Upload model distil-whisper_distil-large-v3_turbo_600MB - QLoRA compressed encoder variant

distil-whisper_distil-large-v3_turbo_600MB/AudioEncoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94b7b86bea508311eb4f6f60fa2692bdd4ed566ecd54a80f463f22ece82d8a33
+size 243

distil-whisper_distil-large-v3_turbo_600MB/AudioEncoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed121e756d637301566cc86a4787e2adece71c438774ed1e5652a76d4f9ed62b
+size 348

distil-whisper_distil-large-v3_turbo_600MB/AudioEncoder.mlmodelc/metadata.json

@@ -0,0 +1,70 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Mixed (Float16, Palettized (4 bits))",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1280 × 1 × 1500)",
+        "shortDescription" : "",
+        "shape" : "[1, 1280, 1, 1500]",
+        "name" : "encoder_output_embeds",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.mul" : 2690,
+      "Ios17.transpose" : 32,
+      "Ios17.rsqrt" : 65,
+      "Ios16.constexprLutToDense" : 192,
+      "Ios17.conv" : 578,
+      "Ios17.sub" : 65,
+      "Ios16.einsum" : 5120,
+      "Ios17.concat" : 672,
+      "Ios17.add" : 322,
+      "Ios16.reduceMean" : 130,
+      "Ios17.batchNorm" : 65,
+      "Ios17.sliceByIndex" : 4480,
+      "Ios16.softmax" : 2560,
+      "Ios16.gelu" : 34
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "isUpdatable" : "0",
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.2.0",
+      "com.github.apple.coremltools.version" : "7.1"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 128 × 1 × 3000)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 1, 3000]",
+        "name" : "melspectrogram_features",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "AudioEncoder_mixedBitPalettized_4_0_bit",
+    "method" : "predict"
+  }
+]

distil-whisper_distil-large-v3_turbo_600MB/AudioEncoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98410d88d5243d74e5e5210febfbb84dd7a9392793f0fb320b46ab8b92c52a15
+size 353908416

distil-whisper_distil-large-v3_turbo_600MB/MelSpectrogram.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:091a361134891f94e613562771beea0d93a9aefbc6984ba86c60f856e07a508f
+size 243

distil-whisper_distil-large-v3_turbo_600MB/MelSpectrogram.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3c5778c86d6fbc6a9817a56dbcac05a946a4d95c77f6db8355572f3be9e9a68
+size 329

distil-whisper_distil-large-v3_turbo_600MB/MelSpectrogram.mlmodelc/metadata.json

@@ -0,0 +1,71 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 128 × 1 × 3000)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 1, 3000]",
+        "name" : "melspectrogram_features",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 7,
+    "mlProgramOperationTypeHistogram" : {
+      "Pad" : 1,
+      "Ios16.mul" : 2,
+      "SliceByIndex" : 1,
+      "Ios16.sub" : 1,
+      "Ios16.log" : 1,
+      "Ios16.conv" : 2,
+      "Ios16.add" : 3,
+      "Ios16.square" : 2,
+      "Ios16.matmul" : 1,
+      "Squeeze" : 2,
+      "Ios16.maximum" : 1,
+      "ExpandDims" : 4,
+      "Ios16.reduceMax" : 1,
+      "Identity" : 1,
+      "Ios16.reshape" : 2
+    },
+    "computePrecision" : "Mixed (Float16, Int32)",
+    "isUpdatable" : "0",
+    "availability" : {
+      "macOS" : "13.0",
+      "tvOS" : "16.0",
+      "visionOS" : "1.0",
+      "watchOS" : "9.0",
+      "iOS" : "16.0",
+      "macCatalyst" : "16.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.2.1",
+      "com.github.apple.coremltools.version" : "7.1"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 480000)",
+        "shortDescription" : "",
+        "shape" : "[480000]",
+        "name" : "audio",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "MelSpectrogram",
+    "method" : "predict"
+  }
+]

