputs:u}=this._prepare_decoder_input_ids_for_generation({batch_size:u[l].dims.at(0),model_input_name:l,model_kwargs:u,decoder_start_token_id:r.decoder_start_token_id,bos_token_id:r.bos_token_id,generation_config:r}):f=u[l];let m=f.dims.at(-1);r.max_new_tokens!==null&&(r.max_length=m+r.max_new_tokens);const c=this._get_logits_processor(r,m,n),y=this._get_stopping_criteria(r,a),w=u[l].dims.at(0),v=Ji.getSampler(r),k=new Array(w).fill(0),$=f.tolist();i&&i.put($);let C=null;for(;;){u=this.prepare_inputs_for_generation($,u,r);const A=await this.forward(u),B=A.logits.slice(null,-1,null),R=c($,B),D=[];for(let j=0;j<R.dims.at(0);++j){const ie=R[j],te=v(ie);for(const[oe,re]of te){const M=BigInt(oe);k[j]+=re,$[j].push(M),D.push([M])}}if(i&&i.put(D),y($).every(j=>j)){r.return_dict_in_generate&&(C=this.getPastKeyValues(A,u.past_key_values,!1));break}u=this._update_model_kwargs_for_generation({generated_input_ids:D,outputs:A,model_inputs:u,is_encoder_decoder:p})}i&&i.end();const T=new pe("int64",$.flat(),[$.length,$[0].length]);return r.return_dict_in_generate?{sequences:T,past_key_values:C}:T}addAttentionsToBeam(e,r){if(this.config.is_encoder_decoder){if(!r.cross_attentions||r.cross_attentions.length===0)throw Error("`output_attentions` is true, but the model did not produce cross-attentions. This is most likely because the model was not exported with `output_attentions=True`.");e.cross_attentions||(e.cross_attentions=[]),e.cross_attentions.push(r.cross_attentions)}if(!r.decoder_attentions||r.decoder_attentions.length===0)throw Error("`output_attentions` is true, but the model did not produce decoder-attentions. This is most likely because the model was not exported with `output_attentions=True`.");e.decoder_attentions||(e.decoder_attentions=[]),e.decoder_attentions.push(r.decoder_attentions)}groupBeams(e){const r=Object.create(null);for(const n of e)r[n.id]===void 0?r[n.id]=[n]:r[n.id].push(n);return Object.values(r)}getPastKeyValues(e,r,n=!0){const a=Object.create(null);for(const i in e)if(i.startsWith("present")){let s=i.replace("present","past_key_values");if(r&&i.includes("encoder"))a[s]=r[s];else{if(n&&r){const o=r[s];o.location==="gpu-buffer"&&o.dispose()}a[s]=e[i]}}return a}getAttentions(e){const r=Object.create(null);for(const n of["cross_attentions","decoder_attentions"]){const a=[];for(const i in e)if(i.startsWith(n)){const s=i.split(".").pop();a[s]=e[i]}r[n]=a}return r}addPastKeyValues(e,r){if(r)Object.assign(e,r);else{const n=this.custom_config.kv_cache_dtype??"float32",a=n==="float16"?new Uint16Array:[],i=xg(this.config);for(const s in i)e[s]=new pe(n,a,i[s])}}}class Wt{}class Ea extends ee{}class Kv extends Ea{}class Yv extends Ea{async _call(e){return new gt(await super._call(e))}}class Xv extends Ea{async _call(e){return new Me(await super._call(e))}}class Qv extends Ea{async _call(e){return new mt(await super._call(e))}}class Jv extends Ea{async _call(e){return new xt(await super._call(e))}}class Zv extends ee{}class e2 extends Zv{}class Ca extends ee{}class t2 extends Ca{}class r2 extends Ca{async _call(e){return new gt(await super._call(e))}}class n2 extends Ca{async _call(e){return new Me(await super._call(e))}}class a2 extends Ca{async _call(e){return new mt(await super._call(e))}}class i2 extends Ca{async _call(e){return new xt(await super._call(e))}}class Ta extends ee{}class s2 extends Ta{}class o2 extends Ta{async _call(e){return new gt(await super._call(e))}}class u2 extends Ta{async _call(e){return new Me(await super._call(e))}}class l2 extends Ta{async _call(e){return new mt(await super._call(e))}}class d2 extends Ta{async _call(e){return new xt(await super._call(e))}}class Ia extends ee{}class c2 extends Ia{}class p2 extends Ia{async _call(e){return new gt(await super._call(e))}}class h2 extends Ia{async _call(e){return new Me(await super._call(e))}}class f2 extends Ia{async _call(e){return new mt(await super._call(e))}}class m2 extends Ia{async _call(e){return new xt(await super._call(e))}}class Aa extends ee{}class g2 extends Aa{}class _2 extends Aa{async _call(e){return new gt(await super._call(e))}}class y2 extends Aa{async _call(e){return new Me(await super._call(e))}}class w2 extends Aa{async _call(e){return new mt(await super._call(e))}}class b2 extends Aa{async _call(e){return new xt(await super._call(e))}}class Ma extends ee{}class v2 extends Ma{}class $2 extends Ma{async _call(e){return new gt(await super._call(e))}}class x2 extends Ma{async _call(e){return new Me(await super._call(e))}}class S2 extends Ma{async _call(e){return new mt(await super._call(e))}}class k2 extends Ma{async _call(e){return new xt(await super._call(e))}}class Oa extends ee{}class E2 extends Oa{}class C2 extends Oa{async _call(e){return new gt(await super._call(e))}}class T2 extends Oa{async _call(e){return new Me(await super._call(e))}}class I2 extends Oa{async _call(e){return new mt(await super._call(e))}}class A2 extends Oa{async _call(e){return new xt(await super._call(e))}}class za extends ee{}class M2 extends za{}class O2 extends za{async _call(e){return new Me(await super._call(e))}}class z2 extends za{async _call(e){return new mt(await super._call(e))}}class P2 extends za{async _call(e){return new xt(await super._call(e))}}class R2 extends za{async _call(e){return new gt(await super._call(e))}}class Zi extends ee{}class B2 extends Zi{}class D2 extends Zi{async _call(e){return new gt(await super._call(e))}}class N2 extends Zi{async _call(e){return new Me(await super._call(e))}}class F2 extends Zi{async _call(e){return new mt(await super._call(e))}}class es extends ee{}class L2 extends es{}class U2 extends es{async _call(e){return new gt(await super._call(e))}}class W2 extends es{async _call(e){return new Me(await super._call(e))}}class V2 extends es{async _call(e){return new xt(await super._call(e))}}class Pa extends ee{}class G2 extends Pa{}class H2 extends Pa{async _call(e){return new gt(await super._call(e))}}class j2 extends Pa{async _call(e){return new Me(await super._call(e))}}class q2 extends Pa{async _call(e){return new mt(await super._call(e))}}class K2 extends Pa{async _call(e){return new xt(await super._call(e))}}class ts extends ee{}class Y2 extends ts{}class X2 extends ts{async _call(e){return new gt(await super._call(e))}}class Q2 extends ts{async _call(e){return new Me(await super._call(e))}}class J2 extends ts{async _call(e){return new xt(await super._call(e))}}class rs extends ee{}class Z2 extends rs{}class e1 extends rs{async _call(e){return new Me(await super._call(e))}}class t1 extends rs{async _call(e){return new xt(await super._call(e))}}class r1 extends rs{async _call(e){return new gt(await super._call(e))}}class Ag extends ee{forward_params=["input_ids","attention_mask","encoder_outputs","decoder_input_ids","decoder_attention_mask","past_key_values"];constructor(e,r,n){super(e,r),this.generation_config=n}}class n1 extends Ag{}class a1 extends Ag{}class Mg extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class i1 extends Mg{}class s1 extends Mg{}class Og extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class o1 extends Og{}class u1 extends Og{}class pu extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class l1 extends pu{}class d1 extends pu{}class c1 extends pu{async _call(e){return new Me(await super._call(e))}}class ns extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class p1 extends ns{}class h1 extends ns{}class f1 extends ns{async _call(e){return new Me(await super._call(e))}}class m1 extends ns{}class zg extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class g1 extends zg{}class _1 extends zg{}class Pg extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class y1 extends Pg{}class w1 extends Pg{}class Ra extends ee{}class b1 extends Ra{}class v1 extends Ra{async _call(e){return new gt(await super._call(e))}}class $1 extends Ra{async _call(e){return new Me(await super._call(e))}}class x1 extends Ra{async _call(e){return new mt(await super._call(e))}}class S1 extends Ra{async _call(e){return new xt(await super._call(e))}}class Ba extends ee{}class k1 extends Ba{}class E1 extends Ba{async _call(e){return new gt(await super._call(e))}}class C1 extends Ba{async _call(e){return new Me(await super._call(e))}}class T1 extends Ba{async _call(e){return new mt(await super._call(e))}}class I1 extends Ba{async _call(e){return new xt(await super._call(e))}}class Da extends ee{}class A1 extends Da{}class M1 extends Da{async _call(e){return new gt(await super._call(e))}}class O1 extends Da{async _call(e){return new Me(await super._call(e))}}class z1 extends Da{async _call(e){return new mt(await super._call(e))}}class P1 extends Da{async _call(e){return new xt(await super._call(e))}}class Rg extends ee{}class R1 extends Rg{}class B1 extends Rg{}class Bg extends ee{requires_attention_mask=!1;main_input_name="input_features";forward_params=["input_features","attention_mask","decoder_input_ids","decoder_attention_mask","past_key_values"];constructor(e,r,n){super(e,r),this.generation_config=n}}class D1 extends Bg{}class N1 extends Bg{_prepare_generation_config(e,r){return super._prepare_generation_config(e,r,Wv)}_retrieve_init_tokens(e){const r=[e.decoder_start_token_id];let n=e.language;const a=e.task;if(e.is_multilingual){n||(console.warn("No language specified - defaulting to English (en)."),n="en");const s=`<|${dg(n)}|>`;r.push(e.lang_to_id[s]),r.push(e.task_to_id[a??"transcribe"])}else if(n||a)throw new Error("Cannot specify `task` or `language` for an English-only model. If the model is intended to be multilingual, pass `is_multilingual=true` to generate, or update the generation config.");return!e.return_timestamps&&e.no_timestamps_token_id&&r.at(-1)!