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
import os
import asyncio

# ✅ Set a safe and writable HF cache directory
os.environ["HF_HOME"] = "./hf_home"
os.makedirs(os.environ["HF_HOME"], exist_ok=True)

# ✅ Model and tokenizer (only loaded 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 (use GPU if available)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

# ✅ FastAPI app
app = FastAPI()

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

# ✅ Request schema
class Question(BaseModel):
    question: str

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

# ✅ Streaming generator
async def generate_response_chunks(prompt: str):
    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 prompt
    inputs = tokenizer(qwen_prompt, return_tensors="pt").to(device)

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

    # Yield chunks word by word (simulating stream)
    for word in reply.split():
        yield word + " "
        await asyncio.sleep(0.01)  # slight delay for streaming effect

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