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
import os
import asyncio

# Set cache directories
cache_dir = "/tmp/hf_home"
os.environ["HF_HOME"] = cache_dir
os.environ["TRANSFORMERS_CACHE"] = cache_dir
os.environ["HUGGINGFACE_HUB_CACHE"] = cache_dir

# Create cache directory with proper permissions
os.makedirs(cache_dir, exist_ok=True)
os.chmod(cache_dir, 0o777)

# Load model and tokenizer
model_name = "Qwen/Qwen2.5-0.5B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(
    model_name,
    trust_remote_code=True,
    cache_dir=cache_dir
)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    trust_remote_code=True,
    cache_dir=cache_dir,
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
)

# Set device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

# Initialize FastAPI
app = FastAPI()

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

# Input model
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."

async def generate_response_chunks(prompt: str):
    # Create the chat template
    messages = [
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": prompt}
    ]
    
    # Apply chat template
    qwen_prompt = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )
    
    # Tokenize and generate
    inputs = tokenizer(qwen_prompt, return_tensors="pt").to(device)
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        do_sample=True,
        temperature=0.7,
        top_p=0.9,
        pad_token_id=tokenizer.eos_token_id
    )
    
    # Decode and clean the output
    full_output = tokenizer.decode(outputs[0], skip_special_tokens=False)
    
    # Extract only the assistant's response
    response = full_output[len(qwen_prompt):].split(tokenizer.eos_token)[0].strip()
    
    # Stream the response
    for word in response.split():
        yield word + " "
        await asyncio.sleep(0.05)

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