==e.no_timestamps_token_id?r.push(e.no_timestamps_token_id):e.return_timestamps&&r.at(-1)===e.no_timestamps_token_id&&(console.warn("<|notimestamps|> prompt token is removed from generation_config since `return_timestamps` is set to `true`."),r.pop()),r.filter(i=>i!=null)}async generate({inputs:e=null,generation_config:r=null,logits_processor:n=null,stopping_criteria:a=null,...i}){r=this._prepare_generation_config(r,i);const s=this._retrieve_init_tokens(r);return r.return_timestamps&&(n??=new Ao,n.push(new Cv(r,s))),await super.generate({inputs:e,generation_config:r,logits_processor:n,decoder_input_ids:s,...i})}_extract_token_timestamps(e,r,n=null,a=.02){if(!e.cross_attentions)throw new Error("Model outputs must contain cross attentions to extract timestamps. This is most likely because the model was not exported with `output_attentions=True`.");let i=this.config.median_filter_width;i===void 0&&(console.warn("Model config has no `median_filter_width`, using default value of 7."),i=7);const s=e.cross_attentions.map(l=>{let p=Array.from({length:this.config.decoder_layers},(v,k)=>dr(l.map($=>$[k]),2)),f=_a(r.map(([v,k])=>n?p[v].slice(null,k,null,[0,n]):p[v].slice(null,k)));f=f.transpose(1,0,2,3);let[m,c]=sw(f,-2,0,!0),y=f.clone();for(let v=0;v<y.dims[0];++v){let k=y[v];for(let $=0;$<k.dims[0];++$){let C=k[$];const T=m[v][$][0],A=c[v][$][0];for(let B=0;B<C.dims[0];++B){let R=C[B];for(let D=0;D<R.data.length;++D)R.data[D]=(R.data[D]-A.data[D])/T.data[D];R.data.set($_(R.data,i))}}}return nu(y,1)}),o=[e.sequences.length,e.sequences[0].length],u=new pe("float32",new Float32Array(o[0]*o[1]),o);for(let l=0;l<o[0];++l){const p=s[l].neg().squeeze_(0);let[f,m]=ow(p),c=Array.from({length:f.length-1},(v,k)=>f[k+1]-f[k]),y=ct([1],c).map(v=>!!v),w=[];for(let v=0;v<y.length;++v)y[v]&&w.push(m[v]*a);u[l].data.set(w,1)}return u}}class F1 extends ee{main_input_name="pixel_values";forward_params=["pixel_values","input_ids","encoder_hidden_states","past_key_values"];constructor(e,r,n){super(e,r),this.generation_config=n}}class L1 extends ee{forward_params=["input_ids","pixel_values","attention_mask","position_ids","past_key_values"];constructor(e,r,n){super(e,r),this.generation_config=n}}class Dg extends L1{async encode_image({pixel_values:e}){const r=(await Or(this.sessions.vision_encoder,{pixel_values:e})).image_features;return this.config.num_image_tokens||(console.warn(`The number of image tokens was not set in the model configuration. Setting it to the number of features detected by the vision encoder (${r.dims[1]}).`),this.config.num_image_tokens=r.dims[1]),r}async encode_text({input_ids:e}){return(await Or(this.sessions.embed_tokens,{input_ids:e})).inputs_embeds}_merge_input_ids_with_image_features({inputs_embeds:e,image_features:r,input_ids:n,attention_mask:a}){const i=this.config.image_token_index,o=n.tolist().map(m=>m.findIndex(c=>c==i)),u=o.every(m=>m===-1),l=o.every(m=>m!==-1);if(!u&&!l)throw new Error("Every input should contain either 0 or 1 image token.");if(u)return{inputs_embeds:e,attention_mask:a};const p=[],f=[];for(let m=0;m<o.length;++m){const c=o[m],y=e[m],w=r[m],v=a[m];p.push(dr([y.slice([0,c]),w,y.slice([c+1,y.dims[0]])],0)),f.push(dr([v.slice([0,c]),xa([w.dims[0]]),v.slice([c+1,v.dims[0]])],0))}return{inputs_embeds:_a(p,0),attention_mask:_a(f,0)}}}class U1 extends Dg{}class as extends ee{}class W1 extends as{}class Ng extends as{static async from_pretrained(e,r={}){return r.model_file_name??="text_model",super.from_pretrained(e,r)}}class Fg extends as{static async from_pretrained(e,r={}){return r.model_file_name??="vision_model",super.from_pretrained(e,r)}}class Lg extends ee{}class V1 extends Lg{}class G1 extends Lg{static async from_pretrained(e,r={}){return r.model_file_name??="text_model",super.from_pretrained(e,r)}}class H1 extends as{static async from_pretrained(e,r={}){return r.model_file_name??="vision_model",super.from_pretrained(e,r)}}class j1 extends ee{}class q1 extends j1{}class Ug extends ee{}class K1 extends Ug{}class Y1 extends Ug{}class Wg extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class X1 extends Wg{}class Q1 extends Wg{}class Vg extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class J1 extends Vg{}class Z1 extends Vg{}class Gg extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class e$ extends Gg{}class t$ extends Gg{}class Hg extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class r$ extends Hg{}class n$ extends Hg{}class jg extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class a$ extends jg{}class i$ extends jg{}class qg extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class s$ extends qg{}class o$ extends qg{}class Kg extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class u$ extends Kg{}class l$ extends Kg{}class Yg extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class d$ extends Yg{}class c$ extends Yg{}class Xg extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class p$ extends Xg{}class h$ extends Xg{}class Qg extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class f$ extends Qg{}class m$ extends Qg{}class Jg extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class g$ extends Jg{}class _$ extends Jg{}class Zg extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class y$ extends Zg{}class w$ extends Zg{}class e0 extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class b$ extends e0{}class v$ extends e0{}class t0 extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class $$ extends t0{}class x$ extends t0{}class r0 extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class S$ extends r0{}class k$ extends r0{}class n0 extends ee{}class E$ extends n0{}class C$ extends n0{async _call(e){return new Me(await super._call(e))}}class a0 extends ee{}class T$ extends a0{}class I$ extends a0{async _call(e){return new Me(await super._call(e))}}class A$ extends ee{}class M$ extends A${async _call(e){return new VS(await super._call(e))}}class i0 extends ee{}class O$ extends i0{}class z$ extends i0{async _call(e){return new Me(await super._call(e))}}class s0 extends ee{}class P$ extends s0{}class R$ extends s0{async _call(e){return new Me(await super._call(e))}}class o0 extends ee{}class B$ extends o0{}class D$ extends o0{}class u0 extends ee{}class N$ extends u0{}class F$ extends u0{}class l0 extends ee{}class L$ extends l0{}class U$ extends l0{async _call(e){return new Me(await super._call(e))}}class hu extends ee{}class W$ extends hu{}class V$ extends hu{async _call(e){return new d0(await super._call(e))}}class G$ extends hu{async _call(e){return new H$(await super._call(e))}}class d0 extends Wt{constructor({logits:e,pred_boxes:r}){super(),this.logits=e,this.pred_boxes=r}}class H$ extends Wt{constructor({logits:e,pred_boxes:r,pred_masks:n}){super(),this.logits=e,this.pred_boxes=r,this.pred_masks=n}}class c0 extends ee{}class j$ extends c0{}class q$ extends c0{async _call(e){return new K$(await super._call(e))}}class K$ extends d0{}class p0 extends ee{}class Y$ extends p0{}class X$ extends p0{async _call(e){return new Me(await super._call(e))}}class h0 extends ee{}class Q$ extends h0{}class J$ extends h0{async _call(e){return new Me(await super._call(e))}}class f0 extends ee{}class Z$ extends f0{}class ex extends f0{async _call(e){return new Me(await super._call(e))}}class m0 extends ee{}class tx extends m0{}class rx extends m0{}class g0 extends ee{}class nx extends g0{}class ax extends g0{}class ix extends ee{}class sx extends ix{}class _0 extends ee{}class ox extends _0{}class ux extends _0{}class lx extends ee{}class dx extends lx{}class y0 extends ee{}class cx extends y0{}class px extends y0{async _call(e){return new Me(await super._call(e))}}class w0 extends ee{}class hx extends w0{}class fx extends w0{async _call(e){return new Me(await super._call(e))}}class b0 extends ee{}class mx extends b0{}class gx extends b0{async _call(e){return new Me(await super._call(e))}}class v0 extends ee{}class _x extends v0{}class yx extends v0{async _call(e){return new wx(await super._call(e))}}class wx extends Wt{constructor({logits:e,pred_boxes:r}){super(),this.logits=e,this.pred_boxes=r}}class bx extends ee{}class vx extends bx{async get_image_embeddings({pixel_values:e}){return await wa(this,{pixel_values:e})}async forward(e){if((!e.image_embeddings||!e.image_positional_embeddings)&&(e={...e,...await