distil-whisper_distil-large-v3_turbo_600MB/MelSpectrogram.mlmodelc/model.mil

@@ -0,0 +1,66 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "2.2.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "7.1"}})]
+{
+    func main<ios16>(tensor<fp16, [480000]> audio) {
+            tensor<int32, [3]> var_10 = const()[name = tensor<string, []>("op_10"), val = tensor<int32, [3]>([1, 1, 480000])];
+            tensor<fp16, [1, 1, 480000]> input_1_cast_fp16 = reshape(shape = var_10, x = audio)[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<int32, [6]> input_3_pad_0 = const()[name = tensor<string, []>("input_3_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 200, 200])];
+            tensor<string, []> input_3_mode_0 = const()[name = tensor<string, []>("input_3_mode_0"), val = tensor<string, []>("reflect")];
+            tensor<fp16, []> input_3_constant_val_0_to_fp16 = const()[name = tensor<string, []>("input_3_constant_val_0_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 1, 480400]> input_3_cast_fp16 = pad(constant_val = input_3_constant_val_0_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<int32, [1]> var_22 = const()[name = tensor<string, []>("op_22"), val = tensor<int32, [1]>([480400])];
+            tensor<fp16, [480400]> input_cast_fp16 = reshape(shape = var_22, x = input_3_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = tensor<string, []>("expand_dims_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 480400]> expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = input_cast_fp16)[name = tensor<string, []>("expand_dims_0_cast_fp16")];
+            tensor<int32, [1]> expand_dims_3 = const()[name = tensor<string, []>("expand_dims_3"), val = tensor<int32, [1]>([160])];
+            tensor<int32, [1]> expand_dims_4_axes_0 = const()[name = tensor<string, []>("expand_dims_4_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 480400]> expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = expand_dims_0_cast_fp16)[name = tensor<string, []>("expand_dims_4_cast_fp16")];
+            tensor<string, []> conv_0_pad_type_0 = const()[name = tensor<string, []>("conv_0_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> conv_0_pad_0 = const()[name = tensor<string, []>("conv_0_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_0_dilations_0 = const()[name = tensor<string, []>("conv_0_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> conv_0_groups_0 = const()[name = tensor<string, []>("conv_0_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [201, 1, 400]> expand_dims_1_to_fp16 = const()[name = tensor<string, []>("expand_dims_1_to_fp16"), val = tensor<fp16, [201, 1, 400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [1, 201, 3001]> conv_0_cast_fp16 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1_to_fp16, x = expand_dims_4_cast_fp16)[name = tensor<string, []>("conv_0_cast_fp16")];
+            tensor<string, []> conv_1_pad_type_0 = const()[name = tensor<string, []>("conv_1_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> conv_1_pad_0 = const()[name = tensor<string, []>("conv_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_1_dilations_0 = const()[name = tensor<string, []>("conv_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> conv_1_groups_0 = const()[name = tensor<string, []>("conv_1_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [201, 1, 400]> expand_dims_2_to_fp16 = const()[name = tensor<string, []>("expand_dims_2_to_fp16"), val = tensor<fp16, [201, 1, 400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160960)))];
+            tensor<fp16, [1, 201, 3001]> conv_1_cast_fp16 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2_to_fp16, x = expand_dims_4_cast_fp16)[name = tensor<string, []>("conv_1_cast_fp16")];
+            tensor<int32, [1]> squeeze_0_axes_0 = const()[name = tensor<string, []>("squeeze_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [201, 3001]> squeeze_0_cast_fp16 = squeeze(axes = squeeze_0_axes_0, x = conv_0_cast_fp16)[name = tensor<string, []>("squeeze_0_cast_fp16")];
+            tensor<int32, [1]> squeeze_1_axes_0 = const()[name = tensor<string, []>("squeeze_1_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [201, 3001]> squeeze_1_cast_fp16 = squeeze(axes = squeeze_1_axes_0, x = conv_1_cast_fp16)[name = tensor<string, []>("squeeze_1_cast_fp16")];
+            tensor<fp16, [201, 3001]> square_0_cast_fp16 = square(x = squeeze_0_cast_fp16)[name = tensor<string, []>("square_0_cast_fp16")];
+            tensor<fp16, [201, 3001]> square_1_cast_fp16 = square(x = squeeze_1_cast_fp16)[name = tensor<string, []>("square_1_cast_fp16")];
+            tensor<fp16, [201, 3001]> add_1_cast_fp16 = add(x = square_0_cast_fp16, y = square_1_cast_fp16)[name = tensor<string, []>("add_1_cast_fp16")];
+            tensor<fp16, [201, 3001]> magnitudes_1_cast_fp16 = identity(x = add_1_cast_fp16)[name = tensor<string, []>("magnitudes_1_cast_fp16")];
+            tensor<int32, [2]> magnitudes_begin_0 = const()[name = tensor<string, []>("magnitudes_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> magnitudes_end_0 = const()[name = tensor<string, []>("magnitudes_end_0"), val = tensor<int32, [2]>([201, 3000])];
+            tensor<bool, [2]> magnitudes_end_mask_0 = const()[name = tensor<string, []>("magnitudes_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp16, [201, 3000]> magnitudes_cast_fp16 = slice_by_index(begin = magnitudes_begin_0, end = magnitudes_end_0, end_mask = magnitudes_end_mask_0, x = magnitudes_1_cast_fp16)[name = tensor<string, []>("magnitudes_cast_fp16")];
+            tensor<bool, []> mel_spec_1_transpose_x_0 = const()[name = tensor<string, []>("mel_spec_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> mel_spec_1_transpose_y_0 = const()[name = tensor<string, []>("mel_spec_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [128, 201]> mel_filters_to_fp16 = const()[name = tensor<string, []>("mel_filters_to_fp16"), val = tensor<fp16, [128, 201]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321856)))];
+            tensor<fp16, [128, 3000]> mel_spec_1_cast_fp16 = matmul(transpose_x = mel_spec_1_transpose_x_0, transpose_y = mel_spec_1_transpose_y_0, x = mel_filters_to_fp16, y = magnitudes_cast_fp16)[name = tensor<string, []>("mel_spec_1_cast_fp16")];
+            tensor<fp16, []> var_41_to_fp16 = const()[name = tensor<string, []>("op_41_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [128, 3000]> mel_spec_cast_fp16 = add(x = mel_spec_1_cast_fp16, y = var_41_to_fp16)[name = tensor<string, []>("mel_spec_cast_fp16")];
+            tensor<fp16, []> log_0_epsilon_0_to_fp16 = const()[name = tensor<string, []>("log_0_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [128, 3000]> log_0_cast_fp16 = log(epsilon = log_0_epsilon_0_to_fp16, x = mel_spec_cast_fp16)[name = tensor<string, []>("log_0_cast_fp16")];
+            tensor<fp16, []> mul_0_y_0_to_fp16 = const()[name = tensor<string, []>("mul_0_y_0_to_fp16"), val = tensor<fp16, []>(0x1.bccp-2)];
+            tensor<fp16, [128, 3000]> mul_0_cast_fp16 = mul(x = log_0_cast_fp16, y = mul_0_y_0_to_fp16)[name = tensor<string, []>("mul_0_cast_fp16")];
+            tensor<bool, []> var_44_keep_dims_0 = const()[name = tensor<string, []>("op_44_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, []> var_44_cast_fp16 = reduce_max(keep_dims = var_44_keep_dims_0, x = mul_0_cast_fp16)[name = tensor<string, []>("op_44_cast_fp16")];
+            tensor<fp16, []> var_46_to_fp16 = const()[name = tensor<string, []>("op_46_to_fp16"), val = tensor<fp16, []>(0x1p+3)];
+            tensor<fp16, []> var_47_cast_fp16 = sub(x = var_44_cast_fp16, y = var_46_to_fp16)[name = tensor<string, []>("op_47_cast_fp16")];
+            tensor<fp16, [128, 3000]> log_spec_3_cast_fp16 = maximum(x = mul_0_cast_fp16, y = var_47_cast_fp16)[name = tensor<string, []>("log_spec_3_cast_fp16")];
+            tensor<fp16, []> var_50_to_fp16 = const()[name = tensor<string, []>("op_50_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
+            tensor<fp16, [128, 3000]> var_51_cast_fp16 = add(x = log_spec_3_cast_fp16, y = var_50_to_fp16)[name = tensor<string, []>("op_51_cast_fp16")];
+            tensor<fp16, []> _inversed_log_spec_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_log_spec_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-2)];
+            tensor<fp16, [128, 3000]> _inversed_log_spec_cast_fp16 = mul(x = var_51_cast_fp16, y = _inversed_log_spec_y_0_to_fp16)[name = tensor<string, []>("_inversed_log_spec_cast_fp16")];
+            tensor<int32, [1]> var_55_axes_0 = const()[name = tensor<string, []>("op_55_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 128, 3000]> var_55_cast_fp16 = expand_dims(axes = var_55_axes_0, x = _inversed_log_spec_cast_fp16)[name = tensor<string, []>("op_55_cast_fp16")];
+            tensor<int32, [1]> var_62_axes_0 = const()[name = tensor<string, []>("op_62_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 128, 1, 3000]> melspectrogram_features = expand_dims(axes = var_62_axes_0, x = var_55_cast_fp16)[name = tensor<string, []>("op_62_cast_fp16")];
+        } -> (melspectrogram_features);
+}

distil-whisper_distil-large-v3_turbo_600MB/MelSpectrogram.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a80bf139bea628069316903af7fb674a8d319233ab98f9d74b1c840915a4ef0
+size 373376

distil-whisper_distil-large-v3_turbo_600MB/TextDecoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:394520c761dbfb69b05c7c2d49380bbece53e92f6bd19ffabcd46f6aaa2193ad
+size 243

distil-whisper_distil-large-v3_turbo_600MB/TextDecoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9b45ba978c8fcc372a7b52150f83b3ddbe3376ec044e3439e9e227d0076950f
+size 593

distil-whisper_distil-large-v3_turbo_600MB/TextDecoder.mlmodelc/metadata.json

@@ -0,0 +1,154 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1 × 51866)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 51866]",
+        "name" : "logits",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 2560 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 2560, 1, 1]",
+        "name" : "key_cache_updates",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 2560 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 2560, 1, 1]",
+        "name" : "value_cache_updates",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 7,
+    "mlProgramOperationTypeHistogram" : {
+      "Split" : 2,
+      "Concat" : 2,
+      "Ios16.rsqrt" : 7,
+      "Ios16.mul" : 26,
+      "Squeeze" : 1,
+      "Ios16.sub" : 8,
+      "Transpose" : 1,
+      "Ios16.conv" : 20,
+      "Ios16.add" : 20,
+      "Ios16.linear" : 1,
+      "Ios16.matmul" : 8,
+      "Ios16.gelu" : 2,
+      "Ios16.reduceMean" : 14,
+      "ExpandDims" : 6,
+      "Ios16.batchNorm" : 7,
+      "Ios16.gather" : 2,
+      "Ios16.reshape" : 16,
+      "Ios16.softmax" : 4
+    },
+    "computePrecision" : "Mixed (Float16, Int32)",
+    "isUpdatable" : "0",
+    "availability" : {
+      "macOS" : "13.0",
+      "tvOS" : "16.0",
+      "visionOS" : "1.0",
+      "watchOS" : "9.0",
+      "iOS" : "16.0",
+      "macCatalyst" : "16.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.2.1",
+      "com.github.apple.coremltools.version" : "7.1"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "input_ids",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "cache_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 2560 × 1 × 224)",
+        "shortDescription" : "",
+        "shape" : "[1, 2560, 1, 224]",
+        "name" : "key_cache",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 2560 × 1 × 224)",
+        "shortDescription" : "",
+        "shape" : "[1, 2560, 1, 224]",
+        "name" : "value_cache",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 224)",
+        "shortDescription" : "",
+        "shape" : "[1, 224]",
+        "name" : "kv_cache_update_mask",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1280 × 1 × 1500)",
+        "shortDescription" : "",
+        "shape" : "[1, 1280, 1, 1500]",
+        "name" : "encoder_output_embeds",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 224)",
+        "shortDescription" : "",
+        "shape" : "[1, 224]",
+        "name" : "decoder_key_padding_mask",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "TextDecoder",
+    "method" : "predict"
+  }
+]

