Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,272 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from langchain.llms import LlamaCpp
|
3 |
+
import os
|
4 |
+
import json
|
5 |
+
import torch
|
6 |
+
import logging
|
7 |
+
from typing import Optional, List, Dict, Any
|
8 |
+
from fastapi import FastAPI, HTTPException, Request
|
9 |
+
from fastapi.responses import JSONResponse
|
10 |
+
from pydantic import BaseModel
|
11 |
+
import uvicorn
|
12 |
+
import time
|
13 |
+
from threading import Lock
|
14 |
+
|
15 |
+
# Configure logging
|
16 |
+
logging.basicConfig(level=logging.INFO)
|
17 |
+
logger = logging.getLogger(__name__)
|
18 |
+
|
19 |
+
class ChatCompletionRequest(BaseModel):
|
20 |
+
model: str
|
21 |
+
messages: List[Dict[str, str]]
|
22 |
+
temperature: Optional[float] = 0.7
|
23 |
+
max_tokens: Optional[int] = 2048
|
24 |
+
stream: Optional[bool] = False
|
25 |
+
|
26 |
+
class QwenModel:
|
27 |
+
def __init__(self, model_path: str):
|
28 |
+
"""Initialize the Qwen model with automatic device detection."""
|
29 |
+
try:
|
30 |
+
# Check for GPU availability
|
31 |
+
self.has_gpu = torch.cuda.is_available()
|
32 |
+
self.device_count = torch.cuda.device_count() if self.has_gpu else 0
|
33 |
+
logger.info(f"GPU available: {self.has_gpu}, Device count: {self.device_count}")
|
34 |
+
|
35 |
+
# Configure model parameters based on available hardware
|
36 |
+
n_gpu_layers = 40 if self.has_gpu else 0
|
37 |
+
logger.info(f"Using {'GPU' if self.has_gpu else 'CPU'} for inference")
|
38 |
+
|
39 |
+
self.llm = LlamaCpp(
|
40 |
+
model_path=model_path,
|
41 |
+
n_gpu_layers=n_gpu_layers,
|
42 |
+
n_ctx=4096,
|
43 |
+
n_batch=512 if self.has_gpu else 128, # Reduced batch size for CPU
|
44 |
+
verbose=True,
|
45 |
+
temperature=0.7,
|
46 |
+
max_tokens=2048,
|
47 |
+
top_p=0.95,
|
48 |
+
top_k=50,
|
49 |
+
f16_kv=self.has_gpu, # Only use f16 when GPU is available
|
50 |
+
use_mlock=True, # Pin memory for better performance
|
51 |
+
use_mmap=True,
|
52 |
+
)
|
53 |
+
|
54 |
+
# Thread lock for concurrent API requests
|
55 |
+
self.lock = Lock()
|
56 |
+
|
57 |
+
except Exception as e:
|
58 |
+
logger.error(f"Failed to initialize model: {str(e)}")
|
59 |
+
raise
|
60 |
+
|
61 |
+
def generate_cot_prompt(self, messages: List[Dict[str, str]]) -> str:
|
62 |
+
"""Generate a chain-of-thought prompt from message history."""
|
63 |
+
conversation = []
|
64 |
+
for msg in messages:
|
65 |
+
role = msg.get("role", "")
|
66 |
+
content = msg.get("content", "")
|
67 |
+
|
68 |
+
if role == "system":
|
69 |
+
conversation.append(f"System: {content}")
|
70 |
+
elif role == "user":
|
71 |
+
conversation.append(f"Human: {content}")
|
72 |
+
elif role == "assistant":
|
73 |
+
conversation.append(f"Assistant: {content}")
|
74 |
+
|
75 |
+
last_user_msg = next((msg["content"] for msg in reversed(messages)
|
76 |
+
if msg["role"] == "user"), None)
|
77 |
+
|
78 |
+
if not last_user_msg:
|
79 |
+
raise ValueError("No user message found in the conversation")
|
80 |
+
|
81 |
+
cot_template = f"""Previous conversation:
|
82 |
+
{chr(10).join(conversation)}
|
83 |
+
|
84 |
+
Let's approach the latest question step-by-step:
|
85 |
+
|
86 |
+
1. Understanding the question:
|
87 |
+
{last_user_msg}
|
88 |
+
|
89 |
+
2. Breaking down components:
|
90 |
+
- Key elements to consider
|
91 |
+
- Specific information requested
|
92 |
+
- Relevant constraints
|
93 |
+
|
94 |
+
3. Reasoning process:
|
95 |
+
- Systematic approach
|
96 |
+
- Applicable knowledge
|
97 |
+
- Potential challenges
|
98 |
+
|
99 |
+
4. Step-by-step solution:
|
100 |
+
|
101 |
+
"""
|
102 |
+
return cot_template
|
103 |
+
|
104 |
+
def process_response(self, response: str) -> str:
|
105 |
+
"""Process and format the model's response."""