this.get_image_embeddings(e)}),!e.input_labels&&e.input_points){const n=e.input_points.dims.slice(0,-1),a=n.reduce((i,s)=>i*s,1);e.input_labels=new pe("int64",new BigInt64Array(a).fill(1n),n)}const r={image_embeddings:e.image_embeddings,image_positional_embeddings:e.image_positional_embeddings};return e.input_points&&(r.input_points=e.input_points),e.input_labels&&(r.input_labels=e.input_labels),e.input_boxes&&(r.input_boxes=e.input_boxes),await Or(this.sessions.prompt_encoder_mask_decoder,r)}async _call(e){return new $x(await super._call(e))}}class $x extends Wt{constructor({iou_scores:e,pred_masks:r}){super(),this.iou_scores=e,this.pred_masks=r}}class $0 extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class xx extends $0{}class Sx extends $0{}class x0 extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class kx extends x0{}class Ex extends x0{}class an extends ee{}class Cx extends an{}class Tx extends an{async _call(e){return new On(await super._call(e))}}class Ix extends an{async _call(e){return new Me(await super._call(e))}}class Ax extends an{async _call(e){return new mt(await super._call(e))}}class fu extends ee{}class Mx extends fu{}class Ox extends fu{async _call(e){return new On(await super._call(e))}}class zx extends fu{async _call(e){return new Me(await super._call(e))}}class is extends ee{}class Px extends is{}class Rx extends is{async _call(e){return new On(await super._call(e))}}class Bx extends is{async _call(e){return new Me(await super._call(e))}}class Dx extends is{async _call(e){return new mt(await super._call(e))}}class mu extends ee{}class Nx extends mu{}class Fx extends mu{async _call(e){return new On(await super._call(e))}}class Lx extends mu{async _call(e){return new Me(await super._call(e))}}class Ux extends an{}class Wx extends an{async _call(e){return new On(await super._call(e))}}class Vx extends an{async _call(e){return new Me(await super._call(e))}}class Na extends ee{}class Gx extends Na{}class Hx extends Na{async _call(e){return new On(await super._call(e))}}class jx extends Na{async _call(e){return new Me(await super._call(e))}}class qx extends Na{async _call(e){return new WS(await super._call(e))}}class Kx extends Na{async _call(e){return new mt(await super._call(e))}}class S0 extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class Yx extends S0{}class Xx extends S0{async generate_speech(e,r,{threshold:n=.5,minlenratio:a=0,maxlenratio:i=20,vocoder:s=null}={}){const o={input_ids:e},{encoder_outputs:u,encoder_attention_mask:l}=await wa(this,o),p=u.dims[1]/this.config.reduction_factor,f=Math.floor(p*i),m=Math.floor(p*a),c=this.config.num_mel_bins;let y=[],w=null,v=null,k=0;for(;;){++k;const T=Ig(!!v);let A;v?A=v.output_sequence_out:A=new pe("float32",new Float32Array(c),[1,1,c]);let B={use_cache_branch:T,output_sequence:A,encoder_attention_mask:l,speaker_embeddings:r,encoder_hidden_states:u};this.addPastKeyValues(B,w),v=await Or(this.sessions.decoder_model_merged,B),w=this.getPastKeyValues(v,w);const{prob:R,spectrum:D}=v;if(y.push(D),k>=m&&(Array.from(R.data).filter(K=>K>=n).length>0||k>=f))break}const $=dr(y),{waveform:C}=await Or(s.sessions.model,{spectrogram:$});return{spectrogram:$,waveform:C}}}class Qx extends ee{main_input_name="spectrogram"}class Jx extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class Zx extends Jx{}class k0 extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class eS extends k0{}class tS extends k0{}class E0 extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class rS extends E0{}class nS extends E0{}class C0 extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class aS extends C0{}class iS extends C0{}class gu extends ee{}class sS extends gu{}class oS extends gu{static async from_pretrained(e,r={}){return r.model_file_name??="text_model",super.from_pretrained(e,r)}}class uS extends gu{static async from_pretrained(e,r={}){return r.model_file_name??="audio_model",super.from_pretrained(e,r)}}class lS extends ee{}class T0 extends lS{async _call(e){return new GS(await super._call(e))}}class I0 extends ee{}class dS extends I0{}class cS extends I0{}class A0 extends ee{constructor(e,r,n){super(e,r),this.generation_config=n}}class pS extends A0{}class hS extends A0{}class M0 extends ee{}class fS extends M0{}class mS extends M0{async _call(e){return new Me(await super._call(e))}}class O0 extends ee{forward_params=["input_ids","attention_mask","encoder_outputs","decoder_input_ids","decoder_attention_mask","past_key_values"];constructor(e,r,n){super(e,r),this.generation_config=n}_apply_and_filter_by_delay_pattern_mask(e){const[r,n]=e.dims,a=this.config.decoder.num_codebooks,i=n-a;let s=0;for(let l=0;l<e.size;++l){if(e.data[l]===this.config.decoder.pad_token_id)continue;const p=l%n,f=Math.floor(l/n)%a,m=p-f;m>0&&m<=i&&(e.data[s++]=e.data[l])}const o=Math.floor(r/a),u=s/(o*a);return new pe(e.type,e.data.slice(0,s),[o,a,u])}prepare_inputs_for_generation(e,r,n){let a=structuredClone(e);for(let s=0;s<a.length;++s)for(let o=0;o<a[s].length;++o)s%this.config.decoder.num_codebooks>=o&&(a[s][o]=BigInt(this.config.decoder.pad_token_id));return n.guidance_scale!==null&&n.guidance_scale>1&&(a=a.concat(a)),super.prepare_inputs_for_generation(a,r,n)}async generate(e){const r=await super.generate(e),n=this._apply_and_filter_by_delay_pattern_mask(r).unsqueeze_(0),{audio_values:a}=await Or(this.sessions.encodec_decode,{audio_codes:n});return a}}class gS{static MODEL_CLASS_MAPPINGS=null;static BASE_IF_FAIL=!1;static async from_pretrained(e,{progress_callback:r=null,config:n=null,cache_dir:a=null,local_files_only:i=!1,revision:s="main",model_file_name:o=null,subfolder:u="onnx",device:l=null,dtype:p=null,use_external_data_format:f=null,session_options:m={}}={}){let c={progress_callback:r,config:n,cache_dir:a,local_files_only:i,revision:s,model_file_name:o,subfolder:u,device:l,dtype:p,use_external_data_format:f,session_options:m};if(c.config=await Sg.from_pretrained(e,c),!this.MODEL_CLASS_MAPPINGS)throw new Error("`MODEL_CLASS_MAPPINGS` not implemented for this type of `AutoClass`: "+this.name);for(let y of this.MODEL_CLASS_MAPPINGS){const w=y.get(c.config.model_type);if(w)return await w[1].from_pretrained(e,c)}if(this.BASE_IF_FAIL)return console.warn(`Unknown model class "${c.config.model_type}", attempting to construct from base class.`),await ee.from_pretrained(e,c);throw Error(`Unsupported model type: ${c.config.model_type}`)}}const _S=new Map([["bert",["BertModel",Kv]],["nomic_bert",["NomicBertModel",e2]],["roformer",["RoFormerModel",t2]],["electra",["ElectraModel",c2]],["esm",["EsmModel",B2]],["convbert",["ConvBertModel",s2]],["camembert",["CamembertModel",g2]],["deberta",["DebertaModel",v2]],["deberta-v2",["DebertaV2Model",E2]],["mpnet",["MPNetModel",G2]],["albert",["AlbertModel",Z2]],["distilbert",["DistilBertModel",M2]],["roberta",["RobertaModel",b1]],["xlm",["XLMModel",k1]],["xlm-roberta",["XLMRobertaModel",A1]],["clap",["ClapModel",sS]],["clip",["CLIPModel",W1]],["clipseg",["CLIPSegModel",K1]],["chinese_clip",["ChineseCLIPModel",q1]],["siglip",["SiglipModel",V1]],["mobilebert",["MobileBertModel",L2]],["squeezebert",["SqueezeBertModel",Y2]],["wav2vec2",["Wav2Vec2Model",Cx]],["wav2vec2-bert",["Wav2Vec2BertModel",Nx]],["unispeech",["UniSpeechModel",Mx]],["unispeech-sat",["UniSpeechSatModel",Px]],["hubert",["HubertModel",Ux]],["wavlm",["WavLMModel",Gx]],["audio-spectrogram-transformer",["ASTModel",R1]],["vits",["VitsModel",T0]],["detr",["DetrModel",W$]],["table-transformer",["TableTransformerModel",j$]],["vit",["ViTModel",E$]],["fastvit",["FastViTModel",T$]],["mobilevit",["MobileViTModel",O$]],["mobilevitv2",["MobileViTV2Model",P$]],["owlvit",["OwlViTModel",B$]],["owlv2",["Owlv2Model",N$]],["beit",["BeitModel",L$]],["deit",["DeiTModel",Y$]],["convnext",["ConvNextModel",cx]],["convnextv2",["ConvNextV2Model",hx]],["dinov2",["Dinov2Model",mx]],["resnet",["ResNetModel",Q$]],["swin",["SwinModel",Z$]],["swin2sr",["Swin2SRModel",tx]],["donut-swin",["DonutSwinModel",dx]],["yolos",["YolosModel",_x]],["dpt",["DPTModel",nx]],["glpn",["GLPNModel",ox]],["hifigan",["SpeechT5HifiGan",Qx]],["efficientnet",["EfficientNetModel",fS]]]),yS=new Map([["t5",["T5Model",n1]],["longt5",["LongT5Model",i1]],["mt5",["MT5Model",o1]],["bart",["BartModel",l1]],["mbart",["MBartModel",p1]],["marian",["MarianModel",xx]],["whisper",["WhisperModel",D1]],["m2m_100",["M2M100Model",kx]],["blenderbot",["BlenderbotModel",g1]],["blenderbot-small",["BlenderbotSmallModel",y1]]]),wS=new Map([["bloom",["BloomModel",b$]],["gpt2",["GPT2Model",X1]],["gptj",["GPTJModel",r$]],["gpt_bigcode",["GPTBigCodeModel",a$]],["gpt_neo",["GPTNeoModel",J1]],["gpt_neox",["GPTNeoXModel",e$]],["codegen",["CodeGenModel",s$]],["llama",["LlamaModel",u$]],["gemma",["GemmaModel",d$]],["openelm",["OpenELMModel",p$]],["qwen2",["Qwen2Model",f$]],["phi",["PhiModel",g$]],["phi3",["Phi3Model",y$]],["mpt",["MptModel",$$]],["opt",["OPTModel",S$]],["mistral",["MistralModel",eS]],["starcoder2",["Starcoder2Model",rS]],["falcon",["FalconModel",aS]],["stablelm",["StableLmModel",pS]]]),z0=new