app.py

import os
import shutil
import torch
import gradio as gr
from huggingface_hub import HfApi, whoami, ModelCard
from gradio_huggingfacehub_search import HuggingfaceHubSearch
from textwrap import dedent
from pathlib import Path

from tempfile import TemporaryDirectory

from huggingface_hub.file_download import repo_folder_name
from optimum.exporters.tasks import TasksManager
from optimum.intel.utils.constant import _TASK_ALIASES
from optimum.intel.openvino.utils import _HEAD_TO_AUTOMODELS
from optimum.exporters import TasksManager

from optimum.intel.utils.modeling_utils import _find_files_matching_pattern
from optimum.intel import (
    OVModelForAudioClassification,
    OVModelForCausalLM,
    OVModelForFeatureExtraction,
    OVModelForImageClassification,
    OVModelForMaskedLM,
    OVModelForQuestionAnswering,
    OVModelForSeq2SeqLM,
    OVModelForSequenceClassification,
    OVModelForTokenClassification,
    OVStableDiffusionPipeline,
    OVStableDiffusionXLPipeline,
    OVLatentConsistencyModelPipeline,
    OVModelForPix2Struct,
    OVWeightQuantizationConfig,
)
from optimum.intel.openvino.modeling_diffusion import OVStableDiffusionPipelineBase


def export(
    model_id: str,
    private_repo: bool,
    oauth_token: gr.OAuthToken,
):
    if oauth_token.token is None:
        raise ValueError("You must be logged in to use this space")

    model_name = model_id.split("/")[-1]
    username = whoami(oauth_token.token)["name"]
    new_repo_id = f"{username}/{model_name}-openvino"

    task = TasksManager.infer_task_from_model(model_id)
    if task not in _HEAD_TO_AUTOMODELS:
        raise ValueError(
            f"The task '{task}' is not supported, only {_HEAD_TO_AUTOMODELS.keys()} tasks are supported"
        )

    if task == "text2text-generation":
        raise ValueError("Export of Seq2Seq models is currently disabled.")

    auto_model_class = _HEAD_TO_AUTOMODELS[task]
    ov_files = _find_files_matching_pattern(
        model_id,
        pattern=r"(.*)?openvino(.*)?\_model.xml",
        use_auth_token=oauth_token.token,
    )

    if len(ov_files) > 0:
        raise Exception(f"Model {model_id} is already converted, skipping..")

    api = HfApi(token=oauth_token.token)

    with TemporaryDirectory() as d:
        folder = os.path.join(d, repo_folder_name(repo_id=model_id, repo_type="models"))
        os.makedirs(folder)
        try:
            api.snapshot_download(
                repo_id=model_id, local_dir=folder, allow_patterns=["*.json"]
            )

            ov_model = eval(auto_model_class).from_pretrained(model_id, export=True)
            ov_model.save_pretrained(folder)

            """
            if not isinstance(ov_model, OVStableDiffusionPipelineBase):
                model = TasksManager.get_model_from_task(task, model_id)
                exporter_config_class = TasksManager.get_exporter_config_constructor(
                    exporter="openvino",
                    model=model,
                    task=task,
                    model_name=model_id,
                    model_type=model.config.model_type.replace("_", "-"),
                )
                openvino_config = exporter_config_class(model.config)
                inputs = openvino_config.generate_dummy_inputs(framework="pt")
                ov_outputs = ov_model(**inputs)
                outputs = model(**inputs)

                for output_name in ov_outputs:
                    if isinstance(outputs, torch.Tensor) and not torch.allclose(outputs[output_name], ov_outputs[output_name], atol=1e-3):
                        raise ValueError(
                            "The exported model does not have the same outputs as the original model. Export interrupted."
                        )
            """

            new_repo_url = api.create_repo(repo_id=new_repo_id, exist_ok=True, private=private_repo)
            new_repo_id = new_repo_url.repo_id
            print("Repo created successfully!", new_repo_url)

            folder = Path(folder)
            folder_parts = len(folder.parts)
            for file_path in folder.glob("**/*"):
                name = Path(*file_path.parts[folder_parts:])
                if not file_path.is_file() or any(part_name.startswith(".") for part_name in name.parts):
                    continue
                try:
                    api.upload_file(
                        path_or_fileobj=file_path,
                        path_in_repo=str(name),
                        repo_id=new_repo_id,
                    )
                except Exception as e:
                    raise Exception(f"Error uploading file {file_path}: {e}")

            try:
                card = ModelCard.load(model_id, token=oauth_token.token)
            except:
                card = ModelCard("")

            if card.data.tags is None:
                card.data.tags = []
            card.data.tags.append("openvino")
            card.data.base_model = model_id
            card.text = dedent(
                f"""
                This model was converted to OpenVINO from [`{model_id}`](https://huggingface.co/{model_id}) using [optimum-intel](https://github.com/huggingface/optimum-intel)
                via the [export](https://huggingface.co/spaces/echarlaix/openvino-export) space.

                First make sure you have optimum-intel installed:

                ```bash
                pip install optimum[openvino]
                ```

                To load your model you can do as follows:

                ```python
                from optimum.intel import {auto_model_class}

                model_id = "{new_repo_id}"
                model = {auto_model_class}.from_pretrained(model_id)
                ```
                """
            )
            card_path = os.path.join(folder, "README.md")
            card.save(card_path)

            api.upload_file(
                path_or_fileobj=card_path,
                path_in_repo="README.md",
                repo_id=new_repo_id,
            )
            return f"This model was successfully exported, find it under your repo {new_repo_url}'"
        finally:
            shutil.rmtree(folder, ignore_errors=True)


model_id = HuggingfaceHubSearch(
    label="Hub Model ID",
    placeholder="Search for model id on the hub",
    search_type="model",
)
private_repo = gr.Checkbox(
    value=False,
    label="Private Repo",
    info="Create a private repo under your username",
)
interface = gr.Interface(
    fn=export,
    inputs=[
        model_id,
        private_repo,
    ],
    outputs=[
        gr.Markdown(label="output"),
    ],
    title="Export your model to OpenVINO",
    description="This space converts your model to the OpenVINO format using [optimum-intel](https://huggingface.co/docs/optimum/main/intel/openvino/inference) The resulting model will then be pushed on the Hub under your HF user namespace",
    api_name=False,
)

with gr.Blocks() as demo:
    gr.Markdown("You must be logged in to use this space")
    gr.LoginButton(min_width=250)
    interface.render()

demo.launch()