Upload 2 files
Browse files- app.py +128 -540
- requirements.txt +6 -19
app.py
CHANGED
|
@@ -1,333 +1,176 @@
|
|
| 1 |
"""
|
| 2 |
-
Nanbeige4.1-3B
|
| 3 |
-
|
| 4 |
"""
|
| 5 |
|
| 6 |
import os
|
| 7 |
import json
|
| 8 |
-
import
|
| 9 |
-
from typing import AsyncGenerator, Dict,
|
| 10 |
from contextlib import asynccontextmanager
|
| 11 |
-
|
| 12 |
-
from fastapi
|
|
|
|
| 13 |
from fastapi.middleware.cors import CORSMiddleware
|
| 14 |
-
from pydantic import BaseModel
|
| 15 |
-
from transformers import
|
| 16 |
from threading import Thread
|
| 17 |
-
import
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
|
| 19 |
-
# Global model
|
| 20 |
model = None
|
| 21 |
tokenizer = None
|
| 22 |
|
| 23 |
-
# Model configuration
|
| 24 |
-
MODEL_ID = "Nanbeige/Nanbeige4.1-3B"
|
| 25 |
-
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
| 26 |
-
DEFAULT_MAX_TOKENS = 2048
|
| 27 |
-
DEFAULT_TEMPERATURE = 0.6
|
| 28 |
-
DEFAULT_TOP_P = 0.95
|
| 29 |
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
content: str = Field(..., description="Message content")
|
| 34 |
-
tool_calls: Optional[List[Dict]] = Field(None, description="Tool calls from assistant")
|
| 35 |
-
tool_call_id: Optional[str] = Field(None, description="Tool call ID for tool responses")
|
| 36 |
|
| 37 |
|
| 38 |
class ChatRequest(BaseModel):
|
| 39 |
-
messages: List[
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
top_p: float = Field(default=DEFAULT_TOP_P, ge=0.0, le=1.0)
|
| 45 |
-
stop: Optional[List[str]] = Field(default=None, description="Stop sequences")
|
| 46 |
-
|
| 47 |
|
| 48 |
-
class ChatResponse(BaseModel):
|
| 49 |
-
id: str
|
| 50 |
-
object: str = "chat.completion"
|
| 51 |
-
created: int
|
| 52 |
-
model: str
|
| 53 |
-
choices: List[Dict]
|
| 54 |
-
usage: Optional[Dict] = None
|
| 55 |
|
| 56 |
-
|
| 57 |
-
def
|
| 58 |
-
"""
|
| 59 |
global model, tokenizer
|
| 60 |
-
|
| 61 |
-
print(
|
| 62 |
-
|
| 63 |
-
tokenizer = AutoTokenizer.from_pretrained(
|
| 64 |
-
MODEL_ID,
|
| 65 |
-
trust_remote_code=True,
|
| 66 |
-
padding_side="left"
|
| 67 |
-
)
|
| 68 |
-
|
| 69 |
-
# Set pad token if not present
|
| 70 |
-
if tokenizer.pad_token is None:
|
| 71 |
-
tokenizer.pad_token = tokenizer.eos_token
|
| 72 |
-
|
| 73 |
model = AutoModelForCausalLM.from_pretrained(
|
| 74 |
-
|
| 75 |
-
torch_dtype=torch.float16 if DEVICE == "cuda" else torch.float32,
|
| 76 |
-
device_map="auto" if DEVICE == "cuda" else None,
|
| 77 |
trust_remote_code=True,
|
| 78 |
-
|
|
|
|
| 79 |
)
|
| 80 |
-
|
| 81 |
-
if
|
| 82 |
-
model = model.to(
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
print(f"Model loaded successfully on {DEVICE}")
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
@asynccontextmanager
|
| 89 |
-
async def lifespan(app: FastAPI):
|
| 90 |
-
"""Application lifespan manager."""