distil-whisper_distil-large-v3_turbo_600MB/TextDecoder.mlmodelc/model.mil

@@ -0,0 +1,389 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "2.2.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "7.1"}})]
+{
+    func main<ios16>(tensor<int32, [1]> cache_length, tensor<fp16, [1, 224]> decoder_key_padding_mask, tensor<fp16, [1, 1280, 1, 1500]> encoder_output_embeds, tensor<int32, [1]> input_ids, tensor<fp16, [1, 2560, 1, 224]> key_cache, tensor<fp16, [1, 224]> kv_cache_update_mask, tensor<fp16, [1, 2560, 1, 224]> value_cache) {
+            tensor<int32, []> var_20_axis_0 = const()[name = tensor<string, []>("op_20_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> var_20_batch_dims_0 = const()[name = tensor<string, []>("op_20_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [51866, 1280]> embed_tokens_weight_to_fp16 = const()[name = tensor<string, []>("embed_tokens_weight_to_fp16"), val = tensor<fp16, [51866, 1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [1, 1280]> var_20_cast_fp16 = gather(axis = var_20_axis_0, batch_dims = var_20_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor<string, []>("op_20_cast_fp16")];
+            tensor<int32, []> var_24_axis_0 = const()[name = tensor<string, []>("op_24_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> var_24_batch_dims_0 = const()[name = tensor<string, []>("op_24_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [448, 1280]> embed_positions_weight_to_fp16 = const()[name = tensor<string, []>("embed_positions_weight_to_fp16"), val = tensor<fp16, [448, 1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132777088)))];
+            tensor<fp16, [1, 1280]> var_24_cast_fp16 = gather(axis = var_24_axis_0, batch_dims = var_24_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor<string, []>("op_24_cast_fp16")];
+            tensor<fp16, [1, 1280]> hidden_states_1_cast_fp16 = add(x = var_20_cast_fp16, y = var_24_cast_fp16)[name = tensor<string, []>("hidden_states_1_cast_fp16")];
+            tensor<int32, [1]> var_38_axes_0 = const()[name = tensor<string, []>("op_38_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1280, 1]> var_38_cast_fp16 = expand_dims(axes = var_38_axes_0, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("op_38_cast_fp16")];
+            tensor<int32, [1]> inputs_1_axes_0 = const()[name = tensor<string, []>("inputs_1_axes_0"), val = tensor<int32, [1]>([3])];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_38_cast_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")];
+            tensor<int32, [2]> tile_0 = const()[name = tensor<string, []>("tile_0"), val = tensor<int32, [2]>([1280, 1280])];
+            tensor<int32, []> var_43_axis_0 = const()[name = tensor<string, []>("op_43_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1280, 1, 224]> var_43_cast_fp16_0, tensor<fp16, [1, 1280, 1, 224]> var_43_cast_fp16_1 = split(axis = var_43_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor<string, []>("op_43_cast_fp16")];
+            tensor<int32, [2]> tile_1 = const()[name = tensor<string, []>("tile_1"), val = tensor<int32, [2]>([1280, 1280])];
+            tensor<int32, []> var_48_axis_0 = const()[name = tensor<string, []>("op_48_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 1280, 1, 224]> var_48_cast_fp16_0, tensor<fp16, [1, 1280, 1, 224]> var_48_cast_fp16_1 = split(axis = var_48_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor<string, []>("op_48_cast_fp16")];
+            tensor<int32, []> var_56 = const()[name = tensor<string, []>("op_56"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_63 = const()[name = tensor<string, []>("op_63"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_64 = const()[name = tensor<string, []>("op_64"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_76 = const()[name = tensor<string, []>("op_76"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> channels_mean_1_cast_fp16 = reduce_mean(axes = var_76, keep_dims = var_64, x = inputs_1_cast_fp16)[name = tensor<string, []>("channels_mean_1_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> zero_mean_1_cast_fp16 = sub(x = inputs_1_cast_fp16, y = channels_mean_1_cast_fp16)[name = tensor<string, []>("zero_mean_1_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = zero_mean_1_cast_fp16)[name = tensor<string, []>("zero_mean_sq_1_cast_fp16")];
+            tensor<int32, [1]> var_80 = const()[name = tensor<string, []>("op_80"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> var_81_cast_fp16 = reduce_mean(axes = var_80, keep_dims = var_64, x = zero_mean_sq_1_cast_fp16)[name = tensor<string, []>("op_81_cast_fp16")];
+            tensor<fp16, []> var_82_to_fp16 = const()[name = tensor<string, []>("op_82_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1]> var_83_cast_fp16 = add(x = var_81_cast_fp16, y = var_82_to_fp16)[name = tensor<string, []>("op_83_cast_fp16")];
+            tensor<fp16, []> denom_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [1, 1, 1, 1]> denom_1_cast_fp16 = rsqrt(epsilon = denom_1_epsilon_0_to_fp16, x = var_83_cast_fp16)[name = tensor<string, []>("denom_1_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> out_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = denom_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
+            tensor<fp16, [1280]> obj_1_mean_0_to_fp16 = const()[name = tensor<string, []>("obj_1_mean_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133924032)))];
+            tensor<fp16, [1280]> obj_1_variance_0_to_fp16 = const()[name = tensor<string, []>("obj_1_variance_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133926656)))];
+            tensor<fp16, [1280]> obj_1_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_1_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133929280)))];
+            tensor<fp16, [1280]> obj_1_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_1_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133931904)))];
+            tensor<fp16, []> obj_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor<string, []>("obj_1_cast_fp16")];
+            tensor<int32, [2]> var_98 = const()[name = tensor<string, []>("op_98"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_100 = const()[name = tensor<string, []>("op_100"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_1_pad_type_0 = const()[name = tensor<string, []>("query_1_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_1_pad_0 = const()[name = tensor<string, []>("query_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133934528)))];
+            tensor<fp16, [1280]> layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137211392)))];
+            tensor<fp16, [1, 1280, 1, 1]> query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = var_100, groups = var_63, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = var_98, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")];
+            tensor<int32, [2]> var_104 = const()[name = tensor<string, []>("op_104"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_106 = const()[name = tensor<string, []>("op_106"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> current_key_1_pad_type_0 = const()[name = tensor<string, []>("current_key_1_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> current_key_1_pad_0 = const()[name = tensor<string, []>("current_key_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137214016)))];
+            tensor<fp16, [1, 1280, 1, 1]> current_key_1_cast_fp16 = conv(dilations = var_106, groups = var_63, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = var_104, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("current_key_1_cast_fp16")];
+            tensor<int32, [2]> var_111 = const()[name = tensor<string, []>("op_111"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_113 = const()[name = tensor<string, []>("op_113"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> current_value_1_pad_type_0 = const()[name = tensor<string, []>("current_value_1_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> current_value_1_pad_0 = const()[name = tensor<string, []>("current_value_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140490880)))];
+            tensor<fp16, [1280]> layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143767744)))];
+            tensor<fp16, [1, 1280, 1, 1]> current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = var_113, groups = var_63, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = var_111, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("current_value_1_cast_fp16")];
+            tensor<int32, [1]> var_117_axes_0 = const()[name = tensor<string, []>("op_117_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 224]> var_117_cast_fp16 = expand_dims(axes = var_117_axes_0, x = kv_cache_update_mask)[name = tensor<string, []>("op_117_cast_fp16")];
+            tensor<int32, [1]> var_118_axes_0 = const()[name = tensor<string, []>("op_118_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 224]> var_118_cast_fp16 = expand_dims(axes = var_118_axes_0, x = var_117_cast_fp16)[name = tensor<string, []>("op_118_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 224]> var_120_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_118_cast_fp16)[name = tensor<string, []>("op_120_cast_fp16")];
+            tensor<fp16, []> var_57_to_fp16 = const()[name = tensor<string, []>("op_57_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1, 1, 1, 224]> var_121_cast_fp16 = sub(x = var_57_to_fp16, y = var_118_cast_fp16)[name = tensor<string, []>("op_121_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 224]> var_122_cast_fp16 = mul(x = var_43_cast_fp16_0, y = var_121_cast_fp16)[name = tensor<string, []>("op_122_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 224]> key_1_cast_fp16 = add(x = var_120_cast_fp16, y = var_122_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 224]> var_124_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_118_cast_fp16)[name = tensor<string, []>("op_124_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 224]> var_126_cast_fp16 = mul(x = var_48_cast_fp16_0, y = var_121_cast_fp16)[name = tensor<string, []>("op_126_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 224]> value_1_cast_fp16 = add(x = var_124_cast_fp16, y = var_126_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")];
+            tensor<int32, [4]> var_129 = const()[name = tensor<string, []>("op_129"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1]> var_130_cast_fp16 = reshape(shape = var_129, x = query_1_cast_fp16)[name = tensor<string, []>("op_130_cast_fp16")];
+            tensor<fp16, []> var_131_to_fp16 = const()[name = tensor<string, []>("op_131_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1]> var_132_cast_fp16 = mul(x = var_130_cast_fp16, y = var_131_to_fp16)[name = tensor<string, []>("op_132_cast_fp16")];
+            tensor<int32, [4]> var_133 = const()[name = tensor<string, []>("op_133"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 224]> var_134_cast_fp16 = reshape(shape = var_133, x = key_1_cast_fp16)[name = tensor<string, []>("op_134_cast_fp16")];
+            tensor<bool, []> mh_w_1_transpose_x_0 = const()[name = tensor<string, []>("mh_w_1_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_1_transpose_y_0 = const()[name = tensor<string, []>("mh_w_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1, 224]> mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_132_cast_fp16, y = var_134_cast_fp16)[name = tensor<string, []>("mh_w_1_cast_fp16")];
+            tensor<int32, [1]> var_138_axes_0 = const()[name = tensor<string, []>("op_138_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 224]> var_138_cast_fp16 = expand_dims(axes = var_138_axes_0, x = decoder_key_padding_mask)[name = tensor<string, []>("op_138_cast_fp16")];
+            tensor<int32, [1]> var_139_axes_0 = const()[name = tensor<string, []>("op_139_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 224]> var_139_cast_fp16 = expand_dims(axes = var_139_axes_0, x = var_138_cast_fp16)[name = tensor<string, []>("op_139_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 224]> mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_139_cast_fp16)[name = tensor<string, []>("mh_w_3_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 224]> var_142_cast_fp16 = softmax(axis = var_56, x = mh_w_3_cast_fp16)[name = tensor<string, []>("op_142_cast_fp16")];
+            tensor<int32, [4]> var_143 = const()[name = tensor<string, []>("op_143"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 224]> var_144_cast_fp16 = reshape(shape = var_143, x = value_1_cast_fp16)[name = tensor<string, []>("op_144_cast_fp16")];
+            tensor<bool, []> attn_1_transpose_x_0 = const()[name = tensor<string, []>("attn_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_1_transpose_y_0 = const()[name = tensor<string, []>("attn_1_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_144_cast_fp16, y = var_142_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")];