Map([["speecht5",["SpeechT5ForSpeechToText",Yx]],["whisper",["WhisperForConditionalGeneration",N1]]]),bS=new Map([["speecht5",["SpeechT5ForTextToSpeech",Xx]]]),vS=new Map([["vits",["VitsModel",T0]],["musicgen",["MusicgenForConditionalGeneration",O0]]]),$S=new Map([["bert",["BertForSequenceClassification",Xv]],["roformer",["RoFormerForSequenceClassification",n2]],["electra",["ElectraForSequenceClassification",h2]],["esm",["EsmForSequenceClassification",N2]],["convbert",["ConvBertForSequenceClassification",u2]],["camembert",["CamembertForSequenceClassification",y2]],["deberta",["DebertaForSequenceClassification",x2]],["deberta-v2",["DebertaV2ForSequenceClassification",T2]],["mpnet",["MPNetForSequenceClassification",j2]],["albert",["AlbertForSequenceClassification",e1]],["distilbert",["DistilBertForSequenceClassification",O2]],["roberta",["RobertaForSequenceClassification",$1]],["xlm",["XLMForSequenceClassification",C1]],["xlm-roberta",["XLMRobertaForSequenceClassification",O1]],["bart",["BartForSequenceClassification",c1]],["mbart",["MBartForSequenceClassification",f1]],["mobilebert",["MobileBertForSequenceClassification",W2]],["squeezebert",["SqueezeBertForSequenceClassification",Q2]]]),xS=new Map([["bert",["BertForTokenClassification",Qv]],["roformer",["RoFormerForTokenClassification",a2]],["electra",["ElectraForTokenClassification",f2]],["esm",["EsmForTokenClassification",F2]],["convbert",["ConvBertForTokenClassification",l2]],["camembert",["CamembertForTokenClassification",w2]],["deberta",["DebertaForTokenClassification",S2]],["deberta-v2",["DebertaV2ForTokenClassification",I2]],["mpnet",["MPNetForTokenClassification",q2]],["distilbert",["DistilBertForTokenClassification",z2]],["roberta",["RobertaForTokenClassification",x1]],["xlm",["XLMForTokenClassification",T1]],["xlm-roberta",["XLMRobertaForTokenClassification",z1]]]),P0=new Map([["t5",["T5ForConditionalGeneration",a1]],["longt5",["LongT5ForConditionalGeneration",s1]],["mt5",["MT5ForConditionalGeneration",u1]],["bart",["BartForConditionalGeneration",d1]],["mbart",["MBartForConditionalGeneration",h1]],["marian",["MarianMTModel",Sx]],["m2m_100",["M2M100ForConditionalGeneration",Ex]],["blenderbot",["BlenderbotForConditionalGeneration",_1]],["blenderbot-small",["BlenderbotSmallForConditionalGeneration",w1]]]),R0=new Map([["bloom",["BloomForCausalLM",v$]],["gpt2",["GPT2LMHeadModel",Q1]],["gptj",["GPTJForCausalLM",n$]],["gpt_bigcode",["GPTBigCodeForCausalLM",i$]],["gpt_neo",["GPTNeoForCausalLM",Z1]],["gpt_neox",["GPTNeoXForCausalLM",t$]],["codegen",["CodeGenForCausalLM",o$]],["llama",["LlamaForCausalLM",l$]],["gemma",["GemmaForCausalLM",c$]],["openelm",["OpenELMForCausalLM",h$]],["qwen2",["Qwen2ForCausalLM",m$]],["phi",["PhiForCausalLM",_$]],["phi3",["Phi3ForCausalLM",w$]],["mpt",["MptForCausalLM",x$]],["opt",["OPTForCausalLM",k$]],["mbart",["MBartForCausalLM",m1]],["mistral",["MistralForCausalLM",tS]],["starcoder2",["Starcoder2ForCausalLM",nS]],["falcon",["FalconForCausalLM",iS]],["trocr",["TrOCRForCausalLM",Zx]],["stablelm",["StableLmForCausalLM",hS]]]),SS=new Map([["bert",["BertForMaskedLM",Yv]],["roformer",["RoFormerForMaskedLM",r2]],["electra",["ElectraForMaskedLM",p2]],["esm",["EsmForMaskedLM",D2]],["convbert",["ConvBertForMaskedLM",o2]],["camembert",["CamembertForMaskedLM",_2]],["deberta",["DebertaForMaskedLM",$2]],["deberta-v2",["DebertaV2ForMaskedLM",C2]],["mpnet",["MPNetForMaskedLM",H2]],["albert",["AlbertForMaskedLM",r1]],["distilbert",["DistilBertForMaskedLM",R2]],["roberta",["RobertaForMaskedLM",v1]],["xlm",["XLMWithLMHeadModel",E1]],["xlm-roberta",["XLMRobertaForMaskedLM",M1]],["mobilebert",["MobileBertForMaskedLM",U2]],["squeezebert",["SqueezeBertForMaskedLM",X2]]]),kS=new Map([["bert",["BertForQuestionAnswering",Jv]],["roformer",["RoFormerForQuestionAnswering",i2]],["electra",["ElectraForQuestionAnswering",m2]],["convbert",["ConvBertForQuestionAnswering",d2]],["camembert",["CamembertForQuestionAnswering",b2]],["deberta",["DebertaForQuestionAnswering",k2]],["deberta-v2",["DebertaV2ForQuestionAnswering",A2]],["mpnet",["MPNetForQuestionAnswering",K2]],["albert",["AlbertForQuestionAnswering",t1]],["distilbert",["DistilBertForQuestionAnswering",P2]],["roberta",["RobertaForQuestionAnswering",S1]],["xlm",["XLMForQuestionAnswering",I1]],["xlm-roberta",["XLMRobertaForQuestionAnswering",P1]],["mobilebert",["MobileBertForQuestionAnswering",V2]],["squeezebert",["SqueezeBertForQuestionAnswering",J2]]]),B0=new Map([["vision-encoder-decoder",["VisionEncoderDecoderModel",F1]]]),ES=new Map([["llava",["LlavaForConditionalGeneration",Dg]],["moondream1",["Moondream1ForConditionalGeneration",U1]]]),CS=new Map([["vit",["ViTForImageClassification",C$]],["fastvit",["FastViTForImageClassification",I$]],["mobilevit",["MobileViTForImageClassification",z$]],["mobilevitv2",["MobileViTV2ForImageClassification",R$]],["beit",["BeitForImageClassification",U$]],["deit",["DeiTForImageClassification",X$]],["convnext",["ConvNextForImageClassification",px]],["convnextv2",["ConvNextV2ForImageClassification",fx]],["dinov2",["Dinov2ForImageClassification",gx]],["resnet",["ResNetForImageClassification",J$]],["swin",["SwinForImageClassification",ex]],["segformer",["SegformerForImageClassification",dS]],["efficientnet",["EfficientNetForImageClassification",mS]]]),TS=new Map([["detr",["DetrForObjectDetection",V$]],["table-transformer",["TableTransformerForObjectDetection",q$]],["yolos",["YolosForObjectDetection",yx]]]),IS=new Map([["owlvit",["OwlViTForObjectDetection",D$]],["owlv2",["Owlv2ForObjectDetection",F$]]]),AS=new Map([["detr",["DetrForSegmentation",G$]],["clipseg",["CLIPSegForImageSegmentation",Y1]]]),MS=new Map([["segformer",["SegformerForSemanticSegmentation",cS]]]),OS=new Map([["sam",["SamModel",vx]]]),zS=new Map([["wav2vec2",["Wav2Vec2ForCTC",Tx]],["wav2vec2-bert",["Wav2Vec2BertForCTC",Fx]],["unispeech",["UniSpeechForCTC",Ox]],["unispeech-sat",["UniSpeechSatForCTC",Rx]],["wavlm",["WavLMForCTC",Hx]],["hubert",["HubertForCTC",Wx]]]),PS=new Map([["wav2vec2",["Wav2Vec2ForSequenceClassification",Ix]],["wav2vec2-bert",["Wav2Vec2BertForSequenceClassification",Lx]],["unispeech",["UniSpeechForSequenceClassification",zx]],["unispeech-sat",["UniSpeechSatForSequenceClassification",Bx]],["wavlm",["WavLMForSequenceClassification",jx]],["hubert",["HubertForSequenceClassification",Vx]],["audio-spectrogram-transformer",["ASTForAudioClassification",B1]]]),RS=new Map([["wavlm",["WavLMForXVector",qx]]]),BS=new Map([["unispeech-sat",["UniSpeechSatForAudioFrameClassification",Dx]],["wavlm",["WavLMForAudioFrameClassification",Kx]],["wav2vec2",["Wav2Vec2ForAudioFrameClassification",Ax]]]),DS=new Map([["vitmatte",["VitMatteForImageMatting",M$]]]),NS=new Map([["swin2sr",["Swin2SRForImageSuperResolution",rx]]]),FS=new Map([["dpt",["DPTForDepthEstimation",ax]],["depth_anything",["DepthAnythingForDepthEstimation",sx]],["glpn",["GLPNForDepthEstimation",ux]]]),LS=new Map([["clip",["CLIPVisionModelWithProjection",Fg]],["siglip",["SiglipVisionModel",H1]]]),D0=[[_S,ve.EncoderOnly],[yS,ve.EncoderDecoder],[wS,ve.DecoderOnly],[$S,ve.EncoderOnly],[xS,ve.EncoderOnly],[P0,ve.Seq2Seq],[z0,ve.Seq2Seq],[R0,ve.DecoderOnly],[SS,ve.EncoderOnly],[kS,ve.EncoderOnly],[B0,ve.Vision2Seq],[ES,ve.ImageTextToText],[CS,ve.EncoderOnly],[AS,ve.EncoderOnly],[MS,ve.EncoderOnly],[DS,ve.EncoderOnly],[NS,ve.EncoderOnly],[FS,ve.EncoderOnly],[TS,ve.EncoderOnly],[IS,ve.EncoderOnly],[OS,ve.MaskGeneration],[zS,ve.EncoderOnly],[PS,ve.EncoderOnly],[bS,ve.Seq2Seq],[vS,ve.EncoderOnly],[RS,ve.EncoderOnly],[BS,ve.EncoderOnly],[LS,ve.EncoderOnly]];for(const[t,e]of D0)for(const[r,n]of t.values())qi.set(r,e),pa.set(n,r),Eg.set(r,n);const US=[["MusicgenForConditionalGeneration",O0,ve.Musicgen],["CLIPTextModelWithProjection",Ng,ve.EncoderOnly],["SiglipTextModel",G1,ve.EncoderOnly],["ClapTextModelWithProjection",oS,ve.EncoderOnly],["ClapAudioModelWithProjection",uS,ve.EncoderOnly]];for(const[t,e,r]of US)qi.set(t,r),pa.set(e,t),Eg.set(t,e);class N3 extends gS{static MODEL_CLASS_MAPPINGS=D0.map(e=>e[0]);static BASE_IF_FAIL=!0}class Me extends Wt{constructor({logits:e}){super(),this.logits=e}}class WS extends Wt{constructor({logits:e,embeddings:r}){super(),this.logits=e,this.embeddings=r}}class mt extends Wt{constructor({logits:e}){super(),this.logits=e}}class gt extends Wt{constructor({logits:e}){super(),this.logits=e}}class xt extends Wt{constructor({start_logits:e,end_logits:r}){super(),this.start_logits=e,this.end_logits=r}}class On extends Wt{constructor({logits:e}){super(),this.logits=e}}class VS extends Wt{constructor({alphas:e}){super(),this.alphas=e}}class GS extends Wt{constructor({waveform:e,spectrogram:r}){super(),this.waveform=e,this.spectrogram=r}}const Nt=typeof self<"u",HS=Nt&&self.constructor.name==="DedicatedWorkerGlobalScope";let Gr,N0,Ir;if(Nt)Gr=(t,e)=>{if(!self.OffscreenCanvas)throw new Error("OffscreenCanvas not supported by this browser.");return new self.OffscreenCanvas(t,e)},Ir=self.createImageBitmap,N0=self.ImageData;else if(Ze)Ir=async t=>{const r=(await t.metadata()).channels,{data:n,info:a}=await t.rotate().raw().toBuffer({resolveWithObject:!0}),i=new tr(new Uint8ClampedArray(n),a.width,a.height,a.channels);return r!