|
| 91 |
-
# Startup
|
| 92 |
-
load_model()
|
| 93 |
yield
|
| 94 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 95 |
|
| 96 |
|
| 97 |
app = FastAPI(
|
| 98 |
title="Nanbeige4.1-3B Inference API",
|
| 99 |
-
description="
|
| 100 |
version="1.0.0",
|
| 101 |
lifespan=lifespan
|
| 102 |
)
|
| 103 |
|
| 104 |
-
# CORS for local agent communication
|
| 105 |
app.add_middleware(
|
| 106 |
CORSMiddleware,
|
| 107 |
-
allow_origins=["*"],
|
| 108 |
allow_credentials=True,
|
| 109 |
allow_methods=["*"],
|
| 110 |
allow_headers=["*"],
|
| 111 |
)
|
| 112 |
|
| 113 |
|
| 114 |
-
def
|
| 115 |
-
"""Format messages
|
| 116 |
-
|
| 117 |
-
|
| 118 |
for msg in messages:
|
| 119 |
if msg.role == "system":
|
| 120 |
-
|
| 121 |
elif msg.role == "user":
|
| 122 |
-
|
| 123 |
elif msg.role == "assistant":
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
content = f"{content}\n<tool_calls>{tool_calls_str}</tool_calls>"
|
| 129 |
-
formatted_messages.append({"role": "assistant", "content": content})
|
| 130 |
-
elif msg.role == "tool":
|
| 131 |
-
formatted_messages.append({
|
| 132 |
-
"role": "tool",
|
| 133 |
-
"content": msg.content,
|
| 134 |
-
"tool_call_id": msg.tool_call_id
|
| 135 |
-
})
|
| 136 |
-
|
| 137 |
-
# Add tools to system message if provided
|
| 138 |
-
if tools:
|
| 139 |
-
tools_description = "\n\nAvailable tools:\n" + json.dumps(tools, indent=2)
|
| 140 |
-
if formatted_messages and formatted_messages[0]["role"] == "system":
|
| 141 |
-
formatted_messages[0]["content"] += tools_description
|
| 142 |
-
else:
|
| 143 |
-
formatted_messages.insert(0, {"role": "system", "content": tools_description})
|
| 144 |
-
|
| 145 |
-
# Apply chat template
|
| 146 |
-
prompt = tokenizer.apply_chat_template(
|
| 147 |
-
formatted_messages,
|
| 148 |
-
tokenize=False,
|
| 149 |
-
add_generation_prompt=True
|
| 150 |
-
)
|
| 151 |
-
|
| 152 |
-
return prompt
|
| 153 |
|
| 154 |
|
| 155 |
-
def
|
| 156 |
-
"""
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
start = response_text.find("<tool_calls>") + len("<tool_calls>")
|
| 164 |
-
end = response_text.find("</tool_calls>")
|
| 165 |
-
tool_calls_json = response_text[start:end]
|
| 166 |
-
tool_calls = json.loads(tool_calls_json)
|
| 167 |
-
content = response_text[:response_text.find("<tool_calls>")].strip()
|
| 168 |
-
except (json.JSONDecodeError, ValueError):
|
| 169 |
-
pass
|
| 170 |
-
|
| 171 |
-
return content, tool_calls
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
def generate_stream(
|
| 175 |
-
prompt: str,
|
| 176 |
-
max_tokens: int,
|
| 177 |
-
temperature: float,
|
| 178 |
-
top_p: float,
|
| 179 |
-
stop: Optional[List[str]]
|
| 180 |
-
) -> AsyncGenerator[str, None]:
|
| 181 |
-
"""Generate streaming response."""
|
| 182 |
-
inputs = tokenizer(prompt, return_tensors="pt", padding=True)
|
| 183 |
-
inputs = {k: v.to(model.device) for k, v in inputs.items()}
|
| 184 |
-
|
| 185 |
streamer = TextIteratorStreamer(
|
| 186 |
tokenizer,
|
| 187 |
skip_prompt=True,
|
| 188 |
skip_special_tokens=True
|
| 189 |
)
|
| 190 |
-
|
| 191 |
-
generation_kwargs =
|
| 192 |
-
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
| 197 |
-
|
| 198 |
-
|
| 199 |
-
|
| 200 |
-
"eos_token_id": tokenizer.eos_token_id,
|
| 201 |
-
}
|
| 202 |
-
|
| 203 |
-
if stop:
|
| 204 |
-
generation_kwargs["stopping_criteria"] = create_stopping_criteria(stop)
|
| 205 |
-
|
| 206 |
# Run generation in separate thread
|
| 207 |
thread = Thread(target=model.generate, kwargs=generation_kwargs)
|
| 208 |
thread.start()
|
| 209 |
-
|
| 210 |
generated_text = ""
|
| 211 |
for new_text in streamer:
|
| 212 |
generated_text += new_text
|
| 213 |
-
#
|
| 214 |
-
|
| 215 |
-
|
| 216 |
-
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
|
| 220 |
-
yield new_text
|
| 221 |
-
|
| 222 |
thread.join()
|
| 223 |
|
| 224 |
|
| 225 |
-
|
| 226 |
-
|
| 227 |
-
|
| 228 |
-
|
| 229 |
-
|
| 230 |
-
|
| 231 |
-
|
| 232 |
-
|
| 233 |
-
|
| 234 |
-
def __call__(self, input_ids, scores, **kwargs):
|
| 235 |
-
generated = self.tokenizer.decode(input_ids[0], skip_special_tokens=True)
|
| 236 |
-
for stop in self.stops:
|
| 237 |
-
if stop in generated:
|
| 238 |
-
return True
|
| 239 |
-
return False
|
| 240 |
-
|
| 241 |
-
return StoppingCriteriaList([StopSequenceCriteria(stop_sequences, tokenizer)])
|
| 242 |
-
|
| 243 |
-
|
| 244 |
-
def generate_non_stream(
|
| 245 |
-
prompt: str,
|
| 246 |
-
max_tokens: int,
|
| 247 |
-
temperature: float,
|
| 248 |
-
top_p: float,
|
| 249 |
-
stop: Optional[List[str]]
|
| 250 |
-
) -> str:
|
| 251 |
-
"""Generate non-streaming response."""