+            tensor<int32, [4]> var_147 = const()[name = tensor<string, []>("op_147"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1]> input_1_cast_fp16 = reshape(shape = var_147, x = attn_1_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<int32, [2]> var_151 = const()[name = tensor<string, []>("op_151"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_153 = const()[name = tensor<string, []>("op_153"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_7_pad_type_0 = const()[name = tensor<string, []>("obj_7_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_7_pad_0 = const()[name = tensor<string, []>("obj_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(143770368)))];
+            tensor<fp16, [1280]> layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147047232)))];
+            tensor<fp16, [1, 1280, 1, 1]> obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = var_153, groups = var_63, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = var_151, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")];
+            tensor<int32, [1]> var_163 = const()[name = tensor<string, []>("op_163"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> channels_mean_3_cast_fp16 = reduce_mean(axes = var_163, keep_dims = var_64, x = inputs_3_cast_fp16)[name = tensor<string, []>("channels_mean_3_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> zero_mean_3_cast_fp16 = sub(x = inputs_3_cast_fp16, y = channels_mean_3_cast_fp16)[name = tensor<string, []>("zero_mean_3_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = zero_mean_3_cast_fp16)[name = tensor<string, []>("zero_mean_sq_3_cast_fp16")];
+            tensor<int32, [1]> var_167 = const()[name = tensor<string, []>("op_167"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> var_168_cast_fp16 = reduce_mean(axes = var_167, keep_dims = var_64, x = zero_mean_sq_3_cast_fp16)[name = tensor<string, []>("op_168_cast_fp16")];
+            tensor<fp16, []> var_169_to_fp16 = const()[name = tensor<string, []>("op_169_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1]> var_170_cast_fp16 = add(x = var_168_cast_fp16, y = var_169_to_fp16)[name = tensor<string, []>("op_170_cast_fp16")];
+            tensor<fp16, []> denom_3_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_3_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [1, 1, 1, 1]> denom_3_cast_fp16 = rsqrt(epsilon = denom_3_epsilon_0_to_fp16, x = var_170_cast_fp16)[name = tensor<string, []>("denom_3_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> out_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = denom_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
+            tensor<fp16, [1280]> obj_9_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_9_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147049856)))];
+            tensor<fp16, [1280]> obj_9_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_9_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147052480)))];
+            tensor<fp16, []> obj_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor<string, []>("obj_9_cast_fp16")];
+            tensor<int32, [2]> var_185 = const()[name = tensor<string, []>("op_185"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_187 = const()[name = tensor<string, []>("op_187"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_3_pad_type_0 = const()[name = tensor<string, []>("query_3_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_3_pad_0 = const()[name = tensor<string, []>("query_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147055104)))];
+            tensor<fp16, [1280]> layers_0_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150331968)))];
+            tensor<fp16, [1, 1280, 1, 1]> query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = var_187, groups = var_63, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = var_185, weight = layers_0_encoder_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor<string, []>("query_3_cast_fp16")];
+            tensor<int32, [2]> var_191 = const()[name = tensor<string, []>("op_191"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_193 = const()[name = tensor<string, []>("op_193"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_3_pad_type_0 = const()[name = tensor<string, []>("key_3_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_3_pad_0 = const()[name = tensor<string, []>("key_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150334592)))];
+            tensor<fp16, [1, 1280, 1, 1500]> key_3_cast_fp16 = conv(dilations = var_193, groups = var_63, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = var_191, weight = layers_0_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_3_cast_fp16")];
+            tensor<int32, [2]> var_198 = const()[name = tensor<string, []>("op_198"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_200 = const()[name = tensor<string, []>("op_200"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_3_pad_type_0 = const()[name = tensor<string, []>("value_3_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_3_pad_0 = const()[name = tensor<string, []>("value_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153611456)))];
+            tensor<fp16, [1280]> layers_0_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156888320)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_3_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_bias_to_fp16, dilations = var_200, groups = var_63, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = var_198, weight = layers_0_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_3_cast_fp16")];
+            tensor<int32, [4]> var_204 = const()[name = tensor<string, []>("op_204"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1]> var_205_cast_fp16 = reshape(shape = var_204, x = query_3_cast_fp16)[name = tensor<string, []>("op_205_cast_fp16")];
+            tensor<fp16, []> var_206_to_fp16 = const()[name = tensor<string, []>("op_206_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1]> var_207_cast_fp16 = mul(x = var_205_cast_fp16, y = var_206_to_fp16)[name = tensor<string, []>("op_207_cast_fp16")];
+            tensor<int32, [4]> var_208 = const()[name = tensor<string, []>("op_208"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_209_cast_fp16 = reshape(shape = var_208, x = key_3_cast_fp16)[name = tensor<string, []>("op_209_cast_fp16")];
+            tensor<bool, []> mh_w_5_transpose_x_0 = const()[name = tensor<string, []>("mh_w_5_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_5_transpose_y_0 = const()[name = tensor<string, []>("mh_w_5_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1, 1500]> mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_207_cast_fp16, y = var_209_cast_fp16)[name = tensor<string, []>("mh_w_5_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 1500]> var_212_cast_fp16 = softmax(axis = var_56, x = mh_w_5_cast_fp16)[name = tensor<string, []>("op_212_cast_fp16")];
+            tensor<int32, [4]> var_213 = const()[name = tensor<string, []>("op_213"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_214_cast_fp16 = reshape(shape = var_213, x = value_3_cast_fp16)[name = tensor<string, []>("op_214_cast_fp16")];
+            tensor<bool, []> attn_3_transpose_x_0 = const()[name = tensor<string, []>("attn_3_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_3_transpose_y_0 = const()[name = tensor<string, []>("attn_3_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_214_cast_fp16, y = var_212_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")];
+            tensor<int32, [4]> var_217 = const()[name = tensor<string, []>("op_217"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1]> input_3_cast_fp16 = reshape(shape = var_217, x = attn_3_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<int32, [2]> var_221 = const()[name = tensor<string, []>("op_221"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_223 = const()[name = tensor<string, []>("op_223"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_11_pad_type_0 = const()[name = tensor<string, []>("obj_11_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_11_pad_0 = const()[name = tensor<string, []>("obj_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156890944)))];
+            tensor<fp16, [1280]> layers_0_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160167808)))];
+            tensor<fp16, [1, 1280, 1, 1]> obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = var_223, groups = var_63, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = var_221, weight = layers_0_encoder_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor<string, []>("inputs_5_cast_fp16")];
+            tensor<int32, [1]> var_229 = const()[name = tensor<string, []>("op_229"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> channels_mean_5_cast_fp16 = reduce_mean(axes = var_229, keep_dims = var_64, x = inputs_5_cast_fp16)[name = tensor<string, []>("channels_mean_5_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> zero_mean_5_cast_fp16 = sub(x = inputs_5_cast_fp16, y = channels_mean_5_cast_fp16)[name = tensor<string, []>("zero_mean_5_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = zero_mean_5_cast_fp16)[name = tensor<string, []>("zero_mean_sq_5_cast_fp16")];
+            tensor<int32, [1]> var_233 = const()[name = tensor<string, []>("op_233"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> var_234_cast_fp16 = reduce_mean(axes = var_233, keep_dims = var_64, x = zero_mean_sq_5_cast_fp16)[name = tensor<string, []>("op_234_cast_fp16")];
+            tensor<fp16, []> var_235_to_fp16 = const()[name = tensor<string, []>("op_235_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1]> var_236_cast_fp16 = add(x = var_234_cast_fp16, y = var_235_to_fp16)[name = tensor<string, []>("op_236_cast_fp16")];
+            tensor<fp16, []> denom_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [1, 1, 1, 1]> denom_5_cast_fp16 = rsqrt(epsilon = denom_5_epsilon_0_to_fp16, x = var_236_cast_fp16)[name = tensor<string, []>("denom_5_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> out_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = denom_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
+            tensor<fp16, [1280]> input_5_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_5_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160170432)))];
+            tensor<fp16, [1280]> input_5_beta_0_to_fp16 = const()[name = tensor<string, []>("input_5_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160173056)))];
+            tensor<fp16, []> input_5_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_5_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> input_5_cast_fp16 = batch_norm(beta = input_5_beta_0_to_fp16, epsilon = input_5_epsilon_0_to_fp16, gamma = input_5_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<int32, [2]> var_247 = const()[name = tensor<string, []>("op_247"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_249 = const()[name = tensor<string, []>("op_249"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_7_pad_type_0 = const()[name = tensor<string, []>("input_7_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_7_pad_0 = const()[name = tensor<string, []>("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_0_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160175680)))];
+            tensor<fp16, [5120]> layers_0_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173282944)))];
+            tensor<fp16, [1, 5120, 1, 1]> input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = var_249, groups = var_63, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = var_247, weight = layers_0_fc1_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<string, []> input_9_mode_0 = const()[name = tensor<string, []>("input_9_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1]> input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
+            tensor<int32, [2]> var_255 = const()[name = tensor<string, []>("op_255"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_257 = const()[name = tensor<string, []>("op_257"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_3_pad_type_0 = const()[name = tensor<string, []>("hidden_states_3_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = tensor<string, []>("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_0_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173293248)))];
+            tensor<fp16, [1280]> layers_0_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186400512)))];
+            tensor<fp16, [1, 1280, 1, 1]> hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = var_257, groups = var_63, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = var_255, weight = layers_0_fc2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("hidden_states_3_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")];
