app.py

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

BASE_ID = "unsloth/Llama-3.2-3B-Instruct-bnb-4bit"
LORA_PATH = "./lora_model"

def load_model():
    tokenizer = AutoTokenizer.from_pretrained(BASE_ID, use_fast=True)

    # 4bit baseline 一般需要 GPU；Space 选 GPU 才建议这样跑
    base_model = AutoModelForCausalLM.from_pretrained(
        BASE_ID,
        device_map="auto",
        torch_dtype=torch.float16,
    )

    model = PeftModel.from_pretrained(base_model, LORA_PATH)
    model.eval()
    return tokenizer, model

tokenizer, model = load_model()

@torch.inference_mode()
def respond(prompt, max_new_tokens=256, temperature=0.7, top_p=0.9):
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    out = model.generate(
        **inputs,
        max_new_tokens=int(max_new_tokens),
        do_sample=True,
        temperature=float(temperature),
        top_p=float(top_p),
        eos_token_id=tokenizer.eos_token_id,
    )
    return tokenizer.decode(out[0], skip_special_tokens=True)

demo = gr.Interface(
    fn=respond,
    inputs=[
        gr.Textbox(lines=6, label="Prompt"),
        gr.Slider(1, 1024, value=256, step=1, label="max_new_tokens"),
        gr.Slider(0.0, 2.0, value=0.7, step=0.05, label="temperature"),
        gr.Slider(0.0, 1.0, value=0.9, step=0.01, label="top_p"),
    ],
    outputs=gr.Textbox(lines=12, label="Output"),
    title="Llama-3.2-1B (4bit) + LoRA Adapter",
)

if __name__ == "__main__":
    demo.launch()