app.py

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import gradio as gr

# --- Models Load (CPU ke liye optimized) ---
BASE_MODEL = "Qwen/Qwen2.5-1.5B"
LORA_ADAPTER = "modular-ai/qwen"

print("Loading base model on CPU... (ye 1-2 min lagega pehli baar)")

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    torch_dtype=torch.float32,      # CPU pe float16 nahi chalta
    device_map="cpu",               # Sirf CPU
    trust_remote_code=True,
    low_cpu_mem_usage=True          # Memory bachaye
)

print("Loading LoRA adapter...")
model = PeftModel.from_pretrained(base_model, LORA_ADAPTER)

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# --- Chat Function (Fast & Safe) ---
def ask_kant(message, history):
    prompt = f"### Instruction: You are Immanuel Kant.\n\n### Input: {message}\n\n### Response:"
    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512)
    
    with torch.no_grad():
        output = model.generate(
            **inputs,
            max_new_tokens=200,
            temperature=0.7,
            do_sample=True,
            top_p=0.9,
            repetition_penalty=1.1,
            pad_token_id=tokenizer.eos_token_id
        )
    
    response = tokenizer.decode(output[0], skip_special_tokens=True)
    bot_reply = response.split("### Response:")[-1].strip()
    return bot_reply

# --- Gradio UI (Simple & Fast) ---
with gr.Blocks() as demo:
    gr.Markdown("# 🧠 **Kant AI** – Qwen2.5-1.5B LoRA")
    gr.Markdown("**Zero GPU | Free | Live Demo**  \nPoochein koi bhi sawal, *Immanuel Kant* jawab denge!")

    chatbot = gr.ChatInterface(
        fn=ask_kant,
        title="",
        examples=[
            "What is freedom?",
            "Kya hai swatantrata?",
            "Explain categorical imperative",
            "Moral law kya hai?"
        ],
        cache_examples=False,
        submit_btn="Ask Kant",
        retry_btn=None,
        clear_btn="Clear"
    )
    
    gr.Markdown("---\n*Model: Qwen2.5-1.5B + LoRA | CPU Only | ~8-12 sec per reply*")

demo.launch()