app.py

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

# Your adapter (LoRA fine-tuned model on Hugging Face)
ADAPTER_ID = "Anabury/My_Finetuned_Phi-4"

# Detect device
USE_GPU = torch.cuda.is_available()

# Pick base model depending on device
if USE_GPU:
    BASE_MODEL = "unsloth/phi-4-unsloth-bnb-4bit"   # fast + quantized
else:
    BASE_MODEL = "unsloth/phi-4"                   # full precision for CPU

print(f"Loading base model: {BASE_MODEL} on {'GPU' if USE_GPU else 'CPU'}")

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)

# Load base model
base = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    device_map="auto" if USE_GPU else None,
    torch_dtype=torch.float16 if USE_GPU else torch.float32,
    trust_remote_code=True
)

# Attach your LoRA adapter
model = PeftModel.from_pretrained(base, ADAPTER_ID)
model.eval()

# Chat function
def chat(message, history):
    # simple prompt, you can swap in chat template later
    inputs = tokenizer(message, return_tensors="pt").to(model.device)
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=256,
            do_sample=True,
            temperature=0.7,
            top_p=0.9,
            pad_token_id=tokenizer.eos_token_id,
        )
    reply = tokenizer.decode(outputs[0], skip_special_tokens=True)
    history.append((message, reply))
    return history, history

# Gradio UI
with gr.Blocks() as demo:
    gr.Markdown("# 🧠 Phi-4 Chatbot (Fine-tuned)")
    chatbot = gr.Chatbot(height=420)
    msg = gr.Textbox(placeholder="Ask me anything…")
    clear = gr.Button("Clear")

    msg.submit(chat, [msg, chatbot], [chatbot, chatbot])
    clear.click(lambda: [], None, chatbot, queue=False)

demo.launch()