Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,114 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import json
|
2 |
+
import os.path as osp
|
3 |
+
import random
|
4 |
+
from typing import Union
|
5 |
+
import os
|
6 |
+
import sys
|
7 |
+
from typing import List
|
8 |
+
import torch
|
9 |
+
import transformers
|
10 |
+
from datasets import load_dataset
|
11 |
+
from transformers import AutoModelForCausalLM, TrainingArguments, Trainer
|
12 |
+
import gradio as gr
|
13 |
+
import torch.nn as nn
|
14 |
+
|
15 |
+
|
16 |
+
from peft import (
|
17 |
+
LoraConfig,
|
18 |
+
get_peft_model,
|
19 |
+
get_peft_model_state_dict,
|
20 |
+
prepare_model_for_int8_training,
|
21 |
+
set_peft_model_state_dict,
|
22 |
+
PeftModel
|
23 |
+
)
|
24 |
+
from transformers import LlamaForCausalLM, LlamaTokenizer
|
25 |
+
|
26 |
+
|
27 |
+
base_model='nickypro/tinyllama-15M'
|
28 |
+
|
29 |
+
|
30 |
+
class Prompter(object):
|
31 |
+
|
32 |
+
def generate_prompt(
|
33 |
+
self,
|
34 |
+
instruction: str,
|
35 |
+
label: Union[None, str] = None,
|
36 |
+
) -> str:
|
37 |
+
|
38 |
+
res = f"{instruction}\nAnswer: "
|
39 |
+
|
40 |
+
if label:
|
41 |
+
res = f"{res}{label}"
|
42 |
+
|
43 |
+
return res
|
44 |
+
|
45 |
+
def get_response(self, output: str) -> str:
|
46 |
+
return output.split("Answer:")[1].strip().replace("/", "\u00F7").replace("*", "\u00D7")
|
47 |
+
|
48 |
+
model = LlamaForCausalLM.from_pretrained(
|
49 |
+
base_model,
|
50 |
+
torch_dtype=torch.float32,
|
51 |
+
device_map="auto",
|
52 |
+
)
|
53 |
+
model = PeftModel.from_pretrained(
|
54 |
+
model,
|
55 |
+
f'checkpoint-16000',
|
56 |
+
torch_dtype=torch.float32,
|
57 |
+
)
|
58 |
+
|
59 |
+
model.eval()
|
60 |
+
if torch.__version__ >= "2" and sys.platform != "win32":
|
61 |
+
model = torch.compile(model)
|
62 |
+
|
63 |
+
tokenizer = LlamaTokenizer.from_pretrained('hf-internal-testing/llama-tokenizer')
|
64 |
+
tokenizer.pad_token_id = 0
|
65 |
+
tokenizer.padding_side = "left"
|
66 |
+
|
67 |
+
|
68 |
+
def generate_answers(instructions, model, tokenizer):
|
69 |
+
prompter = Prompter()
|
70 |
+
raw_answers = []
|
71 |
+
|
72 |
+
for instruction in instructions:
|
73 |
+
prompt = prompter.generate_prompt(instruction)
|
74 |
+
inputs = tokenizer(prompt, return_tensors="pt")
|
75 |
+
|
76 |
+
input_ids = inputs["input_ids"]
|
77 |
+
|
78 |
+
generation_output = model.generate(
|
79 |
+
input_ids=input_ids,
|
80 |
+
return_dict_in_generate=True,
|
81 |
+
output_scores=True,
|
82 |
+
pad_token_id=0,
|
83 |
+
eos_token_id=tokenizer.eos_token_id,
|
84 |
+
max_new_tokens=16
|
85 |
+
)
|
86 |
+
s = generation_output.sequences[0]
|
87 |
+
raw_answers.append(tokenizer.decode(s, skip_special_tokens=True).strip())
|
88 |
+
|
89 |
+
return raw_answers
|
90 |
+
|
91 |
+
|
92 |
+
def evaluate(instruction):
|
93 |
+
return generate_answers([instruction], model, tokenizer)[0]
|
94 |
+
|
95 |
+
|
96 |
+
if __name__ == "__main__":
|
97 |
+
gr.Interface(
|
98 |
+
fn=evaluate,
|
99 |
+
inputs=[
|
100 |
+
gr.components.Textbox(
|
101 |
+
lines=1,
|
102 |
+
label="Arithmetic",
|
103 |
+
placeholder="63303235 + 20239503",
|
104 |
+
)
|
105 |
+
],
|
106 |
+
outputs=[
|
107 |
+
gr.Textbox(
|
108 |
+
lines=5,
|
109 |
+
label="Output",
|
110 |
+
)
|
111 |
+
],
|
112 |
+
title="Arithmetic LLaMA",
|
113 |
+
description="This model is 15M llama model, finetuned on a+b tasks",
|
114 |
+
).queue().launch(server_name="0.0.0.0", share=True)
|