==void 0&&r!==a.channels&&i.convert(r),i};else throw new Error("Unable to load image processing library.");const jS={0:"nearest",1:"lanczos",2:"bilinear",3:"bicubic",4:"box",5:"hamming"},qS=new Map([["png","image/png"],["jpg","image/jpeg"],["jpeg","image/jpeg"],["gif","image/gif"]]);class tr{constructor(e,r,n,a){this.data=e,this.width=r,this.height=n,this.channels=a}get size(){return[this.width,this.height]}static async read(e){if(e instanceof tr)return e;if(typeof e=="string"||e instanceof URL)return await this.fromURL(e);throw new Error(`Unsupported input type: ${typeof e}`)}static fromCanvas(e){if(!Nt)throw new Error("fromCanvas() is only supported in browser environments.");const n=e.getContext("2d").getImageData(0,0,e.width,e.height).data;return new tr(n,e.width,e.height,4)}static async fromURL(e){const r=await fo(e);if(r.status!==200)throw new Error(`Unable to read image from "${e}" (${r.status} ${r.statusText})`);const n=await r.blob();return this.fromBlob(n)}static async fromBlob(e){if(Nt){const r=await Ir(e),n=Gr(r.width,r.height).getContext("2d");return n.drawImage(r,0,0),new this(n.getImageData(0,0,r.width,r.height).data,r.width,r.height,4)}else{const r=Ze(await e.arrayBuffer());return await Ir(r)}}static fromTensor(e,r="CHW"){if(e.dims.length!==3)throw new Error(`Tensor should have 3 dimensions, but has ${e.dims.length} dimensions.`);if(r==="CHW")e=e.transpose(1,2,0);else if(r!=="HWC")throw new Error(`Unsupported channel format: ${r}`);if(!(e.data instanceof Uint8ClampedArray||e.data instanceof Uint8Array))throw new Error(`Unsupported tensor type: ${e.type}`);switch(e.dims[2]){case 1:case 2:case 3:case 4:return new tr(e.data,e.dims[1],e.dims[0],e.dims[2]);default:throw new Error(`Unsupported number of channels: ${e.dims[2]}`)}}grayscale(){if(this.channels===1)return this;const e=new Uint8ClampedArray(this.width*this.height*1);switch(this.channels){case 3:case 4:for(let r=0,n=0;r<this.data.length;r+=this.channels){const a=this.data[r],i=this.data[r+1],s=this.data[r+2];e[n++]=Math.round(.2989*a+.587*i+.114*s)}break;default:throw new Error(`Conversion failed due to unsupported number of channels: ${this.channels}`)}return this._update(e,this.width,this.height,1)}rgb(){if(this.channels===3)return this;const e=new Uint8ClampedArray(this.width*this.height*3);switch(this.channels){case 1:for(let r=0,n=0;r<this.data.length;++r)e[n++]=this.data[r],e[n++]=this.data[r],e[n++]=this.data[r];break;case 4:for(let r=0,n=0;r<this.data.length;r+=4)e[n++]=this.data[r],e[n++]=this.data[r+1],e[n++]=this.data[r+2];break;default:throw new Error(`Conversion failed due to unsupported number of channels: ${this.channels}`)}return this._update(e,this.width,this.height,3)}rgba(){if(this.channels===4)return this;const e=new Uint8ClampedArray(this.width*this.height*4);switch(this.channels){case 1:for(let r=0,n=0;r<this.data.length;++r)e[n++]=this.data[r],e[n++]=this.data[r],e[n++]=this.data[r],e[n++]=255;break;case 3:for(let r=0,n=0;r<this.data.length;r+=3)e[n++]=this.data[r],e[n++]=this.data[r+1],e[n++]=this.data[r+2],e[n++]=255;break;default:throw new Error(`Conversion failed due to unsupported number of channels: ${this.channels}`)}return this._update(e,this.width,this.height,4)}async resize(e,r,{resample:n=2}={}){let a=jS[n]??n;if(Nt){const i=this.channels,s=this.toCanvas(),o=Gr(e,r).getContext("2d");return o.drawImage(s,0,0,e,r),new tr(o.getImageData(0,0,e,r).data,e,r,4).convert(i)}else{let i=this.toSharp();switch(a){case"box":case"hamming":(a==="box"||a==="hamming")&&(console.warn(`Resampling method ${a} is not yet supported. Using bilinear instead.`),a="bilinear");case"nearest":case"bilinear":case"bicubic":i=i.affine([e/this.width,0,0,r/this.height],{interpolator:a});break;case"lanczos":i=i.resize({width:e,height:r,fit:"fill",kernel:"lanczos3"});break;default:throw new Error(`Resampling method ${a} is not supported.`)}return await Ir(i)}}async pad([e,r,n,a]){if(e=Math.max(e,0),r=Math.max(r,0),n=Math.max(n,0),a=Math.max(a,0),e===0&&r===0&&n===0&&a===0)return this;if(Nt){const i=this.channels,s=this.toCanvas(),o=this.width+e+r,u=this.height+n+a,l=Gr(o,u).getContext("2d");return l.drawImage(s,0,0,this.width,this.height,e,n,o,u),new tr(l.getImageData(0,0,o,u).data,o,u,4).convert(i)}else{const i=this.toSharp().extend({left:e,right:r,top:n,bottom:a});return await Ir(i)}}async crop([e,r,n,a]){if(e=Math.max(e,0),r=Math.max(r,0),n=Math.min(n,this.width-1),a=Math.min(a,this.height-1),e===0&&r===0&&n===this.width-1&&a===this.height-1)return this;const i=n-e+1,s=a-r+1;if(Nt){const o=this.channels,u=this.toCanvas(),l=Gr(i,s).getContext("2d");return l.drawImage(u,e,r,i,s,0,0,i,s),new tr(l.getImageData(0,0,i,s).data,i,s,4).convert(o)}else{const o=this.toSharp().extract({left:e,top:r,width:i,height:s});return await Ir(o)}}async center_crop(e,r){if(this.width===e&&this.height===r)return this;const n=(this.width-e)/2,a=(this.height-r)/2;if(Nt){const i=this.channels,s=this.toCanvas(),o=Gr(e,r).getContext("2d");let u=0,l=0,p=0,f=0;return n>=0?u=n:p=-n,a>=0?l=a:f=-a,o.drawImage(s,u,l,e,r,p,f,e,r),new tr(o.getImageData(0,0,e,r).data,e,r,4).convert(i)}else{let i=this.toSharp();if(n>=0&&a>=0)i=i.extract({left:Math.floor(n),top:Math.floor(a),width:e,height:r});else if(n<=0&&a<=0){const s=Math.floor(-a),o=Math.floor(-n);i=i.extend({top:s,left:o,right:e-this.width-o,bottom:r-this.height-s})}else{let s=[0,0],o=0;a<0?(s[0]=Math.floor(-a),s[1]=r-this.height-s[0]):o=Math.floor(a);let u=[0,0],l=0;n<0?(u[0]=Math.floor(-n),u[1]=e-this.width-u[0]):l=Math.floor(n),i=i.extend({top:s[0],bottom:s[1],left:u[0],right:u[1]}).extract({left:l,top:o,width:e,height:r})}return await Ir(i)}}async toBlob(e="image/png",r=1){if(!Nt)throw new Error("toBlob() is only supported in browser environments.");return await this.toCanvas().convertToBlob({type:e,quality:r})}toTensor(e="CHW"){let r=new pe("uint8",new Uint8Array(this.data),[this.height,this.width,this.channels]);if(e!=="HWC")if(e==="CHW")r=r.permute(2,0,1);else throw new Error(`Unsupported channel format: ${e}`);return r}toCanvas(){if(!Nt)throw new Error("toCanvas() is only supported in browser environments.");const e=this.clone().rgba(),r=Gr(e.width,e.height),n=new N0(e.data,e.width,e.height);return r.getContext("2d").putImageData(n,0,0),r}_update(e,r,n,a=null){return this.data=e,this.width=r,this.height=n,a!==null&&(this.channels=a),this}clone(){return new tr(this.data.slice(),this.width,this.height,this.channels)}convert(e){if(this.channels===e)return this;switch(e){case 1:this.grayscale();break;case 3:this.rgb();break;case 4:this.rgba();break;default:throw new Error(`Conversion failed due to unsupported number of channels: ${this.channels}`)}return this}async save(e){if(Nt){if(HS)throw new Error("Unable to save an image from a Web Worker.");const r=e.split(".").pop().toLowerCase(),n=qS.get(r)??"image/png",a=await this.toBlob(n),i=URL.createObjectURL(a),s=document.createElement("a");s.href=i,s.download=e,s.click(),s.remove()}else{if(wt.useFS)return await this.toSharp().toFile(e);throw new Error("Unable to save the image because filesystem is disabled in this environment.")}}toSharp(){if(Nt)throw new Error("toSharp() is only supported in server-side environments.");return Ze(this.data,{raw:{width:this.width,height:this.height,channels:this.channels}})}}function _p(t){if(t<1)return new Float64Array;if(t===1)return new Float64Array([1]);const e=t-1,r=Math.PI/e,n=new Float64Array(t);for(let a=0;a<t;++a){const i=2*a-e;n[a]=.5+.5*Math.cos(r*i)}return n}const KS={htk:t=>2595*Math.log10(1+t/700),kaldi:t=>1127*Math.log(1+t/700),slaney:(t,e=1e3,r=15,n=27/Math.log(6.4))=>t>=e?r+Math.log(t/e)*n:3*t/200};function so(t,e="htk"){const r=KS[e];if(!r)throw new Error('mel_scale should be one of "htk", "slaney" or "kaldi".');return typeof t=="number"?r(t):t.map(n=>r(n))}const YS={htk:t=>700*(10**(t/2595)-1),kaldi:t=>700*(Math.exp(t/1127)-1),slaney:(t,e=1e3,r=15,n=Math.log(6.4)/27)=>t>=r?e*Math.exp(n*(t-r)):200*t/3};function XS(t,e="htk"){const r=YS[e];if(!r)throw new Error('mel_scale should be one of "htk", "slaney" or "kaldi".');return typeof t=="number"?r(t):t.map(n=>r(n))}function QS(t,e){const r=Float64Array.from({length:e.length-1},(s,o)=>e[o+1]-e[o]),n=Array.from({length:t.length},()=>new Array(e.length));for(let s=0;s<t.length;++s){const o=n[s];for(let u=0;u<e.length;++u)o[u]=e[u]-t[s]}const a=e.length-2,i=Array.from({length:a},()=>new Array(t.length));for(let s=0;s<t.length;++s){const o=n[s];for(let u=0;u<a;++u){const l=-o[u]/r[u],p=o[u+2]/r[u+1];i[u][s]=Math.max(0,Math.min(l,p))}}return i}function yp(t,e,r){const n=(e-t)/(r-1);return Float64Array.from({length:r},(a,i)=>t+n*i)}function ba(t,e,r,n,a,i=null,s="htk",o=!1){if(i!==null&&i!=="slaney")throw new Error('norm must be one of null or "slaney"');const u=so(r,s),l=so(n,s),p=yp(u,l,e+2);let f=XS(p,s),m;if(o){const y=a/(t*2);m=so(Float64Array.from({length:t},(w,v)=>v*y),s),f=p}else m=yp(0,Math.floor(a/2),t);const c=QS(m,f);if(i!