+            tensor<int32, []> var_270 = const()[name = tensor<string, []>("op_270"), val = tensor<int32, []>(3)];
+            tensor<int32, []> var_277 = const()[name = tensor<string, []>("op_277"), val = tensor<int32, []>(1)];
+            tensor<bool, []> var_278 = const()[name = tensor<string, []>("op_278"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_290 = const()[name = tensor<string, []>("op_290"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> channels_mean_7_cast_fp16 = reduce_mean(axes = var_290, keep_dims = var_278, x = inputs_7_cast_fp16)[name = tensor<string, []>("channels_mean_7_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> zero_mean_7_cast_fp16 = sub(x = inputs_7_cast_fp16, y = channels_mean_7_cast_fp16)[name = tensor<string, []>("zero_mean_7_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = zero_mean_7_cast_fp16)[name = tensor<string, []>("zero_mean_sq_7_cast_fp16")];
+            tensor<int32, [1]> var_294 = const()[name = tensor<string, []>("op_294"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> var_295_cast_fp16 = reduce_mean(axes = var_294, keep_dims = var_278, x = zero_mean_sq_7_cast_fp16)[name = tensor<string, []>("op_295_cast_fp16")];
+            tensor<fp16, []> var_296_to_fp16 = const()[name = tensor<string, []>("op_296_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1]> var_297_cast_fp16 = add(x = var_295_cast_fp16, y = var_296_to_fp16)[name = tensor<string, []>("op_297_cast_fp16")];
+            tensor<fp16, []> denom_7_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_7_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [1, 1, 1, 1]> denom_7_cast_fp16 = rsqrt(epsilon = denom_7_epsilon_0_to_fp16, x = var_297_cast_fp16)[name = tensor<string, []>("denom_7_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> out_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = denom_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
+            tensor<fp16, [1280]> obj_13_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_13_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186403136)))];
+            tensor<fp16, [1280]> obj_13_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_13_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186405760)))];
+            tensor<fp16, []> obj_13_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_13_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> obj_13_cast_fp16 = batch_norm(beta = obj_13_beta_0_to_fp16, epsilon = obj_13_epsilon_0_to_fp16, gamma = obj_13_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor<string, []>("obj_13_cast_fp16")];
+            tensor<int32, [2]> var_312 = const()[name = tensor<string, []>("op_312"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_314 = const()[name = tensor<string, []>("op_314"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_5_pad_type_0 = const()[name = tensor<string, []>("query_5_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_5_pad_0 = const()[name = tensor<string, []>("query_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186408384)))];
+            tensor<fp16, [1280]> layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189685248)))];
+            tensor<fp16, [1, 1280, 1, 1]> query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = var_314, groups = var_277, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = var_312, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")];
+            tensor<int32, [2]> var_318 = const()[name = tensor<string, []>("op_318"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_320 = const()[name = tensor<string, []>("op_320"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> current_key_pad_type_0 = const()[name = tensor<string, []>("current_key_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> current_key_pad_0 = const()[name = tensor<string, []>("current_key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189687872)))];
+            tensor<fp16, [1, 1280, 1, 1]> current_key_cast_fp16 = conv(dilations = var_320, groups = var_277, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = var_318, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor<string, []>("current_key_cast_fp16")];
+            tensor<int32, [2]> var_325 = const()[name = tensor<string, []>("op_325"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_327 = const()[name = tensor<string, []>("op_327"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> current_value_pad_type_0 = const()[name = tensor<string, []>("current_value_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> current_value_pad_0 = const()[name = tensor<string, []>("current_value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192964736)))];
+            tensor<fp16, [1280]> layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196241600)))];
+            tensor<fp16, [1, 1280, 1, 1]> current_value_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = var_327, groups = var_277, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = var_325, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor<string, []>("current_value_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 224]> var_334_cast_fp16 = mul(x = current_key_cast_fp16, y = var_118_cast_fp16)[name = tensor<string, []>("op_334_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 224]> var_336_cast_fp16 = mul(x = var_43_cast_fp16_1, y = var_121_cast_fp16)[name = tensor<string, []>("op_336_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 224]> key_5_cast_fp16 = add(x = var_334_cast_fp16, y = var_336_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 224]> var_338_cast_fp16 = mul(x = current_value_cast_fp16, y = var_118_cast_fp16)[name = tensor<string, []>("op_338_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 224]> var_340_cast_fp16 = mul(x = var_48_cast_fp16_1, y = var_121_cast_fp16)[name = tensor<string, []>("op_340_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 224]> value_5_cast_fp16 = add(x = var_338_cast_fp16, y = var_340_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")];
+            tensor<int32, [4]> var_343 = const()[name = tensor<string, []>("op_343"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1]> var_344_cast_fp16 = reshape(shape = var_343, x = query_5_cast_fp16)[name = tensor<string, []>("op_344_cast_fp16")];
+            tensor<fp16, []> var_345_to_fp16 = const()[name = tensor<string, []>("op_345_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1]> var_346_cast_fp16 = mul(x = var_344_cast_fp16, y = var_345_to_fp16)[name = tensor<string, []>("op_346_cast_fp16")];
+            tensor<int32, [4]> var_347 = const()[name = tensor<string, []>("op_347"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 224]> var_348_cast_fp16 = reshape(shape = var_347, x = key_5_cast_fp16)[name = tensor<string, []>("op_348_cast_fp16")];
+            tensor<bool, []> mh_w_7_transpose_x_0 = const()[name = tensor<string, []>("mh_w_7_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_7_transpose_y_0 = const()[name = tensor<string, []>("mh_w_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1, 224]> mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_346_cast_fp16, y = var_348_cast_fp16)[name = tensor<string, []>("mh_w_7_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 224]> mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_139_cast_fp16)[name = tensor<string, []>("mh_w_9_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 224]> var_356_cast_fp16 = softmax(axis = var_270, x = mh_w_9_cast_fp16)[name = tensor<string, []>("op_356_cast_fp16")];
+            tensor<int32, [4]> var_357 = const()[name = tensor<string, []>("op_357"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 224]> var_358_cast_fp16 = reshape(shape = var_357, x = value_5_cast_fp16)[name = tensor<string, []>("op_358_cast_fp16")];
+            tensor<bool, []> attn_5_transpose_x_0 = const()[name = tensor<string, []>("attn_5_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_5_transpose_y_0 = const()[name = tensor<string, []>("attn_5_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_358_cast_fp16, y = var_356_cast_fp16)[name = tensor<string, []>("attn_5_cast_fp16")];
+            tensor<int32, [4]> var_361 = const()[name = tensor<string, []>("op_361"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1]> input_11_cast_fp16 = reshape(shape = var_361, x = attn_5_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<int32, [2]> var_365 = const()[name = tensor<string, []>("op_365"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_367 = const()[name = tensor<string, []>("op_367"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_19_pad_type_0 = const()[name = tensor<string, []>("obj_19_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_19_pad_0 = const()[name = tensor<string, []>("obj_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196244224)))];
+            tensor<fp16, [1280]> layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199521088)))];
+            tensor<fp16, [1, 1280, 1, 1]> obj_19_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = var_367, groups = var_277, pad = obj_19_pad_0, pad_type = obj_19_pad_type_0, strides = var_365, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("obj_19_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_19_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")];
+            tensor<int32, [1]> var_377 = const()[name = tensor<string, []>("op_377"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> channels_mean_9_cast_fp16 = reduce_mean(axes = var_377, keep_dims = var_278, x = inputs_9_cast_fp16)[name = tensor<string, []>("channels_mean_9_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> zero_mean_9_cast_fp16 = sub(x = inputs_9_cast_fp16, y = channels_mean_9_cast_fp16)[name = tensor<string, []>("zero_mean_9_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = zero_mean_9_cast_fp16)[name = tensor<string, []>("zero_mean_sq_9_cast_fp16")];
+            tensor<int32, [1]> var_381 = const()[name = tensor<string, []>("op_381"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> var_382_cast_fp16 = reduce_mean(axes = var_381, keep_dims = var_278, x = zero_mean_sq_9_cast_fp16)[name = tensor<string, []>("op_382_cast_fp16")];
+            tensor<fp16, []> var_383_to_fp16 = const()[name = tensor<string, []>("op_383_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1]> var_384_cast_fp16 = add(x = var_382_cast_fp16, y = var_383_to_fp16)[name = tensor<string, []>("op_384_cast_fp16")];
+            tensor<fp16, []> denom_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [1, 1, 1, 1]> denom_9_cast_fp16 = rsqrt(epsilon = denom_9_epsilon_0_to_fp16, x = var_384_cast_fp16)[name = tensor<string, []>("denom_9_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> out_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = denom_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
+            tensor<fp16, [1280]> obj_21_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_21_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199523712)))];
+            tensor<fp16, [1280]> obj_21_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_21_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199526336)))];
+            tensor<fp16, []> obj_21_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_21_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> obj_21_cast_fp16 = batch_norm(beta = obj_21_beta_0_to_fp16, epsilon = obj_21_epsilon_0_to_fp16, gamma = obj_21_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor<string, []>("obj_21_cast_fp16")];
+            tensor<int32, [2]> var_399 = const()[name = tensor<string, []>("op_399"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_401 = const()[name = tensor<string, []>("op_401"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> query_pad_type_0 = const()[name = tensor<string, []>("query_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> query_pad_0 = const()[name = tensor<string, []>("query_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199528960)))];
+            tensor<fp16, [1280]> layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202805824)))];
+            tensor<fp16, [1, 1280, 1, 1]> query_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = var_401, groups = var_277, pad = query_pad_0, pad_type = query_pad_type_0, strides = var_399, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_21_cast_fp16)[name = tensor<string, []>("query_cast_fp16")];
+            tensor<int32, [2]> var_405 = const()[name = tensor<string, []>("op_405"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_407 = const()[name = tensor<string, []>("op_407"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> key_pad_type_0 = const()[name = tensor<string, []>("key_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> key_pad_0 = const()[name = tensor<string, []>("key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202808448)))];
