app.py

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

# Model and device setup
base_model_name = "unsloth/gemma-3-12b-it-unsloth-bnb-4bit"
adapter_name = "adarsh3601/my_gemma3_pt"
device = "cuda" if torch.cuda.is_available() else "cpu"

# Load base model with 4-bit quantization
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    device_map={"": device},
    torch_dtype=torch.float16,  # Keep float16 unless it breaks
    load_in_4bit=True
)

# Load tokenizer and adapter
tokenizer = AutoTokenizer.from_pretrained(base_model_name)
model = PeftModel.from_pretrained(base_model, adapter_name)
model.to(device)

# Chat function with stability safeguards
def chat(message):
    if not message or not message.strip():
        return "Please enter a valid message."

    inputs = tokenizer(message, return_tensors="pt")

    # Safely move to device; only convert float tensors to half
    for k in inputs:
        if inputs[k].dtype == torch.float32:
            inputs[k] = inputs[k].to(device).half()
        else:
            inputs[k] = inputs[k].to(device)

    try:
        with torch.no_grad():
            outputs = model.generate(
                **inputs,
                max_new_tokens=150,
                do_sample=True,
                top_k=50,
                top_p=0.95,
                temperature=0.8
            )

        response = tokenizer.decode(outputs[0], skip_special_tokens=True)
        return response

    except RuntimeError as e:
        return f"An error occurred during generation: {str(e)}"

# Launch Gradio app
iface = gr.Interface(
    fn=chat,
    inputs="text",
    outputs="text",
    title="Gemma Chatbot"
)

iface.launch()