app.py

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

BASE_MODEL = "unsloth/phi-4-unsloth-bnb-4bit"   # base that you finetuned from
ADAPTER_ID = "Anabury/My_Finetuned_Phi-4"       # your adapter repo

# tokenizer (either base or adapter works; use base)
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)

# load base model (4-bit quant is fine on Spaces GPU/CPU)
base = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    device_map="auto",
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    trust_remote_code=True
)

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

def chat(message, history):
    # build a simple prompt; adapt if you have a chat template in your repo
    prompt = message
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    with torch.no_grad():
        output = 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(output[0], skip_special_tokens=True)
    history.append((message, reply))
    return history, history

with gr.Blocks() as demo:
    gr.Markdown("# Phi-4 Chat (LoRA)")
    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()