app.py

import torch
from transformers import LlamaTokenizer, LlamaForCausalLM, GenerationConfig
import gradio as gr
import warnings
import os

# Remove command-line arguments parsing and use hardcoded defaults for simplicity in Spaces
MODEL_PATH = "/model/13B_hf"
LORA_PATH = "checkpoint-3000"
USE_TYPEWRITER = 1
USE_LOCAL = 1

tokenizer = LlamaTokenizer.from_pretrained(MODEL_PATH)

LOAD_8BIT = True
BASE_MODEL = MODEL_PATH
LORA_WEIGHTS = LORA_PATH

lora_bin_path = os.path.join(LORA_PATH, "adapter_model.bin")
if not os.path.exists(lora_bin_path) and USE_LOCAL:
    # ... [rest of the path fixing logic]

# ... [rest of the device and model loading logic]

def generate_prompt(instruction, input=None):
    # ... [rest of the generate_prompt function]

def evaluate(
    input,
    temperature=0.1,
    top_p=0.75,
    top_k=40,
    num_beams=4,
    max_new_tokens=128,
    min_new_tokens=1,
    repetition_penalty=2.0,
    **kwargs,
):
    # ... [rest of the evaluate function]

gr.Interface(
    fn=evaluate,
    inputs=[
        gr.components.Textbox(lines=2, label="Input", placeholder="Tell me about alpacas."),
        # ... [rest of the inputs]
    ],
    outputs=[
        gr.inputs.Textbox(lines=25, label="Output"),
    ],
    title="Chinese-Vicuna 中文小羊驼",
    description="Chatlaw app trained on HK law data",
).launch()