==null&&i==="slaney")for(let y=0;y<e;++y){const w=c[y],v=2/(f[y+2]-f[y]);for(let k=0;k<t;++k)w[k]*=v}return c}function JS(t,e,r){const n=new t.constructor(t.length+e+r),a=t.length-1;for(let i=0;i<t.length;++i)n[e+i]=t[i];for(let i=1;i<=e;++i)n[e-i]=t[Ri(i,a)];for(let i=1;i<=r;++i)n[a+e+i]=t[Ri(a-i,a)];return n}function F0(t,e,r,n,a){if(r<=0)throw new Error("reference must be greater than zero");if(n<=0)throw new Error("min_value must be greater than zero");r=Math.max(n,r);const i=Math.log10(r);for(let s=0;s<t.length;++s)t[s]=e*Math.log10(Math.max(n,t[s])-i);if(a!==null){if(a<=0)throw new Error("db_range must be greater than zero");const s=vr(t)[0]-a;for(let o=0;o<t.length;++o)t[o]=Math.max(t[o],s)}return t}function ZS(t,e=1,r=1e-5,n=null){return F0(t,20,e,r,n)}function e3(t,e=1,r=1e-10,n=null){return F0(t,10,e,r,n)}function ss(t,e,r,n,{fft_length:a=null,power:i=1,center:s=!0,pad_mode:o="reflect",onesided:u=!0,preemphasis:l=null,mel_filters:p=null,mel_floor:f=1e-10,log_mel:m=null,reference:c=1,min_value:y=1e-10,db_range:w=null,remove_dc_offset:v=null,max_num_frames:k=null,do_pad:$=!0,transpose:C=!1}={}){const T=e.length;if(a===null&&(a=r),r>a)throw Error(`frame_length (${r}) may not be larger than fft_length (${a})`);if(T!==r)throw new Error(`Length of the window (${T}) must equal frame_length (${r})`);if(n<=0)throw new Error("hop_length must be greater than zero");if(i===null&&p!==null)throw new Error("You have provided `mel_filters` but `power` is `None`. Mel spectrogram computation is not yet supported for complex-valued spectrogram. Specify `power` to fix this issue.");if(s){if(o!=="reflect")throw new Error(`pad_mode="${o}" not implemented yet.`);const P=Math.floor((a-1)/2)+1;t=JS(t,P,P)}const A=Math.floor(1+Math.floor((t.length-r)/n)),B=u?Math.floor(a/2)+1:a;let R=A,D=A;k!==null&&(k>A?$&&(D=k):D=R=k);const K=new v_(a),j=new Float64Array(a),ie=new Float64Array(K.outputBufferSize),te=new Array(R);for(let P=0;P<R;++P){const H=P*n;for(let G=0;G<r;++G)j[G]=t[H+G];if(v){let G=0;for(let N=0;N<r;++N)G+=j[N];const ne=G/r;for(let N=0;N<r;++N)j[N]-=ne}if(l!==null){for(let G=r-1;G>=1;--G)j[G]-=l*j[G-1];j[0]*=1-l}for(let G=0;G<e.length;++G)j[G]*=e[G];K.realTransform(ie,j);const le=new Array(B);for(let G=0;G<le.length;++G){const ne=G<<1;le[G]=ie[ne]**2+ie[ne+1]**2}te[P]=le}if(i!==null&&i!==2){const P=2/i;for(let H=0;H<te.length;++H){const le=te[H];for(let G=0;G<le.length;++G)le[G]**=P}}const oe=p.length,re=new Float32Array(oe*D),M=C?[D,oe]:[oe,D];for(let P=0;P<oe;++P){const H=p[P];for(let le=0;le<R;++le){const G=te[le];let ne=0;for(let N=0;N<B;++N)ne+=H[N]*G[N];re[C?le*oe+P:P*R+le]=Math.max(f,ne)}}if(i!==null&&m!==null){const P=Math.min(re.length,R*oe);switch(m){case"log":for(let H=0;H<P;++H)re[H]=Math.log(re[H]);break;case"log10":for(let H=0;H<P;++H)re[H]=Math.log10(re[H]);break;case"dB":if(i===1)ZS(re,c,y,w);else if(i===2)e3(re,c,y,w);else throw new Error(`Cannot use log_mel option '${m}' with power ${i}`);break;default:throw new Error(`log_mel must be one of null, 'log', 'log10' or 'dB'. Got '${m}'`)}}return{data:re,dims:M}}function os(t,e,{periodic:r=!0,frame_length:n=null,center:a=!0}={}){const i=r?t+1:t;let s;switch(e){case"boxcar":s=new Float64Array(i).fill(1);break;case"hann":case"hann_window":s=_p(i);break;case"povey":s=_p(i).map(o=>Math.pow(o,.85));break;default:throw new Error(`Unknown window type ${e}.`)}if(r&&(s=s.subarray(0,t)),n===null)return s;if(t>n)throw new Error(`Length of the window (${t}) may not be larger than frame_length (${n})`);return s}function t3([t,e,r,n]){return[t-r/2,e-n/2,t+r/2,e+n/2]}function _u(t,e=.5,r=null,n=!1){const a=t.logits,i=t.pred_boxes,[s,o,u]=a.dims;if(r!==null&&r.length!==s)throw Error("Make sure that you pass in as many target sizes as the batch dimension of the logits");let l=[];for(let p=0;p<s;++p){let f=r!==null?r[p]:null,m={boxes:[],classes:[],scores:[]},c=a[p],y=i[p];for(let w=0;w<o;++w){let v=c[w],k=[],$;if(n){$=v.sigmoid().data;for(let C=0;C<$.length;++C)$[C]>e&&k.push(C)}else{let C=vr(v.data)[1];if(C===u-1||($=Tn(v.data),$[C]<e))continue;k.push(C)}for(const C of k){let T=y[w].data;T=t3(T),f!==null&&(T=T.map((A,B)=>A*f[(B+1)%2])),m.boxes.push(T),m.classes.push(C),m.scores.push($[C])}}l.push(m)}return l}function Fa(t,e){if(!(t instanceof Float32Array||t instanceof Float64Array))throw new Error(`${e} expects input to be a Float32Array or a Float64Array, but got ${t?.constructor?.name??typeof t} instead. If using the feature extractor directly, remember to use \`read_audio(url, sampling_rate)\` to obtain the raw audio data of the file/url.`)}function wp(t,e,r=0,n=null){const a=t/e;let i=x_(a)*e;return n!==null&&i>n&&(i=Math.floor(a)*e),i<r&&(i=Math.ceil(a)*e),i}function oo([t,e],r){return[Math.max(Math.floor(t/r),1)*r,Math.max(Math.floor(e/r),1)*r]}class sn extends $t{constructor(e){super(),this.config=e}}class Xe extends sn{constructor(e){super(e),this.image_mean=this.config.image_mean??this.config.mean,this.image_std=this.config.image_std??this.config.std,this.resample=this.config.resample??2,this.do_rescale=this.config.do_rescale??!0,this.rescale_factor=this.config.rescale_factor??1/255,this.do_normalize=this.config.do_normalize,this.do_resize=this.config.do_resize,this.do_thumbnail=this.config.do_thumbnail,this.size=this.config.size,this.size_divisibility=this.config.size_divisibility??this.config.size_divisor,this.do_center_crop=this.config.do_center_crop,this.crop_size=this.config.crop_size,this.do_convert_rgb=this.config.do_convert_rgb??!0,this.do_crop_margin=this.config.do_crop_margin,this.pad_size=this.config.pad_size,this.do_pad=this.config.do_pad,this.do_pad&&!this.pad_size&&this.size&&this.size.width!==void 0&&this.size.height!==void 0&&(this.pad_size=this.size),this.do_flip_channel_order=this.config.do_flip_channel_order??!1}async thumbnail(e,r,n=2){const a=e.height,i=e.width,s=r.height,o=r.width;let u=Math.min(a,s),l=Math.min(i,o);return u===a&&l===i?e:(a>i?l=Math.floor(i*u/a):i>a&&(u=Math.floor(a*l/i)),await e.resize(l,u,{resample:n}))}async crop_margin(e,r=200){const n=e.clone().grayscale(),a=Mp(n.data)[0],s=vr(n.data)[0]-a;if(s===0)return e;const o=r/255;let u=n.width,l=n.height,p=0,f=0;const m=n.data;for(let c=0;c<n.height;++c){const y=c*n.width;for(let w=0;w<n.width;++w)(m[y+w]-a)/s<o&&(u=Math.min(u,w),l=Math.min(l,c),p=Math.max(p,w),f=Math.max(f,c))}return e=await e.crop([u,l,p,f]),e}pad_image(e,r,n,{mode:a="constant",center:i=!1,constant_values:s=0}={}){const[o,u,l]=r;let p,f;if(typeof n=="number"?(p=n,f=n):(p=n.width,f=n.height),p!==u||f!==o){const m=new Float32Array(p*f*l);if(Array.isArray(s))for(let w=0;w<m.length;++w)m[w]=s[w%l];else s!==0&&m.fill(s);const[c,y]=i?[Math.floor((p-u)/2),Math.floor((f-o)/2)]:[0,0];for(let w=0;w<o;++w){const v=(w+y)*p,k=w*u;for(let $=0;$<u;++$){const C=(v+$+c)*l,T=(k+$)*l;for(let A=0;A<l;++A)m[C+A]=e[T+A]}}if(a==="symmetric"){if(i)throw new Error("`center` padding is not supported when `mode` is set to `symmetric`.");const w=o-1,v=u-1;for(let k=0;k<f;++k){const $=k*p,C=Ri(k,w)*u;for(let T=0;T<p;++T){if(k<o&&T<u)continue;const A=($+T)*l,B=(C+Ri(T,v))*l;for(let R=0;R<l;++R)m[A+R]=e[B+R]}}}e=m,r=[f,p,l]}return[e,r]}rescale(e){for(let r=0;r<e.length;++r)e[r]=this.rescale_factor*e[r]}get_resize_output_image_size(e,r){const[n,a]=e.size;let i,s;if(this.do_thumbnail){const{height:o,width:u}=r;i=Math.min(o,u)}else Number.isInteger(r)?(i=r,s=this.config.max_size??i):r!==void 0&&(i=r.shortest_edge,s=r.longest_edge);if(i!==void 0||s!==void 0){const o=i===void 0?1:Math.max(i/n,i/a),u=n*o,l=a*o,p=s===void 0?1:Math.min(s/u,s/l);let f=Math.floor(Number((u*p).toFixed(2))),m=Math.floor(Number((l*p).toFixed(2)));return this.size_divisibility!==void 0&&([f,m]=oo([f,m],this.size_divisibility)),[f,m]}else if(r!==void 0&&r.width!==void 0&&r.height!==void 0){let o=r.width,u=r.height;if(this.config.keep_aspect_ratio&&this.config.ensure_multiple_of){let l=u/a,p=o/n;Math.abs(1-p)<Math.abs(1-l)?l=p:p=l,u=wp(l*a,this.config.ensure_multiple_of),o=wp(p*n,this.config.ensure_multiple_of)}return[o,u]}else{if(this.size_divisibility!==void 0)return oo([n,a],this.size_divisibility);throw new Error(`Could not resize image due to unsupported \`this.size\` option in config: ${JSON.stringify(r)}`)}}async resize(e){const[r,n]=this.get_resize_output_image_size(e,this.size);return await e.resize(r,n,{resample:this.resample})}async preprocess(e,{do_normalize:r=null,do_pad:n=null,do_convert_rgb:a=null,do_convert_grayscale:i=null,do_flip_channel_order:s=null}={}){this.do_crop_margin&&(e=await this.crop_margin(e));const[o,u]=e.size;if(a??this.do_convert_rgb?e=e.rgb():i&&(e=e.grayscale()),this.do_resize&&(e=await this.resize(e)),this.do_thumbnail&&(e=await this.thumbnail(e,this.size,this.resample)),this.do_center_crop){let c,y;Number.isInteger(this.crop_size)?