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 in 4-bit with float16
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    device_map="auto",
    torch_dtype=torch.float16,
    load_in_4bit=True
)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model_name)

# Load fine-tuned adapter
model = PeftModel.from_pretrained(base_model, adapter_name)
model.to(device)

# Chat function
def chat(message):
    try:
        # Tokenize input (do NOT convert to .half())
        inputs = tokenizer(message, return_tensors="pt").to(device)
        
        # Generate output
        outputs = model.generate(
            **inputs,
            max_new_tokens=150,
            do_sample=True,
            temperature=0.7,
            top_p=0.95
        )

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

    except Exception as e:
        print("Unexpected error:", e)
        return "An error occurred during generation."

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