main.py

from fastapi import FastAPI
from pydantic import BaseModel
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import StreamingResponse
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load Qwen model and tokenizer (once)
model_name = "Qwen/Qwen2.5-0.5B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)

# Set device
device = torch.device("cpu")  # Or "cuda" if using GPU
model.to(device)

# FastAPI app
app = FastAPI()

# CORS settings
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

# Request body model
class Question(BaseModel):
    question: str

# System prompt (your custom instructions)
SYSTEM_PROMPT = "You are Orion, an intelligent AI assistant created by Abdullah Ali, a 13-year-old from Lahore. Respond kindly and wisely."

# Chat response generator
async def generate_response_chunks(prompt: str):
    # Build prompt using Qwen's expected format
    qwen_prompt = (
        f"<|im_start|>system\n{SYSTEM_PROMPT}<|im_end|>\n"
        f"<|im_start|>user\n{prompt}<|im_end|>\n"
        f"<|im_start|>assistant\n"
    )

    # Tokenize input
    inputs = tokenizer(qwen_prompt, return_tensors="pt").to(device)

    # Generate response
    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=True,
        temperature=0.7,
        top_p=0.9,
        pad_token_id=tokenizer.eos_token_id
    )

    # Decode and yield line by line
    full_output = tokenizer.decode(outputs[0], skip_special_tokens=True)
    reply = full_output.split("<|im_start|>assistant\n")[-1].strip()

    for chunk in reply.split():
        yield chunk + " "

@app.post("/ask")
async def ask(question: Question):
    return StreamingResponse(
        generate_response_chunks(question.question),
        media_type="text/plain"
    )