|
| 252 |
-
inputs = tokenizer(prompt, return_tensors="pt", padding=True)
|
| 253 |
-
inputs = {k: v.to(model.device) for k, v in inputs.items()}
|
| 254 |
-
|
| 255 |
-
with torch.no_grad():
|
| 256 |
-
outputs = model.generate(
|
| 257 |
-
input_ids=inputs["input_ids"],
|
| 258 |
-
attention_mask=inputs["attention_mask"],
|
| 259 |
-
max_new_tokens=max_tokens,
|
| 260 |
-
temperature=temperature,
|
| 261 |
-
top_p=top_p,
|
| 262 |
-
do_sample=temperature > 0,
|
| 263 |
-
pad_token_id=tokenizer.pad_token_id,
|
| 264 |
-
eos_token_id=tokenizer.eos_token_id,
|
| 265 |
-
)
|
| 266 |
-
|
| 267 |
-
generated = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
|
| 268 |
|
| 269 |
-
# Apply stop sequences
|
| 270 |
-
if stop:
|
| 271 |
-
for s in stop:
|
| 272 |
-
if s in generated:
|
| 273 |
-
generated = generated[:generated.find(s)]
|
| 274 |
-
break
|
| 275 |
|
| 276 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 277 |
|
| 278 |
|
| 279 |
-
@app.post("/chat"
|
| 280 |
-
async def
|
| 281 |
"""
|
| 282 |
-
|
| 283 |
-
Compatible with OpenAI-style API for easy integration.
|
| 284 |
"""
|
| 285 |
-
|
| 286 |
-
|
| 287 |
-
|
| 288 |
-
|
|
|
|
|
|
|
| 289 |
if request.stream:
|
| 290 |
-
|
| 291 |
-
generated = ""
|
| 292 |
-
async for chunk in generate_stream(
|
| 293 |
-
prompt,
|
| 294 |
-
request.max_tokens,
|
| 295 |
-
request.temperature,
|
| 296 |
-
request.top_p,
|
| 297 |
-
request.stop
|
| 298 |
-
):
|
| 299 |
-
generated += chunk
|
| 300 |
-
data = {
|
| 301 |
-
"id": f"chatcmpl-{int(time.time())}",
|
| 302 |
-
"object": "chat.completion.chunk",
|
| 303 |
-
"created": int(time.time()),
|
| 304 |
-
"model": MODEL_ID,
|
| 305 |
-
"choices": [{
|
| 306 |
-
"index": 0,
|
| 307 |
-
"delta": {"content": chunk},
|
| 308 |
-
"finish_reason": None
|
| 309 |
-
}]
|
| 310 |
-
}
|
| 311 |
-
yield f"data: {json.dumps(data)}\n\n"
|
| 312 |
-
|
| 313 |
-
# Final chunk
|
| 314 |
-
content, tool_calls = parse_tool_calls(generated)
|
| 315 |
-
final_data = {
|
| 316 |
-
"id": f"chatcmpl-{int(time.time())}",
|
| 317 |
-
"object": "chat.completion.chunk",
|
| 318 |
-
"created": int(time.time()),
|
| 319 |
-
"model": MODEL_ID,
|
| 320 |
-
"choices": [{
|
| 321 |
-
"index": 0,
|
| 322 |
-
"delta": {},
|
| 323 |
-
"finish_reason": "stop"
|
| 324 |
-
}]
|
| 325 |
-
}
|
| 326 |
-
yield f"data: {json.dumps(final_data)}\n\n"
|
| 327 |
-
yield "data: [DONE]\n\n"
|
| 328 |
-
|
| 329 |
return StreamingResponse(
|
| 330 |
-
|
| 331 |
media_type="text/event-stream",
|
| 332 |
headers={
|
| 333 |
"Cache-Control": "no-cache",
|
|
@@ -335,288 +178,33 @@ async def chat_completion(request: ChatRequest):
|
|
| 335 |
"X-Accel-Buffering": "no"
|
| 336 |
}
|
| 337 |
)
|
| 338 |
-
|
| 339 |
else:
|
| 340 |
-
|
| 341 |
-
|
| 342 |
-
|
| 343 |
-
|
| 344 |
-
|
| 345 |
-
|
| 346 |
-
|
| 347 |
-
|
| 348 |
-
|
| 349 |
-
|
| 350 |
-
|
| 351 |
-
input_tokens = len(tokenizer.encode(prompt))
|
| 352 |
-
output_tokens = len(tokenizer.encode(generated))
|
| 353 |
-
|
| 354 |
-
response = ChatResponse(
|
| 355 |
-
id=f"chatcmpl-{int(time.time())}",
|
| 356 |
-
object="chat.completion",
|
| 357 |
-
created=int(time.time()),
|
| 358 |
-
model=MODEL_ID,
|
| 359 |
-
choices=[{
|
| 360 |
-
"index": 0,
|
| 361 |
-
"message": {
|
| 362 |
-
"role": "assistant",
|
| 363 |
-
