app.py

from fastapi import FastAPI, Form
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

app = FastAPI()

MODEL_NAME = "meituan/DeepSeek-R1-Distill-Qwen-32B"  # example
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, torch_dtype=torch.float16)  
# Possibly load quantized or INT8 version

@app.post("/chat")
async def chat(prompt: str = Form(...)):
    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    outputs = model.generate(**inputs, max_new_tokens=200)
    reply = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return {"reply": reply}