app/main.py

from fastapi import FastAPI, Request, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import os

# === CONFIG ===
HF_MODEL = "Qwen/Qwen2.5-0.5B-Instruct"
ADAPTER_PATH = "adapter"  # folder where your LoRA is saved
API_KEY = os.getenv("API_KEY", "your-secret-key")  # Set in HF Space secrets

# === FastAPI Setup ===
app = FastAPI()

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],  # adjust if needed
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

# === Load Model & Tokenizer (CPU only) ===
print("🔧 Loading model on CPU...")
tokenizer = AutoTokenizer.from_pretrained(HF_MODEL, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(HF_MODEL, torch_dtype=torch.float32, trust_remote_code=True)
model = PeftModel.from_pretrained(model, ADAPTER_PATH)
model = model.to("cpu")
model.eval()
print("✅ Model ready on CPU.")

# === Request Schema ===
class ChatRequest(BaseModel):
    prompt: str
    api_key: str

@app.get("/")
def root():
    return {"message": "✅ Qwen2.5 Chat API running."}

@app.post("/chat")
def chat(req: ChatRequest):
    if req.api_key != API_KEY:
        raise HTTPException(status_code=401, detail="Invalid API Key")

    input_text = f"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n{req.prompt}<|im_end|>\n<|im_start|>assistant\n"

    inputs = tokenizer(input_text, return_tensors="pt").to("cpu")
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.7,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)

    # Extract response after assistant tag
    final_resp = response.split("<|im_start|>assistant\n")[-1].strip()
    return {"response": final_resp}