(c=this.crop_size,y=this.crop_size):(c=this.crop_size.width,y=this.crop_size.height),e=await e.center_crop(c,y)}const l=[e.height,e.width];let p=Float32Array.from(e.data),f=[e.height,e.width,e.channels];if(this.do_rescale&&this.rescale(p),r??this.do_normalize){let c=this.image_mean;Array.isArray(this.image_mean)||(c=new Array(e.channels).fill(c));let y=this.image_std;if(Array.isArray(this.image_std)||(y=new Array(e.channels).fill(c)),c.length!==e.channels||y.length!==e.channels)throw new Error(`When set to arrays, the length of \`image_mean\` (${c.length}) and \`image_std\` (${y.length}) must match the number of channels in the image (${e.channels}).`);for(let w=0;w<p.length;w+=e.channels)for(let v=0;v<e.channels;++v)p[w+v]=(p[w+v]-c[v])/y[v]}if(n??this.do_pad){if(this.pad_size)[p,f]=this.pad_image(p,[e.height,e.width,e.channels],this.pad_size);else if(this.size_divisibility){const[c,y]=oo([f[1],f[0]],this.size_divisibility);[p,f]=this.pad_image(p,f,{width:c,height:y})}}if(s??this.do_flip_channel_order){if(f[2]!==3)throw new Error("Flipping channel order is only supported for RGB images.");for(let c=0;c<p.length;c+=3){const y=p[c];p[c]=p[c+2],p[c+2]=y}}const m=new pe("float32",p,f).permute(2,0,1);return{original_size:[u,o],reshaped_input_size:l,pixel_values:m}}async _call(e,...r){Array.isArray(e)||(e=[e]);const n=await Promise.all(e.map(i=>this.preprocess(i)));return{pixel_values:_a(n.map(i=>i.pixel_values),0),original_sizes:n.map(i=>i.original_size),reshaped_input_sizes:n.map(i=>i.reshaped_input_size)}}}class r3 extends Xe{post_process_semantic_segmentation(e,r=null){const n=e.logits,a=n.dims[0];if(r!==null&&r.length!==a)throw Error("Make sure that you pass in as many target sizes as the batch dimension of the logits");const i=[];for(let s=0;s<a;++s){const o=r!==null?r[s]:null;let u=n[s];o!==null&&(u=og(u,o,"bilinear",!1));const[l,p]=o??u.dims.slice(-2),f=new pe("int32",new Int32Array(l*p),[l,p]),m=u[0].data,c=f.data;for(let k=1;k<u.dims[0];++k){const $=u[k].data;for(let C=0;C<$.length;++C)$[C]>m[C]&&(m[C]=$[C],c[C]=k)}const y=new Array(u.dims[0]),w=f.data;for(let k=0;k<w.length;++k){const $=w[k];y[$]=$}const v=y.filter(k=>k!==void 0);i.push({segmentation:f,labels:v})}return i}}class L0 extends Xe{}class n3 extends L0{}class a3 extends Xe{}class i3 extends Xe{}class U0 extends Xe{}class s3 extends U0{}class o3 extends Xe{}class u3 extends Xe{}class W0 extends Xe{constructor(e){super(e),this.crop_pct=this.config.crop_pct??224/256}async resize(e){const r=this.size?.shortest_edge;if(r===void 0)throw new Error("Size dictionary must contain 'shortest_edge' key.");if(r<384){const n=Math.floor(r/this.crop_pct),[a,i]=this.get_resize_output_image_size(e,{shortest_edge:n});e=await e.resize(a,i,{resample:this.resample}),e=await e.center_crop(r,r)}else e=await e.resize(r,r,{resample:this.resample});return e}}class l3 extends W0{}class d3 extends Xe{}class c3 extends Xe{}class p3 extends Xe{constructor(e){super(e),this.include_top=this.config.include_top??!0,this.include_top&&(this.image_std=this.image_std.map(r=>r*r))}}class V0 extends Xe{}class h3 extends V0{}class G0 extends Xe{post_process_object_detection(...e){return _u(...e)}}class f3 extends G0{}class m3 extends Xe{}class g3 extends Xe{}class H0 extends Xe{pad_image(e,r,n,a={}){const[i,s,o]=r;let u=this.image_mean;Array.isArray(this.image_mean)||(u=new Array(o).fill(u));let l=this.image_std;Array.isArray(l)||(l=new Array(o).fill(u));const p=u.map((f,m)=>-f/l[m]);return super.pad_image(e,r,n,{center:!0,constant_values:p,...a})}}class _3 extends H0{}class y3 extends Xe{async _call(e){const r=await super._call(e),n=[r.pixel_values.dims[0],64,64],a=new pe("int64",new BigInt64Array(n.reduce((i,s)=>i*s)).fill(1n),n);return{...r,pixel_mask:a}}post_process_object_detection(...e){return _u(...e)}remove_low_and_no_objects(e,r,n,a){let i=[],s=[],o=[];for(let u=0;u<e.dims[0];++u){let l=e[u],p=r[u],f=vr(l.data)[1];if(f===a)continue;let c=Tn(l.data)[f];c>n&&(i.push(p),s.push(c),o.push(f))}return[i,s,o]}check_segment_validity(e,r,n,a=.5,i=.8){let s=[],o=0,u=0;const l=r[n].data;for(let f=0;f<e.length;++f)e[f]===n&&(s.push(f),++o),l[f]>=a&&++u;let p=o>0&&u>0;return p&&(p=o/u>i),[p,s]}compute_segments(e,r,n,a,i,s=null,o=null){let[u,l]=o??e[0].dims,p=new pe("int32",new Int32Array(u*l),[u,l]),f=[];if(o!==null)for(let v=0;v<e.length;++v)e[v]=og(e[v],o,"bilinear",!1);let m=new Int32Array(e[0].data.length),c=new Float32Array(e[0].data.length);for(let v=0;v<e.length;++v){let k=r[v];const $=e[v].data;for(let C=0;C<$.length;++C)$[C]*=k,$[C]>c[C]&&(m[C]=v,c[C]=$[C])}let y=0;const w=p.data;for(let v=0;v<n.length;++v){let k=n[v],[$,C]=this.check_segment_validity(m,e,v,a,i);if($){++y;for(let T of C)w[T]=y;f.push({id:y,label_id:k,score:r[v]})}}return[p,f]}post_process_panoptic_segmentation(e,r=.5,n=.5,a=.8,i=null,s=null){i===null&&(console.warn("`label_ids_to_fuse` unset. No instance will be fused."),i=new Set);const o=e.logits,l=e.pred_masks.sigmoid();let[p,f,m]=o.dims;if(m-=1,s!==null&&s.length!==p)throw Error("Make sure that you pass in as many target sizes as the batch dimension of the logits");let c=[];for(let y=0;y<p;++y){let w=s!==null?s[y]:null,v=o[y],k=l[y],[$,C,T]=this.remove_low_and_no_objects(v,k,r,m);if(T.length===0){let[R,D]=w??k.dims.slice(-2),K=new pe("int32",new Int32Array(R*D).fill(-1),[R,D]);c.push({segmentation:K,segments_info:[]});continue}let[A,B]=this.compute_segments($,C,T,n,a,i,w);c.push({segmentation:A,segments_info:B})}return c}post_process_instance_segmentation(){throw Error("Not implemented yet")}}class w3 extends Xe{post_process_object_detection(...e){return _u(...e)}}class b3 extends Xe{reshape_input_points(e,r,n,a=!1){e=structuredClone(e);let i=qu(e);if(i.length===3)a||(i=[1,...i]),e=[e];else if(i.length!==4)throw Error("The input_points must be a 4D tensor of shape `batch_size`, `point_batch_size`, `nb_points_per_image`, `2`.");for(let s=0;s<e.length;++s){let o=r[s],u=n[s],l=[u[0]/o[0],u[1]/o[1]];for(let p=0;p<e[s].length;++p)for(let f=0;f<e[s][p].length;++f)for(let m=0;m<e[s][p][f].length;++m)e[s][p][f][m]*=l[m%2]}return new pe("float32",Float32Array.from(e.flat(1/0)),i)}add_input_labels(e,r){let n=qu(e);if(n.length===2)n=[1,...n],e=[e];else if(n.length!==3)throw Error("The input_points must be a 4D tensor of shape `batch_size`, `point_batch_size`, `nb_points_per_image`, `2`.");if(n.some((a,i)=>a!==r.dims[i]))throw Error(`The first ${n.length} dimensions of 'input_points' and 'input_labels' must be the same.`);return new pe("int64",e.flat(1/0).map(BigInt),n)}async _call(e,{input_points:r=null,input_labels:n=null,input_boxes:a=null}={}){const i=await super._call(e);if(r&&(i.input_points=this.reshape_input_points(r,i.original_sizes,i.reshaped_input_sizes)),n){if(!i.input_points)throw Error("`input_points` must be provided if `input_labels` are provided.");i.input_labels=this.add_input_labels(n,i.input_points)}return a&&(i.input_boxes=this.reshape_input_points(a,i.original_sizes,i.reshaped_input_sizes,!0)),i}async post_process_masks(e,r,n,{mask_threshold:a=0,binarize:i=!0,pad_size:s=null}={}){const o=[];s=s??this.pad_size;const u=[s.height,s.width];for(let l=0;l<r.length;++l){const p=r[l],f=n[l];let m=await ap(e[l],{mode:"bilinear",size:u});if(m=m.slice(null,null,[0,f[0]],[0,f[1]]),m=await ap(m,{mode:"bilinear",size:p}),i){const c=m.data,y=new Uint8Array(c.length);for(let w=0;w<c.length;++w)c[w]>a&&(y[w]=1);m=new pe("bool",y,m.dims)}o.push(m)}return o}generate_crop_boxes(e,r,{crop_n_layers:n=0,overlap_ratio:a=512/1500,points_per_crop:i=32,crop_n_points_downscale_factor:s=1}={}){}}class v3 extends Xe{pad_image(e,r,n,a={}){const[i,s,o]=r;return super.pad_image(e,r,{width:s+(n-s%n)%n,height:i+(n-i%n)%n},{mode:"symmetric",center:!1,constant_values:-1,...a})}}class $3 extends Xe{async _call(e,r){Array.isArray(e)||(e=[e]),Array.isArray(r)||(r=[r]);const n=await Promise.all(e.map(s=>this.preprocess(s))),a=await Promise.all(r.map(s=>this.preprocess(s,{do_normalize:!1,do_convert_rgb:!1,do_convert_grayscale:!0})));return{pixel_values:_a(n.map((s,o)=>dr([s.pixel_values,a[o].pixel_values],0)),0),original_sizes:n.map(s=>s.original_size),reshaped_input_sizes:n.map(s=>s.reshaped_input_size)}}}class x3 extends sn{constructor(e){super(e),this.config.mel_filters??=ba(Math.floor(1+this.config.n_fft/2),this.config.feature_size,0,8e3,this.config.sampling_rate,"slaney","slaney"),this.window=os(this.config.n_fft,"hann")}_extract_fbank_features(e){const{data:r,dims:n}=ss(e,this.window,this.config.n_fft,this.config.hop_length,{power:2,mel_filters:this.config.mel_filters,log_mel:"log10",max_num_frames:this.config.nb_max_frames}),a=vr(r)[0];for(let i=0;i<r.length;++i)r[i]=(Math.max(r[i],a-8)+4)/4;return{data:r,dims:n}}async _call(e){Fa(e,"WhisperFeatureExtractor");let r;e.length>this.config.n_samples?