"content": content,
|
| 364 |
-
"tool_calls": tool_calls
|
| 365 |
-
},
|
| 366 |
-
"finish_reason": "stop"
|
| 367 |
-
}],
|
| 368 |
-
usage={
|
| 369 |
-
"prompt_tokens": input_tokens,
|
| 370 |
-
"completion_tokens": output_tokens,
|
| 371 |
-
"total_tokens": input_tokens + output_tokens
|
| 372 |
-
}
|
| 373 |
)
|
| 374 |
-
|
| 375 |
-
|
| 376 |
-
|
| 377 |
-
|
| 378 |
-
|
| 379 |
-
|
| 380 |
-
|
| 381 |
-
|
| 382 |
-
|
| 383 |
-
|
| 384 |
-
<head>
|
| 385 |
-
<meta charset="UTF-8">
|
| 386 |
-
<meta name="viewport" content="width=device-width, initial-scale=1.0">
|
| 387 |
-
<title>Nanbeige4.1-3B Chat</title>
|
| 388 |
-
<style>
|
| 389 |
-
* { margin: 0; padding: 0; box-sizing: border-box; }
|
| 390 |
-
body {
|
| 391 |
-
font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif;
|
| 392 |
-
background: #1a1a2e;
|
| 393 |
-
color: #eee;
|
| 394 |
-
min-height: 100vh;
|
| 395 |
-
display: flex;
|
| 396 |
-
flex-direction: column;
|
| 397 |
-
}
|
| 398 |
-
header {
|
| 399 |
-
background: #16213e;
|
| 400 |
-
padding: 1rem 2rem;
|
| 401 |
-
border-bottom: 1px solid #0f3460;
|
| 402 |
-
}
|
| 403 |
-
header h1 { font-size: 1.25rem; color: #e94560; }
|
| 404 |
-
header p { font-size: 0.875rem; color: #888; margin-top: 0.25rem; }
|
| 405 |
-
.chat-container {
|
| 406 |
-
flex: 1;
|
| 407 |
-
display: flex;
|
| 408 |
-
flex-direction: column;
|
| 409 |
-
max-width: 900px;
|
| 410 |
-
width: 100%;
|
| 411 |
-
margin: 0 auto;
|
| 412 |
-
padding: 1rem;
|
| 413 |
-
}
|
| 414 |
-
.messages {
|
| 415 |
-
flex: 1;
|
| 416 |
-
overflow-y: auto;
|
| 417 |
-
padding: 1rem;
|
| 418 |
-
display: flex;
|
| 419 |
-
flex-direction: column;
|
| 420 |
-
gap: 1rem;
|
| 421 |
-
}
|
| 422 |
-
.message {
|
| 423 |
-
max-width: 80%;
|
| 424 |
-
padding: 1rem;
|
| 425 |
-
border-radius: 12px;
|
| 426 |
-
line-height: 1.6;
|
| 427 |
-
}
|
| 428 |
-
.message.user {
|
| 429 |
-
align-self: flex-end;
|
| 430 |
-
background: #e94560;
|
| 431 |
-
color: white;
|
| 432 |
-
}
|
| 433 |
-
.message.assistant {
|
| 434 |
-
align-self: flex-start;
|
| 435 |
-
background: #16213e;
|
| 436 |
-
border: 1px solid #0f3460;
|
| 437 |
-
}
|
| 438 |
-
.message.system {
|
| 439 |
-
align-self: center;
|
| 440 |
-
background: #0f3460;
|
| 441 |
-
font-size: 0.875rem;
|
| 442 |
-
color: #888;
|
| 443 |
-
}
|
| 444 |
-
.input-area {
|
| 445 |
-
display: flex;
|
| 446 |
-
gap: 0.5rem;
|
| 447 |
-
padding: 1rem;
|
| 448 |
-
background: #16213e;
|
| 449 |
-
border-top: 1px solid #0f3460;
|
| 450 |
-
}
|
| 451 |
-
textarea {
|
| 452 |
-
flex: 1;
|
| 453 |
-
padding: 0.75rem 1rem;
|
| 454 |
-
border: 1px solid #0f3460;
|
| 455 |
-
border-radius: 8px;
|
| 456 |
-
background: #1a1a2e;
|
| 457 |
-
color: #eee;
|
| 458 |
-
font-size: 1rem;
|
| 459 |
-
resize: none;
|
| 460 |
-
min-height: 50px;
|
| 461 |
-
max-height: 150px;
|
| 462 |
-
}
|
| 463 |
-
textarea:focus {
|
| 464 |
-
outline: none;
|
| 465 |
-
border-color: #e94560;
|
| 466 |
-
}
|
| 467 |
-
button {
|
| 468 |
-
padding: 0.75rem 1.5rem;
|
| 469 |
-
background: #e94560;
|
| 470 |
-
color: white;
|
| 471 |
-
border: none;
|
| 472 |
-
border-radius: 8px;
|
| 473 |
-
cursor: pointer;
|
| 474 |
-
font-size: 1rem;
|
| 475 |
-
transition: background 0.2s;
|
| 476 |
-
}
|
| 477 |
-
button:hover { background: #d63d56; }
|
| 478 |
-
