app.py

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

# Model loading
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
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    device_map={"": device},
    torch_dtype=torch.float16,
    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
def chat(message):
    inputs = tokenizer(message, return_tensors="pt")
    
    # Move tensors to the correct device and convert only 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)

    outputs = model.generate(**inputs, max_new_tokens=150, do_sample=True)
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return response

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