+            tensor<fp16, [1, 1280, 1, 1500]> key_cast_fp16 = conv(dilations = var_407, groups = var_277, pad = key_pad_0, pad_type = key_pad_type_0, strides = var_405, weight = layers_1_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("key_cast_fp16")];
+            tensor<int32, [2]> var_412 = const()[name = tensor<string, []>("op_412"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_414 = const()[name = tensor<string, []>("op_414"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> value_pad_type_0 = const()[name = tensor<string, []>("value_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> value_pad_0 = const()[name = tensor<string, []>("value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206085312)))];
+            tensor<fp16, [1280]> layers_1_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209362176)))];
+            tensor<fp16, [1, 1280, 1, 1500]> value_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_bias_to_fp16, dilations = var_414, groups = var_277, pad = value_pad_0, pad_type = value_pad_type_0, strides = var_412, weight = layers_1_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("value_cast_fp16")];
+            tensor<int32, [4]> var_418 = const()[name = tensor<string, []>("op_418"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1]> var_419_cast_fp16 = reshape(shape = var_418, x = query_cast_fp16)[name = tensor<string, []>("op_419_cast_fp16")];
+            tensor<fp16, []> var_420_to_fp16 = const()[name = tensor<string, []>("op_420_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1]> var_421_cast_fp16 = mul(x = var_419_cast_fp16, y = var_420_to_fp16)[name = tensor<string, []>("op_421_cast_fp16")];
+            tensor<int32, [4]> var_422 = const()[name = tensor<string, []>("op_422"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_423_cast_fp16 = reshape(shape = var_422, x = key_cast_fp16)[name = tensor<string, []>("op_423_cast_fp16")];
+            tensor<bool, []> mh_w_transpose_x_0 = const()[name = tensor<string, []>("mh_w_transpose_x_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> mh_w_transpose_y_0 = const()[name = tensor<string, []>("mh_w_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 20, 1, 1500]> mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_421_cast_fp16, y = var_423_cast_fp16)[name = tensor<string, []>("mh_w_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 1500]> var_426_cast_fp16 = softmax(axis = var_270, x = mh_w_cast_fp16)[name = tensor<string, []>("op_426_cast_fp16")];
+            tensor<int32, [4]> var_427 = const()[name = tensor<string, []>("op_427"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1500]> var_428_cast_fp16 = reshape(shape = var_427, x = value_cast_fp16)[name = tensor<string, []>("op_428_cast_fp16")];
+            tensor<bool, []> attn_transpose_x_0 = const()[name = tensor<string, []>("attn_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attn_transpose_y_0 = const()[name = tensor<string, []>("attn_transpose_y_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 20, 64, 1]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_428_cast_fp16, y = var_426_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")];
+            tensor<int32, [4]> var_431 = const()[name = tensor<string, []>("op_431"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1]> input_13_cast_fp16 = reshape(shape = var_431, x = attn_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
+            tensor<int32, [2]> var_435 = const()[name = tensor<string, []>("op_435"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_437 = const()[name = tensor<string, []>("op_437"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> obj_23_pad_type_0 = const()[name = tensor<string, []>("obj_23_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> obj_23_pad_0 = const()[name = tensor<string, []>("obj_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209364800)))];
+            tensor<fp16, [1280]> layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212641664)))];
+            tensor<fp16, [1, 1280, 1, 1]> obj_23_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = var_437, groups = var_277, pad = obj_23_pad_0, pad_type = obj_23_pad_type_0, strides = var_435, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("obj_23_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_23_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")];
+            tensor<int32, [1]> var_443 = const()[name = tensor<string, []>("op_443"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> channels_mean_11_cast_fp16 = reduce_mean(axes = var_443, keep_dims = var_278, x = inputs_11_cast_fp16)[name = tensor<string, []>("channels_mean_11_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> zero_mean_11_cast_fp16 = sub(x = inputs_11_cast_fp16, y = channels_mean_11_cast_fp16)[name = tensor<string, []>("zero_mean_11_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = zero_mean_11_cast_fp16)[name = tensor<string, []>("zero_mean_sq_11_cast_fp16")];
+            tensor<int32, [1]> var_447 = const()[name = tensor<string, []>("op_447"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> var_448_cast_fp16 = reduce_mean(axes = var_447, keep_dims = var_278, x = zero_mean_sq_11_cast_fp16)[name = tensor<string, []>("op_448_cast_fp16")];
+            tensor<fp16, []> var_449_to_fp16 = const()[name = tensor<string, []>("op_449_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1]> var_450_cast_fp16 = add(x = var_448_cast_fp16, y = var_449_to_fp16)[name = tensor<string, []>("op_450_cast_fp16")];
+            tensor<fp16, []> denom_11_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_11_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [1, 1, 1, 1]> denom_11_cast_fp16 = rsqrt(epsilon = denom_11_epsilon_0_to_fp16, x = var_450_cast_fp16)[name = tensor<string, []>("denom_11_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> out_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = denom_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
+            tensor<fp16, [1280]> input_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_15_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212644288)))];
+            tensor<fp16, [1280]> input_15_beta_0_to_fp16 = const()[name = tensor<string, []>("input_15_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212646912)))];
+            tensor<fp16, []> input_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> input_15_cast_fp16 = batch_norm(beta = input_15_beta_0_to_fp16, epsilon = input_15_epsilon_0_to_fp16, gamma = input_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
+            tensor<int32, [2]> var_461 = const()[name = tensor<string, []>("op_461"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_463 = const()[name = tensor<string, []>("op_463"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> input_17_pad_type_0 = const()[name = tensor<string, []>("input_17_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_1_fc1_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212649536)))];
+            tensor<fp16, [5120]> layers_1_fc1_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(225756800)))];
+            tensor<fp16, [1, 5120, 1, 1]> input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = var_463, groups = var_277, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = var_461, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
+            tensor<string, []> input_mode_0 = const()[name = tensor<string, []>("input_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1]> input_cast_fp16 = gelu(mode = input_mode_0, x = input_17_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [2]> var_469 = const()[name = tensor<string, []>("op_469"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [2]> var_471 = const()[name = tensor<string, []>("op_471"), val = tensor<int32, [2]>([1, 1])];
+            tensor<string, []> hidden_states_5_pad_type_0 = const()[name = tensor<string, []>("hidden_states_5_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [4]> hidden_states_5_pad_0 = const()[name = tensor<string, []>("hidden_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_1_fc2_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(225767104)))];
+            tensor<fp16, [1280]> layers_1_fc2_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238874368)))];
+            tensor<fp16, [1, 1280, 1, 1]> hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = var_471, groups = var_277, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = var_469, weight = layers_1_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("hidden_states_5_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")];
+            tensor<bool, []> var_481 = const()[name = tensor<string, []>("op_481"), val = tensor<bool, []>(true)];
+            tensor<int32, [1]> var_485 = const()[name = tensor<string, []>("op_485"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> channels_mean_cast_fp16 = reduce_mean(axes = var_485, keep_dims = var_481, x = inputs_cast_fp16)[name = tensor<string, []>("channels_mean_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> zero_mean_cast_fp16 = sub(x = inputs_cast_fp16, y = channels_mean_cast_fp16)[name = tensor<string, []>("zero_mean_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> zero_mean_sq_cast_fp16 = mul(x = zero_mean_cast_fp16, y = zero_mean_cast_fp16)[name = tensor<string, []>("zero_mean_sq_cast_fp16")];
+            tensor<int32, [1]> var_489 = const()[name = tensor<string, []>("op_489"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 1]> var_490_cast_fp16 = reduce_mean(axes = var_489, keep_dims = var_481, x = zero_mean_sq_cast_fp16)[name = tensor<string, []>("op_490_cast_fp16")];
+            tensor<fp16, []> var_491_to_fp16 = const()[name = tensor<string, []>("op_491_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 1, 1]> var_492_cast_fp16 = add(x = var_490_cast_fp16, y = var_491_to_fp16)[name = tensor<string, []>("op_492_cast_fp16")];
+            tensor<fp16, []> denom_epsilon_0_to_fp16 = const()[name = tensor<string, []>("denom_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [1, 1, 1, 1]> denom_cast_fp16 = rsqrt(epsilon = denom_epsilon_0_to_fp16, x = var_492_cast_fp16)[name = tensor<string, []>("denom_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> out_cast_fp16 = mul(x = zero_mean_cast_fp16, y = denom_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
+            tensor<fp16, [1280]> hidden_states_gamma_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238876992)))];
+            tensor<fp16, [1280]> hidden_states_beta_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238879616)))];
+            tensor<fp16, []> hidden_states_epsilon_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")];
+            tensor<int32, [1]> var_502_axes_0 = const()[name = tensor<string, []>("op_502_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1280, 1]> var_502_cast_fp16 = squeeze(axes = var_502_axes_0, x = hidden_states_cast_fp16)[name = tensor<string, []>("op_502_cast_fp16")];
+            tensor<int32, [3]> var_505_perm_0 = const()[name = tensor<string, []>("op_505_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [51866]> linear_0_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_0_bias_0_to_fp16"), val = tensor<fp16, [51866]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238882240)))];
+            tensor<fp16, [1, 1, 1280]> transpose_0 = transpose(perm = var_505_perm_0, x = var_502_cast_fp16)[name = tensor<string, []>("transpose_0")];
+            tensor<fp16, [1, 1, 51866]> logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = transpose_0)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<int32, []> var_509 = const()[name = tensor<string, []>("op_509"), val = tensor<int32, []>(1)];
+            tensor<bool, []> obj_27_interleave_0 = const()[name = tensor<string, []>("obj_27_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 2560, 1, 1]> key_cache_updates = concat(axis = var_509, interleave = obj_27_interleave_0, values = (current_key_1_cast_fp16, current_key_cast_fp16))[name = tensor<string, []>("obj_27_cast_fp16")];
+            tensor<int32, []> var_512 = const()[name = tensor<string, []>("op_512"), val = tensor<int32, []>(1)];
+            tensor<bool, []> obj_interleave_0 = const()[name = tensor<string, []>("obj_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 2560, 1, 1]> value_cache_updates = concat(axis = var_512, interleave = obj_interleave_0, values = (current_value_1_cast_fp16, current_value_cast_fp16))[name = tensor<string, []>("obj_cast_fp16")];
+        } -> (logits, key_cache_updates, value_cache_updates);
+}

