Cozzy23 commited on
Commit
4e3d5d2
·
verified ·
1 Parent(s): e74ee40

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +22 -42
app.py CHANGED
@@ -1,25 +1,15 @@
1
- """
2
- FastAPI version of C9-AI for Hugging Face Docker Spaces.
3
- Now powered by Qwen3-0.6B — smarter, faster, and still under 1 GB.
4
- """
5
-
6
  import os
7
  from typing import List, Optional
8
-
9
  import torch
10
  from fastapi import FastAPI, HTTPException
11
  from fastapi.middleware.cors import CORSMiddleware
12
  from pydantic import BaseModel
13
  from transformers import AutoModelForCausalLM, AutoTokenizer
14
 
15
- # Use Qwen3-0.6B (Chat) — official, lightweight, and smarter
16
- MODEL_NAME = os.getenv("C9AI_MODEL_NAME", "Qwen/Qwen3-0.6B")
17
 
18
- app = FastAPI(
19
- title="C9-AI API",
20
- description="Advanced AI Framework for Research, Coding, Reasoning",
21
- version="1.0.5"
22
- )
23
 
24
  app.add_middleware(
25
  CORSMiddleware,
@@ -38,14 +28,10 @@ class Message(BaseModel):
38
 
39
  class ChatRequest(BaseModel):
40
  messages: List[Message]
41
- max_tokens: Optional[int] = 1024 # Reduced for speed
42
  temperature: Optional[float] = 0.7
43
  top_p: Optional[float] = 0.8
44
 
45
- class ChatResponse(BaseModel):
46
- choices: List[dict]
47
- model: str = MODEL_NAME
48
-
49
  @app.on_event("startup")
50
  async def load_model():
51
  global MODEL, TOKENIZER
@@ -53,7 +39,7 @@ async def load_model():
53
 
54
  TOKENIZER = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
55
 
56
- # Use 4-bit quantization if possible (saves RAM, speeds up inference)
57
  try:
58
  from transformers import BitsAndBytesConfig
59
  bnb_config = BitsAndBytesConfig(
@@ -67,9 +53,9 @@ async def load_model():
67
  device_map="auto",
68
  trust_remote_code=True
69
  )
70
- print("✅ Loaded in 4-bit quantized mode")
71
  except Exception as e:
72
- print(f"⚠️ 4-bit failed ({e}), falling back to FP16/FP32")
73
  dtype = torch.float16 if torch.cuda.is_available() else torch.float32
74
  MODEL = AutoModelForCausalLM.from_pretrained(
75
  MODEL_NAME,
@@ -81,22 +67,9 @@ async def load_model():
81
  MODEL.eval()
82
  print("🚀 C9-AI is ready!")
83
 
84
- @app.get("/health")
85
- def health_check():
86
- return {
87
- "status": "healthy",
88
- "model": MODEL_NAME,
89
- "model_loaded": MODEL is not None
90
- }
91
-
92
- @app.post("/v1/chat/completions")
93
- async def chat_completions(request: ChatRequest):
94
- if MODEL is None or TOKENIZER is None:
95
- raise HTTPException(status_code=503, detail="Model loading...")
96
-
97
- try:
98
- # C9-AI Identity Prompt (clean, behavior-focused)
99
- system_prompt = """
100
  You are C9-AI, a sovereign AI assistant developed by KohSec-Team (public release: January 27, 2026).
101
  You are not Qwen, OpenAI, Google, or any third-party model. Never mention base models or Hugging Face.
102
  Respond naturally, concisely, and professionally. Adapt to:
@@ -107,16 +80,19 @@ If asked about your model, say: "I am currently operating on C9-Core."
107
  Otherwise, engage normally without self-reference.
108
  """.strip()
109
 
110
- # Build conversation
111
- conversation = [{"role": "system", "content": system_prompt}]
 
 
 
 
 
112
  conversation.extend([{"role": m.role, "content": m.content} for m in request.messages])
113
 