button:disabled { background: #666; cursor: not-allowed; }
|
| 479 |
-
.loading {
|
| 480 |
-
display: inline-block;
|
| 481 |
-
width: 20px;
|
| 482 |
-
height: 20px;
|
| 483 |
-
border: 2px solid #0f3460;
|
| 484 |
-
border-top-color: #e94560;
|
| 485 |
-
border-radius: 50%;
|
| 486 |
-
animation: spin 1s linear infinite;
|
| 487 |
-
}
|
| 488 |
-
@keyframes spin { to { transform: rotate(360deg); } }
|
| 489 |
-
.tool-calls {
|
| 490 |
-
margin-top: 0.5rem;
|
| 491 |
-
padding: 0.5rem;
|
| 492 |
-
background: #0f3460;
|
| 493 |
-
border-radius: 6px;
|
| 494 |
-
font-size: 0.8rem;
|
| 495 |
-
font-family: monospace;
|
| 496 |
-
}
|
| 497 |
-
</style>
|
| 498 |
-
</head>
|
| 499 |
-
<body>
|
| 500 |
-
<header>
|
| 501 |
-
<h1>Nanbeige4.1-3B Inference Server</h1>
|
| 502 |
-
<p>Remote LLM service for Enterprise ReAct Agent</p>
|
| 503 |
-
</header>
|
| 504 |
-
<div class="chat-container">
|
| 505 |
-
<div class="messages" id="messages"></div>
|
| 506 |
-
<div class="input-area">
|
| 507 |
-
<textarea id="input" placeholder="Type your message..." rows="1"></textarea>
|
| 508 |
-
<button id="send" onclick="sendMessage()">Send</button>
|
| 509 |
-
</div>
|
| 510 |
-
</div>
|
| 511 |
-
|
| 512 |
-
<script>
|
| 513 |
-
const messages = document.getElementById('messages');
|
| 514 |
-
const input = document.getElementById('input');
|
| 515 |
-
const sendBtn = document.getElementById('send');
|
| 516 |
-
let conversation = [];
|
| 517 |
-
|
| 518 |
-
// Auto-resize textarea
|
| 519 |
-
input.addEventListener('input', () => {
|
| 520 |
-
input.style.height = 'auto';
|
| 521 |
-
input.style.height = Math.min(input.scrollHeight, 150) + 'px';
|
| 522 |
-
});
|
| 523 |
-
|
| 524 |
-
// Enter to send, Shift+Enter for new line
|
| 525 |
-
input.addEventListener('keydown', (e) => {
|
| 526 |
-
if (e.key === 'Enter' && !e.shiftKey) {
|
| 527 |
-
e.preventDefault();
|
| 528 |
-
sendMessage();
|
| 529 |
-
}
|
| 530 |
-
});
|
| 531 |
-
|
| 532 |
-
function addMessage(role, content, toolCalls = null) {
|
| 533 |
-
const div = document.createElement('div');
|
| 534 |
-
div.className = `message ${role}`;
|
| 535 |
-
div.textContent = content;
|
| 536 |
-
if (toolCalls) {
|
| 537 |
-
const toolDiv = document.createElement('div');
|
| 538 |
-
toolDiv.className = 'tool-calls';
|
| 539 |
-
toolDiv.textContent = 'Tool calls: ' + JSON.stringify(toolCalls, null, 2);
|
| 540 |
-
div.appendChild(toolDiv);
|
| 541 |
}
|
| 542 |
-
messages.appendChild(div);
|
| 543 |
-
messages.scrollTop = messages.scrollHeight;
|
| 544 |
}
|
| 545 |
|
| 546 |
-
async function sendMessage() {
|
| 547 |
-
const text = input.value.trim();
|
| 548 |
-
if (!text) return;
|
| 549 |
-
|
| 550 |
-
addMessage('user', text);
|
| 551 |
-
conversation.push({ role: 'user', content: text });
|
| 552 |
-
input.value = '';
|
| 553 |
-
input.style.height = 'auto';
|
| 554 |
-
sendBtn.disabled = true;
|
| 555 |
-
sendBtn.innerHTML = '<span class="loading"></span>';
|
| 556 |
-
|
| 557 |
-
try {
|
| 558 |
-
const response = await fetch('/chat', {
|
| 559 |
-
method: 'POST',
|
| 560 |
-
headers: { 'Content-Type': 'application/json' },
|
| 561 |
-
body: JSON.stringify({
|
| 562 |
-
messages: conversation,
|
| 563 |
-
stream: false,
|
| 564 |
-
max_tokens: 2048,
|
| 565 |
-
temperature: 0.6
|
| 566 |