distil-whisper_distil-large-v3_turbo_600MB/TextDecoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7415eed816b4f2f69ce64ece61328f9f60f96e2201a45cb01caca6d413cc6e94
+size 238986036

distil-whisper_distil-large-v3_turbo_600MB/TextDecoderContextPrefill.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:177b3e51ce243362f79be85ee3d422665b49225bbb1993d72678a79a4788372e
+size 243

distil-whisper_distil-large-v3_turbo_600MB/TextDecoderContextPrefill.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1fb28c51275e24a93535cf363012bbff21f6d486a8238d5fcf93d6288e2f781
+size 380

distil-whisper_distil-large-v3_turbo_600MB/TextDecoderContextPrefill.mlmodelc/metadata.json

@@ -0,0 +1,82 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 2560 × 1 × 3)",
+        "shortDescription" : "",
+        "shape" : "[1, 2560, 1, 3]",
+        "name" : "key_cache_prefill",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 2560 × 1 × 3)",
+        "shortDescription" : "",
+        "shape" : "[1, 2560, 1, 3]",
+        "name" : "value_cache_prefill",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.mul" : 1,
+      "Ios17.cast" : 1,
+      "Ios17.sub" : 1,
+      "Ios17.reshape" : 2,
+      "Ios17.add" : 1,
+      "Ios17.gather" : 2
+    },
+    "computePrecision" : "Mixed (Float16, Int16, Int32)",
+    "isUpdatable" : "0",
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.version" : "7.1",
+      "com.github.apple.coremltools.source" : "torch==2.2.1"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "task",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "language",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "TextDecoderContextPrefill",
+    "method" : "predict"
+  }
+]

distil-whisper_distil-large-v3_turbo_600MB/TextDecoderContextPrefill.mlmodelc/model.mil

