Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,13 +1,69 @@
|
|
1 |
-
import
|
|
|
|
|
2 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
|
|
3 |
|
4 |
-
|
5 |
-
model = AutoModelForCausalLM.from_pretrained("IlyaGusev/saiga_gemma2_10b")
|
6 |
|
7 |
-
def
|
8 |
-
|
9 |
-
|
10 |
-
return tokenizer.decode(output[0], skip_special_tokens=True)
|
11 |
|
12 |
-
|
13 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
from flask import Flask, request, jsonify
|
3 |
+
import torch
|
4 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
5 |
+
from huggingface_hub import login
|
6 |
|
7 |
+
app = Flask(__name__)
|
|
|
8 |
|
9 |
+
def init_model():
|
10 |
+
global model, tokenizer
|
11 |
+
hf_token = os.getenv("HF_TOKEN") # Чтение токена из переменной окружения
|
|
|
12 |
|
13 |
+
if hf_token is None:
|
14 |
+
raise ValueError("Hugging Face token is not set. Please set the HF_TOKEN environment variable.")
|
15 |
+
|
16 |
+
# Аутентификация с использованием токена
|
17 |
+
login(hf_token, add_to_git_credential=True)
|
18 |
+
|
19 |
+
# Загрузка модели и токенизатора без квантования и без распределения на CPU/диск
|
20 |
+
tokenizer = AutoTokenizer.from_pretrained("IlyaGusev/saiga_gemma2_10b", token=hf_token)
|
21 |
+
model = AutoModelForCausalLM.from_pretrained(
|
22 |
+
"IlyaGusev/saiga_gemma2_10b",
|
23 |
+
token=hf_token,
|
24 |
+
torch_dtype=torch.float16, # Использование float16 для уменьшения потребления памяти
|
25 |
+
device_map=None # Не использовать автоматическое распределение на CPU/диск
|
26 |
+
)
|
27 |
+
|
28 |
+
# Явное перемещение модели на GPU, если доступно
|
29 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
30 |
+
model.to(device)
|
31 |
+
|
32 |
+
@app.route('/generate', methods=['POST'])
|
33 |
+
def generate_response():
|
34 |
+
try:
|
35 |
+
data = request.get_json()
|
36 |
+
print(f"Received data: {data}")
|
37 |
+
|
38 |
+
prompt = data.get('prompt', '')
|
39 |
+
max_length = data.get('max_length', 100)
|
40 |
+
temperature = data.get('temperature', 0.7)
|
41 |
+
top_p = data.get('top_p', 0.85)
|
42 |
+
repetition_penalty = data.get('repetition_penalty', 1.1)
|
43 |
+
|
44 |
+
input_ids = tokenizer.encode(prompt, return_tensors="pt").to(model.device)
|
45 |
+
attention_mask = torch.ones_like(input_ids).to(model.device)
|
46 |
+
|
47 |
+
output = model.generate(
|
48 |
+
input_ids,
|
49 |
+
attention_mask=attention_mask,
|
50 |
+
max_length=max_length,
|
51 |
+
temperature=temperature,
|
52 |
+
top_p=top_p,
|
53 |
+
repetition_penalty=repetition_penalty,
|
54 |
+
do_sample=True,
|
55 |
+
num_return_sequences=1,
|
56 |
+
pad_token_id=tokenizer.eos_token_id
|
57 |
+
)
|
58 |
+
print(f"Generated output: {output}")
|
59 |
+
response_text = tokenizer.decode(output[0], skip_special_tokens=True)
|
60 |
+
print(f"Generated response: {response_text}")
|
61 |
+
return jsonify({"response": response_text})
|
62 |
+
|
63 |
+
except Exception as e:
|
64 |
+
print(f"Error: {str(e)}")
|
65 |
+
return jsonify({"response": "Извините, произошла ошибка при генерации ответа."}), 500
|
66 |
+
|
67 |
+
if __name__ == "__main__":
|
68 |
+
init_model()
|
69 |
+
app.run(host='0.0.0.0', port=7860)
|