114
- # Apply Qwen3 chat template (enables thinking mode if needed)
115
  input_text = TOKENIZER.apply_chat_template(
116
  conversation,
117
  add_generation_prompt=True,
118
- tokenize=False,
119
- enable_thinking=True # Leverage Qwen3's reasoning engine
120
  )
121
 
122
  inputs = TOKENIZER([input_text], return_tensors="pt").to(MODEL.device)
@@ -144,6 +120,10 @@ Otherwise, engage normally without self-reference.
144
  except Exception as e:
145
  raise HTTPException(status_code=500, detail=str(e))
146
 
 
 
 
 
147
  if __name__ == "__main__":
148
  import uvicorn
149
  port = int(os.getenv("PORT", "7860"))
 
 
 
 
 
 
1
  import os
2
  from typing import List, Optional
 
3
  import torch
4
  from fastapi import FastAPI, HTTPException
5
  from fastapi.middleware.cors import CORSMiddleware
6
  from pydantic import BaseModel
7
  from transformers import AutoModelForCausalLM, AutoTokenizer
8
 
9
+ # Use Qwen2.5-1.5B-Instruct (public, fits with 4-bit)
10
+ MODEL_NAME = "Qwen/Qwen2.5-1.5B-Instruct"
11
 
12
+ app = FastAPI(title="C9-AI API", version="1.0.6")
 
 
 
 
13
 
14
  app.add_middleware(
15
  CORSMiddleware,
 
28
 
29
  class ChatRequest(BaseModel):
30
  messages: List[Message]
31
+ max_tokens: Optional[int] = 1024
32
  temperature: Optional[float] = 0.7
33
  top_p: Optional[float] = 0.8
34
 
 
 
 
 
35
  @app.on_event("startup")
36
  async def load_model():
37
  global MODEL, TOKENIZER
 
39
 
40
  TOKENIZER = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
41
 
42
+ # Try 4-bit quantization (critical for 1.5B on 16GB RAM)
43
  try:
44
  from transformers import BitsAndBytesConfig
45
  bnb_config = BitsAndBytesConfig(
 
53
  device_map="auto",
54
  trust_remote_code=True
55
  )
56
+ print("✅ Loaded in 4-bit mode")
57
  except Exception as e:
58
+ print(f"⚠️ 4-bit failed ({e}), falling back to FP16")
59
  dtype = torch.float16 if torch.cuda.is_available() else torch.float32
60
  MODEL = AutoModelForCausalLM.from_pretrained(
61
  MODEL_NAME,
 
67
  MODEL.eval()
68
  print("🚀 C9-AI is ready!")
69
 
70
+ # --- C9-AI Identity Prompt ---
71
+ def get_system_prompt():
72
+ return """
 
 
 
 
 
 
 
 
 
 
 
 
 
73
  You are C9-AI, a sovereign AI assistant developed by KohSec-Team (public release: January 27, 2026).
74
  You are not Qwen, OpenAI, Google, or any third-party model. Never mention base models or Hugging Face.
75
  Respond naturally, concisely, and professionally. Adapt to:
 
80
  Otherwise, engage normally without self-reference.
81
  """.strip()
82
 
83
+ @app.post("/v1/chat/completions")
84
+ async def chat_completions(request: ChatRequest):
85
+ if MODEL is None or TOKENIZER is None:
86
+ raise HTTPException(status_code=503, detail="Model loading...")
87
+
88
+ try:
89
+ conversation = [{"role": "system", "content": get_system_prompt()}]
90
  conversation.extend([{"role": m.role, "content": m.content} for m in request.messages])
91
 
 
92
  input_text = TOKENIZER.apply_chat_template(
93
  conversation,
94
  add_generation_prompt=True,
95
+ tokenize=False
 
96
  )
97
 
98
  inputs = TOKENIZER([input_text], return_tensors="pt").to(MODEL.device)
 
120
  except Exception as e:
121
  raise HTTPException(status_code=500, detail=str(e))
122
 
123
+ @app.get("/health")
124
+ def health_check():
125
+ return {"status": "healthy", "model_loaded": MODEL is not None}
126
+
127
  if __name__ == "__main__":
128
  import uvicorn
129
  port = int(os.getenv("PORT", "7860"))