(console.warn("Attempting to extract features for audio longer than 30 seconds. If using a pipeline to extract transcript from a long audio clip, remember to specify `chunk_length_s` and/or `stride_length_s`."),r=e.slice(0,this.config.n_samples)):(r=new Float32Array(this.config.n_samples),r.set(e));const{data:n,dims:a}=this._extract_fbank_features(r);return{input_features:new pe("float32",n,[1,...a])}}}class S3 extends sn{_zero_mean_unit_var_norm(e){const n=e.reduce((i,s)=>i+s,0)/e.length,a=e.reduce((i,s)=>i+(s-n)**2,0)/e.length;return e.map(i=>(i-n)/Math.sqrt(a+1e-7))}async _call(e){Fa(e,"Wav2Vec2FeatureExtractor"),e instanceof Float64Array&&(e=new Float32Array(e));let r=e;this.config.do_normalize&&(r=this._zero_mean_unit_var_norm(r));const n=[1,r.length];return{input_values:new pe("float32",r,n),attention_mask:new pe("int64",new BigInt64Array(r.length).fill(1n),n)}}}class k3 extends sn{constructor(e){super(e);const r=this.config.sampling_rate,n=ba(256,this.config.num_mel_bins,20,Math.floor(r/2),r,null,"kaldi",!0);for(let a=0;a<n.length;++a)n[a].push(0);this.mel_filters=n,this.window=os(400,"povey",{periodic:!1})}_extract_fbank_features(e,r){return e=e.map(n=>n*32768),ss(e,this.window,400,160,{fft_length:512,power:2,center:!1,preemphasis:.97,mel_filters:this.mel_filters,log_mel:"log",mel_floor:1192092955078125e-22,remove_dc_offset:!0,max_num_frames:r,transpose:!0})}async _call(e,{padding:r=!0,pad_to_multiple_of:n=2,do_normalize_per_mel_bins:a=!0,return_attention_mask:i=!0}={}){Fa(e,"SeamlessM4TFeatureExtractor");let{data:s,dims:o}=this._extract_fbank_features(e,this.config.max_length);if(a){const[w,v]=o;for(let k=0;k<v;++k){let $=0;for(let B=0;B<w;++B)$+=s[B*v+k];const C=$/w;let T=0;for(let B=0;B<w;++B)T+=(s[B*v+k]-C)**2;T/=w-1;const A=Math.sqrt(T+1e-7);for(let B=0;B<w;++B){const R=B*v+k;s[R]=(s[R]-C)/A}}}let u;if(r){const[w,v]=o,k=w%n;if(k>0){const $=new Float32Array(v*(w+k));$.set(s),$.fill(this.config.padding_value,s.length);const C=w+k;s=$,o=[C,v],i&&(u=new pe("int64",new BigInt64Array(C),[1,C]),u.data.fill(1n,0,w))}}const[l,p]=o,f=this.config.stride;if(l%f!==0)throw new Error(`The number of frames (${l}) must be a multiple of the stride (${f}).`);const c=new pe("float32",s,o).view(1,Math.floor(l/f),p*f),y={input_features:c};if(i){const w=c.dims[1],v=new BigInt64Array(w);if(u){const k=u.data;for(let $=1,C=0;$<l;$+=f,++C)v[C]=k[$]}else v.fill(1n);y.attention_mask=new pe("int64",v,[1,w])}return y}}class E3 extends sn{constructor(e){super(e);const r=this.config.sampling_rate,n=ba(256,this.config.num_mel_bins,20,Math.floor(r/2),r,null,"kaldi",!0);for(let a=0;a<n.length;++a)n[a].push(0);this.mel_filters=n,this.window=os(400,"hann",{periodic:!1}),this.mean=this.config.mean,this.std=this.config.std}_extract_fbank_features(e,r){return ss(e,this.window,400,160,{fft_length:512,power:2,center:!1,preemphasis:.97,mel_filters:this.mel_filters,log_mel:"log",mel_floor:1192092955078125e-22,remove_dc_offset:!0,max_num_frames:r,transpose:!0})}async _call(e){Fa(e,"ASTFeatureExtractor");const r=this._extract_fbank_features(e,this.config.max_length);if(this.config.do_normalize){const n=this.std*2,a=r.data;for(let i=0;i<a.length;++i)a[i]=(a[i]-this.mean)/n}return{input_values:new pe("float32",r.data,[1,...r.dims])}}}class C3 extends sn{constructor(e){super(e),this.mel_filters=ba(this.config.nb_frequency_bins,this.config.feature_size,this.config.frequency_min,this.config.frequency_max,this.config.sampling_rate,null,"htk"),this.mel_filters_slaney=ba(this.config.nb_frequency_bins,this.config.feature_size,this.config.frequency_min,this.config.frequency_max,this.config.sampling_rate,"slaney","slaney"),this.window=os(this.config.fft_window_size,"hann")}_get_input_mel(e,r,n,a){let i,s=!1;const o=e.length-r;if(o>0)if(n==="rand_trunc"){s=!0;const u=Math.floor(Math.random()*(o+1));e=e.subarray(u,u+r),i=this._extract_fbank_features(e,this.mel_filters_slaney,this.config.nb_max_samples),i.dims=[1,...i.dims]}else throw new Error(`Truncation strategy "${n}" not implemented`);else{if(o<0){let u=new Float64Array(r);if(u.set(e),a==="repeat")for(let l=e.length;l<r;l+=e.length)u.set(e.subarray(0,Math.min(e.length,r-l)),l);else if(a==="repeatpad")for(let l=e.length;l<-o;l+=e.length)u.set(e,l);e=u}if(n==="fusion")throw new Error(`Truncation strategy "${n}" not implemented`);i=this._extract_fbank_features(e,this.mel_filters_slaney,this.config.nb_max_samples),i.dims=[1,...i.dims]}return{...i,longer:s}}_extract_fbank_features(e,r,n=null){return ss(e,this.window,this.config.fft_window_size,this.config.hop_length,{power:2,mel_filters:r,log_mel:"dB",max_num_frames:n,do_pad:!1,transpose:!0})}async _call(e,{max_length:r=null}={}){Fa(e,"ClapFeatureExtractor");const n=this._get_input_mel(e,r??this.config.nb_max_samples,this.config.truncation,this.config.padding);return{input_features:new pe("float32",n.data,[1,...n.dims])}}}class T3 extends sn{}class zn extends $t{constructor(e){super(),this.feature_extractor=e}async _call(e,...r){return await this.feature_extractor(e,...r)}}class I3 extends zn{async _call(...e){return await this.feature_extractor(...e)}post_process_masks(...e){return this.feature_extractor.post_process_masks(...e)}reshape_input_points(...e){return this.feature_extractor.reshape_input_points(...e)}}class A3 extends zn{async _call(e){return await this.feature_extractor(e)}}class M3 extends zn{async _call(e){return await this.feature_extractor(e)}}class O3 extends zn{async _call(e){return await this.feature_extractor(e)}}class z3 extends zn{}class P3{static FEATURE_EXTRACTOR_CLASS_MAPPING={ImageFeatureExtractor:Xe,WhisperFeatureExtractor:x3,ViTFeatureExtractor:d3,MobileViTFeatureExtractor:V0,MobileViTImageProcessor:h3,OwlViTFeatureExtractor:G0,Owlv2ImageProcessor:f3,CLIPFeatureExtractor:U0,CLIPImageProcessor:s3,ChineseCLIPFeatureExtractor:o3,SiglipImageProcessor:u3,ConvNextFeatureExtractor:W0,ConvNextImageProcessor:l3,SegformerFeatureExtractor:r3,BitImageProcessor:a3,DPTImageProcessor:n3,DPTFeatureExtractor:L0,GLPNFeatureExtractor:i3,BeitFeatureExtractor:g3,DeiTFeatureExtractor:m3,DetrFeatureExtractor:y3,YolosFeatureExtractor:w3,DonutFeatureExtractor:H0,NougatImageProcessor:_3,EfficientNetImageProcessor:p3,ViTImageProcessor:c3,VitMatteImageProcessor:$3,SamImageProcessor:b3,Swin2SRImageProcessor:v3,Wav2Vec2FeatureExtractor:S3,SeamlessM4TFeatureExtractor:k3,SpeechT5FeatureExtractor:T3,ASTFeatureExtractor:E3,ClapFeatureExtractor:C3};static PROCESSOR_CLASS_MAPPING={WhisperProcessor:A3,Wav2Vec2ProcessorWithLM:M3,SamProcessor:I3,SpeechT5Processor:O3,OwlViTProcessor:z3};static async from_pretrained(e,{progress_callback:r=null,config:n=null,cache_dir:a=null,local_files_only:i=!1,revision:s="main"}={}){let o=n??await Ar(e,"preprocessor_config.json",!0,{progress_callback:r,config:n,cache_dir:a,local_files_only:i,revision:s}),u=o.feature_extractor_type??o.image_processor_type,l=this.FEATURE_EXTRACTOR_CLASS_MAPPING[u];if(!l)if(o.size!==void 0)console.warn(`Feature extractor type "${u}" not found, assuming ImageFeatureExtractor due to size parameter in config.`),l=Xe;else throw new Error(`Unknown Feature Extractor type: ${u}`);let p=this.PROCESSOR_CLASS_MAPPING[o.processor_class]??zn,f=new l(o);return new p(f)}}Jr.IS_PROCESS_AVAILABLE;wt.backends.onnx.wasm.proxy=!1;const us=document.getElementById("status"),bp=document.getElementById("container"),ua=document.getElementById("video"),Mo=document.getElementById("labels"),zi=document.getElementById("template"),$i=document.getElementById("overlay");us.textContent="Loading model (446MB)...";async function R3(){try{return(await navigator.gpu.requestAdapter()).features.has("shader-f16")}catch{return!1}}const vp=await R3()?"fp16":"fp32",uo="jinaai/jina-clip-v1";let j0,q0,K0,Y0;try{j0=await bv.from_pretrained(uo),q0=await Ng.from_pretrained(uo,{device:"webgpu",dtype:vp}),K0=await P3.from_pretrained("Xenova/clip-vit-base-patch32"),Y0=await Fg.from_pretrained(uo,{device:"webgpu",dtype:vp})}catch(t){throw us.textContent=t.message,alert(t.message),t}Mo.disabled=!1;zi.disabled=!1;us.textContent="Ready";const B3=Math.exp(4.5904),Hr=224,Oo=document.createElement("canvas");Oo.width=Oo.height=Hr;const $p=Oo.getContext("2d",{willReadFrequently:!0});let lo=!1,co,xi,po,xp,Si;function X0(){lo||(lo=!0,async function(){if(po!==Mo.value||xp!==zi.value)if(xi=null,po=Mo.value,xp=zi.value,Si=po.split(/\s*,\s*/).filter(t=>t),Si.length>0){const t=Si.map(n=>zi.value.replaceAll("{}",n)),e=j0(t,{padding:!0,truncation:!0}),{text_embeds:r}=await q0(e);xi=r.normalize().tolist()}else $i.innerHTML="";if(xi){$p.drawImage(ua,0,0,Hr,Hr);const t=$p.getImageData(0,0,Hr,Hr).data,e=new tr(t,Hr,Hr,4),r=await K0(e),{image_embeds:n}=await Y0(r),a=n.normalize().tolist()[0],i=xi.map(o=>w_(o,a)*B3),s=Tn(i).map((o,u)=>[o,u]).sort((o,u)=>u[0]-o[0]);$i.innerHTML="";for(const[o,u]of s)$i.appendChild(document.createTextNode(`${Si[u]}: ${o.toFixed(2)}`)),$i.appendChild(document.createElement("br"))}if(co!==void 0){const t=1e3/(performance.now()-co);us.textContent=`FPS: ${t.toFixed(2)}`}co=performance.now(),lo=!1}()),window.requestAnimationFrame(X0)}navigator.mediaDevices.getUserMedia({video:!0}).then(t=>{ua.srcObject=t,ua.play();const e=t.getVideoTracks()[0],{width:r,height:n}=e.getSettings();ua.width=r,ua.height=n;const a=r/n,[i,s]=a>720/405?[720,720/a]:[405*a,405];bp.style.width=`${i}px`,bp.style.height=`${s}px`,window.requestAnimationFrame(X0)}).catch(t=>{alert(t)});