|
106 |
+
try:
|
107 |
+
response = response.strip()
|
108 |
+
# Add structural markers for better readability
|
109 |
+
if not response.startswith("Step"):
|
110 |
+
response = "Step-by-step solution:\n" + response
|
111 |
+
return response
|
112 |
+
except Exception as e:
|
113 |
+
logger.error(f"Error processing response: {str(e)}")
|
114 |
+
return "Error processing response"
|
115 |
+
|
116 |
+
def generate_response(self,
|
117 |
+
messages: List[Dict[str, str]],
|
118 |
+
temperature: float = 0.7,
|
119 |
+
max_tokens: int = 2048) -> Dict[str, Any]:
|
120 |
+
"""Generate a response using chain-of-thought reasoning."""
|
121 |
+
try:
|
122 |
+
with self.lock: # Thread safety for concurrent API requests
|
123 |
+
# Generate the CoT prompt
|
124 |
+
full_prompt = self.generate_cot_prompt(messages)
|
125 |
+
|
126 |
+
# Get response from model
|
127 |
+
start_time = time.time()
|
128 |
+
response = self.llm(
|
129 |
+
full_prompt,
|
130 |
+
temperature=temperature,
|
131 |
+
max_tokens=max_tokens
|
132 |
+
)
|
133 |
+
end_time = time.time()
|
134 |
+
|
135 |
+
# Process response
|
136 |
+
processed_response = self.process_response(response)
|
137 |
+
|
138 |
+
# Format response in OpenAI-compatible structure
|
139 |
+
return {
|
140 |
+
"id": f"chatcmpl-{int(time.time()*1000)}",
|
141 |
+
"object": "chat.completion",
|
142 |
+
"created": int(time.time()),
|
143 |
+
"model": "qwen-2.5-14b",
|
144 |
+
"choices": [{
|
145 |
+
"index": 0,
|
146 |
+
"message": {
|
147 |
+
"role": "assistant",
|
148 |
+
"content": processed_response
|
149 |
+
},
|
150 |
+
"finish_reason": "stop"
|
151 |
+
}],
|
152 |
+
"usage": {
|
153 |
+
"prompt_tokens": len(full_prompt.split()),
|
154 |
+
"completion_tokens": len(processed_response.split()),
|
155 |
+
"total_tokens": len(full_prompt.split()) + len(processed_response.split())
|
156 |
+
},
|
157 |
+
"system_info": {
|
158 |
+
"device": "gpu" if self.has_gpu else "cpu",
|
159 |
+
"processing_time": round(end_time - start_time, 2)
|
160 |
+
}
|
161 |
+
}
|
162 |
+
except Exception as e:
|
163 |
+
logger.error(f"Error generating response: {str(e)}")
|
164 |
+
raise HTTPException(status_code=500, detail=str(e))
|
165 |
+
|
166 |
+
# Initialize FastAPI
|
167 |
+
app = FastAPI(title="Qwen 2.5 API")
|
168 |
+
|
169 |
+
def create_gradio_interface(model: QwenModel):
|
170 |
+
"""Create and configure the Gradio interface."""
