shimaa22 commited on
Commit
8dc7ba2
·
verified ·
1 Parent(s): f29f8d6

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +30 -20
app.py CHANGED
@@ -2,23 +2,31 @@ from fastapi import FastAPI, HTTPException
2
  from pydantic import BaseModel
3
  import torch
4
  from transformers import AutoModelForCausalLM, AutoTokenizer
 
5
 
6
  app = FastAPI(title="Qwen API Server")
7
 
8
- # إعدادات المودل (تم اختيار 3B ليتناسب مع الـ CPU)
9
  MODEL_NAME = "Qwen/Qwen2.5-3B-Instruct"
10
 
11
- print(f"Loading Model: {MODEL_NAME} ...")
12
- # تحميل الـ Tokenizer والـ Model مرة واحدة فقط عند بدء تشغيل السيرفر
13
- tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
 
14
 
15
- # تحميل على الـ CPU واستخدام نوع بيانات خفيف (float32 أو bfloat16 إذا مدعوم) لتوفير الذاكرة
16
- model = AutoModelForCausalLM.from_pretrained(
17
- MODEL_NAME,
18
- torch_dtype=torch.float32,
19
- device_map="cpu"
20
- )
21
- print("Model successfully loaded on CPU!")
 
 
 
 
 
 
 
22
 
23
  class ChatMessage(BaseModel):
24
  role: str
@@ -29,10 +37,18 @@ class ChatRequest(BaseModel):
29
  max_tokens: int = 512
30
  temperature: float = 0.7
31
 
 
 
 
 
 
 
32
  @app.post("/v1/chat/completions")
33
  async def chat_completion(request: ChatRequest):
 
 
 
34
  try:
35
- # تحويل الرسائل إلى صيغة الـ Chat Template الخاصة بـ Qwen
36
  formatted_messages = [{"role": m.role, "content": m.content} for m in request.messages]
37
 
38
  text = tokenizer.apply_chat_template(
@@ -43,17 +59,15 @@ async def chat_completion(request: ChatRequest):
43
 
44
  inputs = tokenizer([text], return_tensors="pt")
45
 
46
- # توليد الرد
47
  with torch.no_grad():
48
  outputs = model.generate(
49
  **inputs,
50
  max_new_tokens=request.max_tokens,
51
- temperature=request.max_tokens, # أو استخدامه بشكل طبيعي
52
  do_sample=True,
53
  top_p=0.9
54
  )
55
 
56
- # استخراج النص المُولد فقط (بدون المدخلات)
57
  generated_ids = [
58
  output[len(input_ids):] for input_ids, output in zip(inputs.input_ids, outputs)
59
  ]
@@ -72,10 +86,6 @@ async def chat_completion(request: ChatRequest):
72
  except Exception as e:
73
  raise HTTPException(status_code=500, detail=str(e))
74
 
75
- @app.get("/")
76
- def home():
77
- return {"status": "Running", "model": MODEL_NAME}
78
-
79
  if __name__ == "__main__":
80
  import uvicorn
81
- uvicorn.run(app, host="0.0.0.0", port=7860) # Hugging Face يستخدم غالباً البورت 7860
 
2
  from pydantic import BaseModel
3
  import torch
4
  from transformers import AutoModelForCausalLM, AutoTokenizer
5
+ import threading
6
 
7
  app = FastAPI(title="Qwen API Server")
8
 
 
9
  MODEL_NAME = "Qwen/Qwen2.5-3B-Instruct"
10
 
11
+ # متغيرات لحفظ حالة التحميل
12
+ tokenizer = None
13
+ model = None
14
+ is_model_loaded = False
15
 
16
+ def load_model_background():
17
+ global tokenizer, model, is_model_loaded
18
+ print(f"Loading Model in background: {MODEL_NAME} ...")
19
+ tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
20
+ model = AutoModelForCausalLM.from_pretrained(
21
+ MODEL_NAME,
22
+ torch_dtype=torch.float32,
23
+ device_map="cpu"
24
+ )
25
+ is_model_loaded = True
26
+ print("Model successfully loaded on CPU!")
27
+
28
+ # تشغيل تحميل النموذج في Background Thread لكي يبدأ الـ FastAPI فوراً ولا يتم عمل Shutdown
29
+ threading.Thread(target=load_model_background).start()
30
 
31
  class ChatMessage(BaseModel):
32
  role: str
 
37
  max_tokens: int = 512
38
  temperature: float = 0.7
39
 
40
+ @app.get("/")
41
+ def home():
42
+ if not is_model_loaded:
43
+ return {"status": "Loading model...", "model": MODEL_NAME}
44
+ return {"status": "Running", "model": MODEL_NAME}
45
+
46
  @app.post("/v1/chat/completions")
47
  async def chat_completion(request: ChatRequest):
48
+ if not is_model_loaded:
49
+ raise HTTPException(status_code=503, detail="Model is still loading, please try again in a minute.")
50
+
51
  try:
 
52
  formatted_messages = [{"role": m.role, "content": m.content} for m in request.messages]
53
 
54
  text = tokenizer.apply_chat_template(
 
59
 
60
  inputs = tokenizer([text], return_tensors="pt")
61
 
 
62
  with torch.no_grad():
63
  outputs = model.generate(
64
  **inputs,
65
  max_new_tokens=request.max_tokens,
66
+ temperature=request.temperature,
67
  do_sample=True,
68
  top_p=0.9
69
  )
70
 
 
71
  generated_ids = [
72
  output[len(input_ids):] for input_ids, output in zip(inputs.input_ids, outputs)
73
  ]
 
86
  except Exception as e:
87
  raise HTTPException(status_code=500, detail=str(e))
88
 
 
 
 
 
89
  if __name__ == "__main__":
90
  import uvicorn
91
+ uvicorn.run(app, host="0.0.0.0", port=7860)