-
})
|
| 567 |
-
});
|
| 568 |
-
|
| 569 |
-
const data = await response.json();
|
| 570 |
-
const assistantMsg = data.choices[0].message;
|
| 571 |
-
|
| 572 |
-
addMessage('assistant', assistantMsg.content, assistantMsg.tool_calls);
|
| 573 |
-
conversation.push({
|
| 574 |
-
role: 'assistant',
|
| 575 |
-
content: assistantMsg.content,
|
| 576 |
-
tool_calls: assistantMsg.tool_calls
|
| 577 |
-
});
|
| 578 |
-
} catch (error) {
|
| 579 |
-
addMessage('system', 'Error: ' + error.message);
|
| 580 |
-
} finally {
|
| 581 |
-
sendBtn.disabled = false;
|
| 582 |
-
sendBtn.textContent = 'Send';
|
| 583 |
-
}
|
| 584 |
-
}
|
| 585 |
-
|
| 586 |
-
// Initial system message
|
| 587 |
-
addMessage('system', 'Welcome! The model is ready for inference.');
|
| 588 |
-
</script>
|
| 589 |
-
</body>
|
| 590 |
-
</html>
|
| 591 |
-
"""
|
| 592 |
-
|
| 593 |
-
|
| 594 |
-
@app.get("/health")
|
| 595 |
-
async def health_check():
|
| 596 |
-
"""Health check endpoint."""
|
| 597 |
-
return {
|
| 598 |
-
"status": "healthy",
|
| 599 |
-
"model": MODEL_ID,
|
| 600 |
-
"device": DEVICE,
|
| 601 |
-
"model_loaded": model is not None and tokenizer is not None
|
| 602 |
-
}
|
| 603 |
-
|
| 604 |
-
|
| 605 |
-
@app.get("/")
|
| 606 |
-
async def root():
|
| 607 |
-
"""Root endpoint - redirect to chat interface."""
|
| 608 |
-
return {
|
| 609 |
-
"message": "Nanbeige4.1-3B Inference Server",
|
| 610 |
-
"endpoints": {
|
| 611 |
-
"chat": "/chat (POST for API, GET for web interface)",
|
| 612 |
-
"health": "/health"
|
| 613 |
-
},
|
| 614 |
-
"model": MODEL_ID,
|
| 615 |
-
"device": DEVICE
|
| 616 |
-
}
|
| 617 |
-
|
| 618 |
|
| 619 |
if __name__ == "__main__":
|
| 620 |
import uvicorn
|
| 621 |
-
|
| 622 |
-
uvicorn.run(app, host="0.0.0.0", port=port)
|
|
|
|
| 1 |
"""
|
| 2 |
+
HuggingFace Space application for Nanbeige4.1-3B model inference.
|
| 3 |
+
Provides streaming chat completion API.
|
| 4 |
"""
|
| 5 |
|
| 6 |
import os
|
| 7 |
import json
|
| 8 |
+
import asyncio
|
| 9 |
+
from typing import AsyncGenerator, List, Dict, Any, Optional
|
| 10 |
from contextlib import asynccontextmanager
|
| 11 |
+
|
| 12 |
+
from fastapi import FastAPI, HTTPException
|
| 13 |
+
from fastapi.responses import StreamingResponse
|
| 14 |
from fastapi.middleware.cors import CORSMiddleware
|
| 15 |
+
from pydantic import BaseModel
|
| 16 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer
|
| 17 |
from threading import Thread
|
| 18 |
+
import torch
|
| 19 |
+
|
| 20 |
+
# Model configuration
|
| 21 |
+
MODEL_NAME = "Nanbeige/Nanbeige4.1-3B"
|
| 22 |
+
MAX_LENGTH = 8192
|
| 23 |
|
| 24 |
+
# Global model and tokenizer
|
| 25 |
model = None
|
| 26 |
tokenizer = None
|
| 27 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
|
| 29 |
+
class Message(BaseModel):
|
| 30 |
+
role: str
|
| 31 |
+
content: str
|
|
|
|
|
|
|
|
|
|
| 32 |
|
| 33 |
|
| 34 |
class ChatRequest(BaseModel):
|
| 35 |
+
messages: List[Message]
|
| 36 |
+
stream: bool = True
|
| 37 |
+
max_tokens: int = 2048
|
| 38 |
+
temperature: float = 0.6
|
| 39 |
+
tools: Optional[List[Dict]] = None
|
|
|
|
|
|
|
|
|
|
| 40 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 41 |
|
| 42 |
+
@asynccontextmanager
|
| 43 |
+
async def lifespan(app: FastAPI):
|
| 44 |
+
"""Application lifespan handler."""