@@ -0,0 +1,27 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "2.2.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "7.1"}})]
+{
+    func main<ios17>(tensor<int32, [1]> language, tensor<int32, [1]> task) {
+            tensor<int32, []> var_6 = const()[name = tensor<string, []>("op_6"), val = tensor<int32, []>(50259)];
+            tensor<int32, [1]> var_7 = sub(x = language, y = var_6)[name = tensor<string, []>("op_7")];
+            tensor<int32, []> var_8 = const()[name = tensor<string, []>("op_8"), val = tensor<int32, []>(2)];
+            tensor<int32, [1]> var_9 = mul(x = var_7, y = var_8)[name = tensor<string, []>("op_9")];
+            tensor<int32, [1]> input = add(x = var_9, y = task)[name = tensor<string, []>("input")];
+            tensor<int32, []> var_15_axis_0 = const()[name = tensor<string, []>("op_15_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> var_15_batch_dims_0 = const()[name = tensor<string, []>("op_15_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> var_15_validate_indices_0 = const()[name = tensor<string, []>("op_15_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [200, 7680]> key_cache_lut_weight_to_fp16 = const()[name = tensor<string, []>("key_cache_lut_weight_to_fp16"), val = tensor<fp16, [200, 7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<string, []> input_to_int16_dtype_0 = const()[name = tensor<string, []>("input_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1]> cast_6 = cast(dtype = input_to_int16_dtype_0, x = input)[name = tensor<string, []>("cast_6")];
+            tensor<fp16, [1, 7680]> var_15_cast_fp16_cast_int16 = gather(axis = var_15_axis_0, batch_dims = var_15_batch_dims_0, indices = cast_6, validate_indices = var_15_validate_indices_0, x = key_cache_lut_weight_to_fp16)[name = tensor<string, []>("op_15_cast_fp16_cast_int16")];
+            tensor<int32, [4]> var_20 = const()[name = tensor<string, []>("op_20"), val = tensor<int32, [4]>([1, 2560, 1, 3])];
+            tensor<fp16, [1, 2560, 1, 3]> key_cache_prefill = reshape(shape = var_20, x = var_15_cast_fp16_cast_int16)[name = tensor<string, []>("op_21_cast_fp16")];
+            tensor<int32, []> var_25_axis_0 = const()[name = tensor<string, []>("op_25_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> var_25_batch_dims_0 = const()[name = tensor<string, []>("op_25_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> var_25_validate_indices_0 = const()[name = tensor<string, []>("op_25_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [200, 7680]> value_cache_lut_weight_to_fp16 = const()[name = tensor<string, []>("value_cache_lut_weight_to_fp16"), val = tensor<fp16, [200, 7680]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3072128)))];
+            tensor<fp16, [1, 7680]> var_25_cast_fp16_cast_int16 = gather(axis = var_25_axis_0, batch_dims = var_25_batch_dims_0, indices = cast_6, validate_indices = var_25_validate_indices_0, x = value_cache_lut_weight_to_fp16)[name = tensor<string, []>("op_25_cast_fp16_cast_int16")];
+            tensor<int32, [4]> var_30 = const()[name = tensor<string, []>("op_30"), val = tensor<int32, [4]>([1, 2560, 1, 3])];
+            tensor<fp16, [1, 2560, 1, 3]> value_cache_prefill = reshape(shape = var_30, x = var_25_cast_fp16_cast_int16)[name = tensor<string, []>("op_31_cast_fp16")];
+        } -> (key_cache_prefill, value_cache_prefill);
+}

distil-whisper_distil-large-v3_turbo_600MB/TextDecoderContextPrefill.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26266f85b3830f13b43e4fec705e3999329b1af171b23f6eff4e89fe9c137ae0
+size 6144192

distil-whisper_distil-large-v3_turbo_600MB/config.json

@@ -0,0 +1 @@
+{"_name_or_path": "./distil-large-v3", "activation_dropout": 0.0, "activation_function": "gelu", "apply_spec_augment": false, "architectures": ["WhisperForConditionalGeneration"], "attention_dropout": 0.0, "begin_suppress_tokens": [220, 50257], "bos_token_id": 50257, "classifier_proj_size": 256, "d_model": 1280, "decoder_attention_heads": 20, "decoder_ffn_dim": 5120, "decoder_layerdrop": 0.0, "decoder_layers": 2, "decoder_start_token_id": 50258, "dropout": 0.0, "encoder_attention_heads": 20, "encoder_ffn_dim": 5120, "encoder_layerdrop": 0.0, "encoder_layers": 32, "eos_token_id": 50257, "init_std": 0.02, "is_encoder_decoder": true, "mask_feature_length": 10, "mask_feature_min_masks": 0, "mask_feature_prob": 0.0, "mask_time_length": 10, "mask_time_min_masks": 2, "mask_time_prob": 0.05, "max_length": 448, "max_source_positions": 1500, "max_target_positions": 448, "median_filter_width": 7, "model_type": "whisper", "num_hidden_layers": 32, "num_mel_bins": 128, "pad_token_id": 50256, "scale_embedding": false, "torch_dtype": "float16", "transformers_version": "4.38.0.dev0", "use_cache": true, "use_weighted_layer_sum": false, "vocab_size": 51866}

distil-whisper_distil-large-v3_turbo_600MB/generation_config.json

@@ -0,0 +1 @@
+{"alignment_heads": [[7, 0], [10, 17], [12, 18], [13, 12], [16, 1], [17, 14], [19, 11], [21, 4], [24, 1], [25, 6]], "begin_suppress_tokens": [220, 50257], "bos_token_id": 50257, "decoder_start_token_id": 50258, "eos_token_id": 50257, "forced_decoder_ids": [[1, null], [2, 50360]], "is_multilingual": true, "lang_to_id": {"<|af|>": 50327, "<|am|>": 50334, "<|ar|>": 50272, "<|as|>": 50350, "<|az|>": 50304, "<|ba|>": 50355, "<|be|>": 50330, "<|bg|>": 50292, "<|bn|>": 50302, "<|bo|>": 50347, "<|br|>": 50309, "<|bs|>": 50315, "<|ca|>": 50270, "<|cs|>": 50283, "<|cy|>": 50297, "<|da|>": 50285, "<|de|>": 50261, "<|el|>": 50281, "<|en|>": 50259, "<|es|>": 50262, "<|et|>": 50307, "<|eu|>": 50310, "<|fa|>": 50300, "<|fi|>": 50277, "<|fo|>": 50338, "<|fr|>": 50265, "<|gl|>": 50319, "<|gu|>": 50333, "<|haw|>": 50352, "<|ha|>": 50354, "<|he|>": 50279, "<|hi|>": 50276, "<|hr|>": 50291, "<|ht|>": 50339, "<|hu|>": 50286, "<|hy|>": 50312, "<|id|>": 50275, "<|is|>": 50311, "<|it|>": 50274, "<|ja|>": 50266, "<|jw|>": 50356, "<|ka|>": 50329, "<|kk|>": 50316, "<|km|>": 50323, "<|kn|>": 50306, "<|ko|>": 50264, "<|la|>": 50294, "<|lb|>": 50345, "<|ln|>": 50353, "<|lo|>": 50336, "<|lt|>": 50293, "<|lv|>": 50301, "<|mg|>": 50349, "<|mi|>": 50295, "<|mk|>": 50308, "<|ml|>": 50296, "<|mn|>": 50314, "<|mr|>": 50320, "<|ms|>": 50282, "<|mt|>": 50343, "<|my|>": 50346, "<|ne|>": 50313, "<|nl|>": 50271, "<|nn|>": 50342, "<|no|>": 50288, "<|oc|>": 50328, "<|pa|>": 50321, "<|pl|>": 50269, "<|ps|>": 50340, "<|pt|>": 50267, "<|ro|>": 50284, "<|ru|>": 50263, "<|sa|>": 50344, "<|sd|>": 50332, "<|si|>": 50322, "<|sk|>": 50298, "<|sl|>": 50305, "<|sn|>": 50324, "<|so|>": 50326, "<|sq|>": 50317, "<|sr|>": 50303, "<|su|>": 50357, "<|sv|>": 50273, "<|sw|>": 50318, "<|ta|>": 50287, "<|te|>": 50299, "<|tg|>": 50331, "<|th|>": 50289, "<|tk|>": 50341, "<|tl|>": 50348, "<|tr|>": 50268, "<|tt|>": 50351, "<|uk|>": 50280, "<|ur|>": 50290, "<|uz|>": 50337, "<|vi|>": 50278, "<|yi|>": 50335, "<|yo|>": 50325, "<|yue|>": 50358, "<|zh|>": 50260}, "language": "<|en|>", "max_initial_timestamp_index": 50, "max_length": 448, "no_timestamps_token_id": 50364, "pad_token_id": 50257, "prev_sot_token_id": 50362, "return_timestamps": false, "suppress_tokens": [1, 2, 7, 8, 9, 10, 14, 25, 26, 27, 28, 29, 31, 58, 59, 60, 61, 62, 63, 90, 91, 92, 93, 359, 503, 522, 542, 873, 893, 902, 918, 922, 931, 1350, 1853, 1982, 2460, 2627, 3246, 3253, 3268, 3536, 3846, 3961, 4183, 4667, 6585, 6647, 7273, 9061, 9383, 10428, 10929, 11938, 12033, 12331, 12562, 13793, 14157, 14635, 15265, 15618, 16553, 16604, 18362, 18956, 20075, 21675, 22520, 26130, 26161, 26435, 28279, 29464, 31650, 32302, 32470, 36865, 42863, 47425, 49870, 50254, 50258, 50359, 50360, 50361, 50362, 50363], "task": "transcribe", "task_to_id": {"transcribe": 50360, "translate": 50359}, "transformers_version": "4.38.0.dev0"}