app.py

import os
import shutil
import subprocess
import signal
os.environ["GRADIO_ANALYTICS_ENABLED"] = "False"
import gradio as gr

from huggingface_hub import HfApi
from huggingface_hub import ModelCard

from textwrap import dedent

HF_PATH = "https://huggingface.co/"

CONV_TEMPLATES = [
    "llama-3",
    "llama-3_1",
    "chatml",
    "chatml_nosystem",
    "qwen2",
    "open_hermes_mistral",
    "neural_hermes_mistral",
    "llama_default",
    "llama-2",
    "mistral_default",
    "gpt2",
    "codellama_completion",
    "codellama_instruct",
    "vicuna_v1.1",
    "conv_one_shot",
    "redpajama_chat",
    "rwkv_world",
    "rwkv",
    "gorilla",
    "gorilla-openfunctions-v2",
    "guanaco",
    "dolly",
    "oasst",
    "stablelm",
    "stablecode_completion",
    "stablecode_instruct",
    "minigpt",
    "moss",
    "LM",
    "stablelm-3b",
    "gpt_bigcode",
    "wizardlm_7b",
    "wizard_coder_or_math",
    "glm",
    "custom",  # for web-llm only
    "phi-2",
    "phi-3",
    "phi-3-vision",
    "stablelm-2",
    "gemma_instruction",
    "orion",
    "llava",
    "hermes2_pro_llama3",
    "hermes3_llama-3_1",
    "tinyllama_v1_0",
    "aya-23",
]

QUANTIZATIONS = ["q0f16", 
                 "q0f32", 
                 "q3f16_1", 
                 "q4f16_1", 
                 "q4f32_1", 
                 "q4f16_awq"]

def button_click(hf_model_id, conv_template, quantization, oauth_token: gr.OAuthToken | None):
    if not oauth_token.token:
        raise ValueError("Log in to Huggingface to use this")
    
    api = HfApi(token=oauth_token.token)
    model_dir_name = hf_model_id.split("/")[1]
    mlc_model_name = model_dir_name + "-" + quantization + "-" + "MLC"

    os.system("mkdir -p dist/models")
    os.system("git lfs install")

    api.snapshot_download(repo_id=hf_model_id, local_dir=f"./dist/models/{model_dir_name}")

    os.system("mlc_llm convert_weight ./dist/models/" + model_dir_name + "/" + \
              " --quantization " + quantization + \
              " -o dist/" + mlc_model_name)
    
    os.system("mlc_llm gen_config ./dist/models/" + model_dir_name + "/" + \
              " --quantization " + quantization + " --conv-template " + conv_template + \
              " -o dist/" + mlc_model_name + "/")
    
    # push to HF
    user_name = api.whoami()["name"]
    created_repo_url = api.create_repo(repo_id=f"{user_name}/{mlc_model_name}", private=True)
    created_repo_id = created_repo_url.repo_id

    api.upload_large_folder(folder_path=f"./dist/{mlc_model_name}",
                            repo_id=f"{user_name}/{mlc_model_name}",
                            repo_type="model")
    
    # push model card to HF
    card = ModelCard.load(hf_model_id, token=oauth_token.token)
    if not card.data.tags:
        card.data.tags = []
    card.data.tags.append("mlc-ai")
    card.data.tags.append("MLC-Weight-Conversion")
    card.data.base_model = hf_model_id

    card.text = dedent(
        f"""
        # {created_repo_id}
        This model was compiled using MLC-LLM with {quantization} quantization from [{hf_model_id}]({HF_PATH}{hf_model_id}).
        The conversion was done using the [MLC-Weight-Conversion](https://huggingface.co/spaces/mlc-ai/MLC-Weight-Conversion) space.

        To run this model, please first install [MLC-LLM](https://llm.mlc.ai/docs/install/mlc_llm.html#install-mlc-packages).

        To chat with the model on your terminal:
        ```bash
        mlc_llm chat HF://{created_repo_id}
        ```

        For more information on how to use MLC-LLM, please visit the MLC-LLM [documentation](https://llm.mlc.ai/docs/index.html).
        """
    )
    card.save("./dist/README.md")

    api.upload_file(path_or_fileobj="./dist/README.md",
                    path_in_repo="README.md",
                    repo_id=created_repo_id,
                    repo_type="model")

    os.system("rm -rf dist/")
    
    return "Successful"

with gr.Blocks() as demo:
    gr.LoginButton()
    model_id = gr.Textbox(label="HF Model ID")
    conv = gr.Dropdown(CONV_TEMPLATES, label="Conversation Template")
    quant = gr.Dropdown(QUANTIZATIONS, label="Quantization Method")
    btn = gr.Button("Convert to MLC")
    out = gr.Textbox(label="Conversion Result")
    btn.click(fn=button_click , inputs=[model_id, conv, quant], outputs=out)

demo.launch()