|
171 |
+
|
172 |
+
def predict(message: str,
|
173 |
+
temperature: float,
|
174 |
+
max_tokens: int) -> str:
|
175 |
+
messages = [{"role": "user", "content": message}]
|
176 |
+
response = model.generate_response(
|
177 |
+
messages,
|
178 |
+
temperature=temperature,
|
179 |
+
max_tokens=max_tokens
|
180 |
+
)
|
181 |
+
return response["choices"][0]["message"]["content"]
|
182 |
+
|
183 |
+
iface = gr.Interface(
|
184 |
+
fn=predict,
|
185 |
+
inputs=[
|
186 |
+
gr.Textbox(
|
187 |
+
label="Input",
|
188 |
+
placeholder="Enter your question or task here...",
|
189 |
+
lines=5
|
190 |
+
),
|
191 |
+
gr.Slider(
|
192 |
+
minimum=0.1,
|
193 |
+
maximum=1.0,
|
194 |
+
value=0.7,
|
195 |
+
label="Temperature",
|
196 |
+
info="Higher values make the output more random"
|
197 |
+
),
|
198 |
+
gr.Slider(
|
199 |
+
minimum=64,
|
200 |
+
maximum=4096,
|
201 |
+
value=2048,
|
202 |
+
step=64,
|
203 |
+
label="Max Tokens",
|
204 |
+
info="Maximum length of the generated response"
|
205 |
+
)
|
206 |
+
],
|
207 |
+
outputs=gr.Textbox(label="Response", lines=10),
|
208 |
+
title=f"Qwen 2.5 14B Instruct Model ({'GPU' if model.has_gpu else 'CPU'} Mode)",
|
209 |
+
description="""This is a Qwen 2.5 14B model interface with chain-of-thought prompting.
|
210 |
+
The model will break down complex problems and solve them step by step.""",
|
211 |
+
examples=[
|
212 |
+
["Explain how photosynthesis works", 0.7, 2048],
|
213 |
+
["Solve the quadratic equation: x² + 5x + 6 = 0", 0.7, 1024],
|
214 |
+
["What are the implications of Moore's Law for future computing?", 0.8, 2048]
|
215 |
+
]
|
216 |
+
)
|
217 |
+
return iface
|
218 |
+
|
219 |
+
# Global model instance
|
220 |
+
model = None
|
221 |
+
|
222 |
+
@app.on_event("startup")
|
223 |
+
async def startup_event():
|
224 |
+
"""Initialize the model on startup."""
|
225 |
+
global model
|
226 |
+
model_path = "G17c21ds/Qwen2.5-14B-Instruct-Uncensored-Q8_0-GGUF"
|
227 |
+
model = QwenModel(model_path)
|
228 |
+
logger.info("Model initialized successfully")
|
229 |
+
|
230 |
+
@app.post("/v1/chat/completions")
|
231 |
+
async def create_chat_completion(request: ChatCompletionRequest):
|
232 |
+
"""OpenAI-compatible chat completions endpoint."""
|
233 |
+
try:
|
234 |
+
response = model.generate_response(
|
235 |
+
request.messages,
|
236 |
+
temperature=request.temperature,
|
237 |
+
max_tokens=request.max_tokens
|
238 |
+
)
|
239 |
+
return JSONResponse(content=response)
|
240 |
+
except Exception as e:
|
241 |
+
raise HTTPException(status_code=500, detail=str(e))
|
242 |
+
|
243 |
+
def main():
|
244 |
+
"""Main function to initialize and launch the application."""
|
245 |
+
try:
|
246 |
+
global model
|
247 |
+
# Model path
|
248 |
+
model_path = "G17c21ds/Qwen2.5-14B-Instruct-Uncensored-Q8_0-GGUF"
|
249 |
+
|
250 |
+
# Initialize the model if not already initialized
|
251 |
+
if model is None:
|
252 |
+
model = QwenModel(model_path)
|
253 |
+
|
254 |
+
# Create and launch the Gradio interface
|
255 |
+
interface = create_gradio_interface(model)
|
256 |
+
|
257 |
+
# Mount FastAPI app to Gradio
|
258 |
+
app.mount("/", interface.app)
|
259 |
+
|
260 |
+
# Launch with uvicorn
|
261 |
+
uvicorn.run(
|
262 |
+
app,
|
263 |
+
host="0.0.0.0",
|
264 |
+
port=7860,
|
265 |
+
log_level="info"
|
266 |
+
)
|
267 |
+
except Exception as e:
|
268 |
+
logger.error(f"Application failed to start: {str(e)}")
|
269 |
+
raise
|
270 |
+
|
271 |
+
if __name__ == "__main__":
|
272 |
+
main()
|