File size: 4,031 Bytes
ac3772f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 |
import os
# ✅ Hugging Face 캐시/토큰 경로를 쓰기 가능한 위치로 지정 (Spaces에서는 /data가 안전)
os.environ["HF_HOME"] = "/data"
os.environ["TRANSFORMERS_CACHE"] = "/data/transformers"
os.environ["HF_HUB_CACHE"] = "/data/hub"
os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
os.environ.setdefault("PYTORCH_FORCE_MPS_FALLBACK", "1")
import threading
from typing import List, Optional, Dict, Any, Iterator
import torch
from fastapi import FastAPI, Body
from fastapi.responses import StreamingResponse, JSONResponse
from pydantic import BaseModel, Field
from transformers import (
AutoTokenizer,
AutoModelForCausalLM,
TextIteratorStreamer,
)
MODEL_ID = "unsloth/Qwen2.5-1.5B-Instruct" # :contentReference[oaicite:3]{index=3}
try:
torch.set_num_threads(max(1, os.cpu_count() or 1))
except Exception:
pass
# ---- 전역 모델/토크나이저 로드 ----
print(f"[BOOT] Loading {MODEL_ID} on CPU(float32)...")
tokenizer = AutoTokenizer.from_pretrained(
MODEL_ID,
use_fast=False,
trust_remote_code=True,
)
model = AutoModelForCausalLM.from_pretrained(
MODEL_ID,
torch_dtype=torch.float32,
device_map="cpu",
low_cpu_mem_usage=True, # accelerate 필요 (요건 requirements에 반영됨) :contentReference[oaicite:4]{index=4}
trust_remote_code=True,
)
model.eval()
# ---- API 모델 ----
class ChatMessage(BaseModel):
role: str = Field(..., description="system | user | assistant")
content: str
class ChatRequest(BaseModel):
messages: List[ChatMessage]
max_new_tokens: int = 256
temperature: float = 0.7
top_p: float = 0.95
repetition_penalty: float = 1.1
class ChatResponse(BaseModel):
text: str
app = FastAPI(title="Qwen2.5-1.5B CPU API")
@app.get("/")
def health():
return {"status": "ok", "model": MODEL_ID}
def build_prompt(messages: List[Dict[str, str]]) -> str:
# Qwen 계열 권장: chat 템플릿 사용(업로드 스크립트와 동일 컨셉) :contentReference[oaicite:5]{index=5}
return tokenizer.apply_chat_template(
[{"role": m["role"], "content": m["content"]} for m in messages],
tokenize=False,
add_generation_prompt=True,
)
@app.post("/v1/chat", response_model=ChatResponse)
def chat(req: ChatRequest):
prompt = build_prompt([m.dict() for m in req.messages])
inputs = tokenizer(prompt, return_tensors="pt")
with torch.no_grad():
output_ids = model.generate(
**{k: v.to("cpu") for k, v in inputs.items()},
max_new_tokens=req.max_new_tokens,
do_sample=True,
temperature=req.temperature,
top_p=req.top_p,
repetition_penalty=req.repetition_penalty,
eos_token_id=tokenizer.eos_token_id,
pad_token_id=tokenizer.eos_token_id,
)
text = tokenizer.decode(output_ids[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
return ChatResponse(text=text)
def stream_generate(req: ChatRequest) -> Iterator[str]:
prompt = build_prompt([m.dict() for m in req.messages])
inputs = tokenizer(prompt, return_tensors="pt")
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
gen_kwargs = dict(
**{k: v.to("cpu") for k, v in inputs.items()},
max_new_tokens=req.max_new_tokens,
do_sample=True,
temperature=req.temperature,
top_p=req.top_p,
repetition_penalty=req.repetition_penalty,
eos_token_id=tokenizer.eos_token_id,
pad_token_id=tokenizer.eos_token_id,
streamer=streamer,
)
thread = threading.Thread(target=model.generate, kwargs=gen_kwargs)
thread.start()
# NDJSON(한 줄에 { "delta": "..." }) 형태로 전송
for token_text in streamer:
yield f'{{"delta": {token_text.__repr__()}}}\n'
@app.post("/v1/chat/stream")
def chat_stream(req: ChatRequest = Body(...)):
return StreamingResponse(
stream_generate(req),
media_type="application/x-ndjson",
)
|