|
| 45 |
global model, tokenizer
|
| 46 |
+
|
| 47 |
+
print("Loading model...")
|
| 48 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 49 |
model = AutoModelForCausalLM.from_pretrained(
|
| 50 |
+
MODEL_NAME,
|
|
|
|
|
|
|
| 51 |
trust_remote_code=True,
|
| 52 |
+
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
| 53 |
+
device_map="auto" if torch.cuda.is_available() else None
|
| 54 |
)
|
| 55 |
+
|
| 56 |
+
if not torch.cuda.is_available():
|
| 57 |
+
model = model.to("cpu")
|
| 58 |
+
|
| 59 |
+
print("Model loaded successfully!")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 60 |
yield
|
| 61 |
+
|
| 62 |
+
# Cleanup
|
| 63 |
+
del model
|
| 64 |
+
del tokenizer
|
| 65 |
+
torch.cuda.empty_cache()
|
| 66 |
|
| 67 |
|
| 68 |
app = FastAPI(
|
| 69 |
title="Nanbeige4.1-3B Inference API",
|
| 70 |
+
description="Streaming chat completion API for Nanbeige4.1-3B",
|
| 71 |
version="1.0.0",
|
| 72 |
lifespan=lifespan
|
| 73 |
)
|
| 74 |
|
|
|
|
| 75 |
app.add_middleware(
|
| 76 |
CORSMiddleware,
|
| 77 |
+
allow_origins=["*"],
|
| 78 |
allow_credentials=True,
|
| 79 |
allow_methods=["*"],
|
| 80 |
allow_headers=["*"],
|
| 81 |
)
|
| 82 |
|
| 83 |
|
| 84 |
+
def format_messages(messages: List[Message]) -> str:
|
| 85 |
+
"""Format messages into prompt string."""
|
| 86 |
+
formatted = []
|
|
|
|
| 87 |
for msg in messages:
|
| 88 |
if msg.role == "system":
|
| 89 |
+
formatted.append(f"System: {msg.content}")
|
| 90 |
elif msg.role == "user":
|
| 91 |
+
formatted.append(f"User: {msg.content}")
|
| 92 |
elif msg.role == "assistant":
|
| 93 |
+
formatted.append(f"Assistant: {msg.content}")
|
| 94 |
+
|
| 95 |
+
formatted.append("Assistant:")
|
| 96 |
+
return "\n\n".join(formatted)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 97 |
|
| 98 |
|
| 99 |
+
async def stream_tokens(prompt: str, max_tokens: int, temperature: float) -> AsyncGenerator[str, None]:
|
| 100 |
+
"""Stream tokens from the model."""
|
| 101 |
+
global model, tokenizer
|
| 102 |
+
|
| 103 |
+
inputs = tokenizer(prompt, return_tensors="pt")
|
| 104 |
+
if torch.cuda.is_available():
|
| 105 |
+
inputs = inputs.to("cuda")
|
| 106 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 107 |
streamer = TextIteratorStreamer(
|
| 108 |
tokenizer,
|
| 109 |
skip_prompt=True,
|
| 110 |
skip_special_tokens=True
|
| 111 |
)
|
| 112 |
+
|
| 113 |
+
generation_kwargs = dict(
|
| 114 |
+
inputs,
|
| 115 |
+
streamer=streamer,
|
| 116 |
+
max_new_tokens=max_tokens,
|
| 117 |
+
temperature=temperature,
|
| 118 |
+
do_sample=temperature > 0,
|
| 119 |
+
pad_token_id=tokenizer.eos_token_id
|
| 120 |
+
)
|
| 121 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 122 |
# Run generation in separate thread
|
| 123 |
thread = Thread(target=model.generate, kwargs=generation_kwargs)
|
| 124 |
thread.start()
|
| 125 |
+
|
| 126 |
generated_text = ""
|
| 127 |
for new_text in streamer:
|
| 128 |
generated_text += new_text
|
| 129 |
+
# Yield each token
|
| 130 |
+
data = json.dumps({"type": "token", "content": new_text})
|
| 131 |
+
yield f"data: {data}\n\n"
|
| 132 |
+
|
| 133 |
+
# Signal completion
|
| 134 |
+
yield f"data: {json.dumps({'type': 'done', 'content': ''})}\n\n"
|
| 135 |
+
|
|
|
|
|
|
|
| 136 |
thread.join()
|
| 137 |
|
| 138 |
|
| 139 |
+
@app.get("/")
|
| 140 |
+
async def root():
|
| 141 |
+
"""Root endpoint."""
|
| 142 |
+
return {
|
| 143 |
+
"name": "Nanbeige4.1-3B Inference API",
|
| 144 |
+
"version": "1.0.0",
|
| 145 |
+
"model": MODEL_NAME,
|
| 146 |
+
"status": "running"
|
| 147 |
+
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 148 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 149 |
|
| 150 |
+
@app.get("/health")
|
| 151 |
+
async def health():
|
| 152 |
+
"""Health check endpoint."""
|
| 153 |
+
return {
|
| 154 |
+
"status": "healthy",
|
| 155 |
+
"model_loaded": model is not None and tokenizer is not None
|
| 156 |
+
}
|
| 157 |
|
| 158 |
|
| 159 |
+
@app.post("/chat")
|
| 160 |
+
async def chat(request: ChatRequest):
|
| 161 |
"""
|
| 162 |
+
Chat completion endpoint with streaming support.
|
|
|
|
| 163 |
"""
|
| 164 |
+
if model is None or tokenizer is None:
|
| 165 |
+
raise HTTPException(status_code=503, detail="Model not loaded yet")
|
| 166 |
+
|
| 167 |
+
# Format messages into prompt
|
| 168 |
+
prompt = format_messages(request.messages)
|
| 169 |
+
|
| 170 |
if request.stream:
|
| 171 |
+
# Return streaming response
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 172 |
return StreamingResponse(
|
| 173 |
+
stream_tokens(prompt, request.max_tokens, request.temperature),
|
| 174 |
media_type="text/event-stream",
|
| 175 |
headers={
|
| 176 |
"Cache-Control": "no-cache",
|
|
|
|
| 178 |
"X-Accel-Buffering": "no"
|
| 179 |
}
|
| 180 |
)
|
|
|
|
| 181 |
else:
|
| 182 |
+
# Non-streaming response
|
| 183 |
+
inputs = tokenizer(prompt, return_tensors="pt")
|
| 184 |
+
if torch.cuda.is_available():
|
| 185 |
+
inputs = inputs.to("cuda")
|
| 186 |
+
|
| 187 |
+
outputs = model.generate(
|
| 188 |
+
**inputs,
|
| 189 |
+
max_new_tokens=request.max_tokens,
|
| 190 |
+
temperature=request.temperature,
|
| 191 |
+
do_sample=request.temperature > 0,
|
| 192 |
+
pad_token_id=tokenizer.eos_token_id
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 193 |
)
|
| 194 |
+
|
| 195 |
+
response_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 196 |
+
# Extract only the assistant's response
|
| 197 |
+
response_text = response_text[len(prompt):].strip()
|
| 198 |
+
|
| 199 |
+
return {
|
| 200 |
+
"content": response_text,
|
| 201 |
+
"usage": {
|
| 202 |
+
"prompt_tokens": inputs.input_ids.shape[1],
|
| 203 |
+
"completion_tokens": outputs.shape[1] - inputs.input_ids.shape[1]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 204 |
}
|
|
|
|
|
|
|
| 205 |
}
|
| 206 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 207 |
|
| 208 |
if __name__ == "__main__":
|
| 209 |
import uvicorn
|
| 210 |
+
uvicorn.run(app, host="0.0.0.0", port=7860)
|
|
|
requirements.txt
CHANGED
|
@@ -1,19 +1,6 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
pydantic>=2.9.0
|
| 8 |
-
|
| 9 |
-
# ML/Transformers
|
| 10 |
-
torch>=2.1.0
|
| 11 |
-
transformers>=4.40.0
|
| 12 |
-
accelerate>=0.30.0
|
| 13 |
-
|
| 14 |
-
# Utilities
|
| 15 |
-
python-dotenv>=1.0.0
|
| 16 |
-
|
| 17 |
-
# Note: This configuration uses the original Nanbeige4.1-3B model
|
| 18 |
-
# from HuggingFace Hub (Nanbeige/Nanbeige4.1-3B)
|
| 19 |
-
# The model will be downloaded on first startup
|
|
|
|
| 1 |
+
fastapi==0.109.0
|
| 2 |
+
uvicorn[standard]==0.27.0
|
| 3 |
+
transformers==4.37.0
|
| 4 |
+
torch==2.1.2
|
| 5 |
+
accelerate==0.26.0
|
| 6 |
+
pydantic==2.5.3
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|