mlflow--mlflow/mlflow/pytorch/__init__.py

"""
The ``mlflow.pytorch`` module provides an API for logging and loading PyTorch models. This module
exports PyTorch models with the following flavors:

PyTorch (native) format
    This is the main flavor that can be loaded back into PyTorch.
:py:mod:`mlflow.pyfunc`
    Produced for use by generic pyfunc-based deployment tools and batch inference.
"""

import atexit
import importlib
import itertools
import logging
import os
import warnings
from functools import partial
from typing import Any, Literal

import numpy as np
import pandas as pd
import yaml
from packaging.version import Version

import mlflow
from mlflow import pyfunc
from mlflow.environment_variables import (
    MLFLOW_ALLOW_PICKLE_DESERIALIZATION,
    MLFLOW_DEFAULT_PREDICTION_DEVICE,
)
from mlflow.exceptions import MlflowException
from mlflow.ml_package_versions import _ML_PACKAGE_VERSIONS
from mlflow.models import Model, ModelSignature
from mlflow.models.model import MLMODEL_FILE_NAME
from mlflow.models.signature import _infer_signature_from_input_example
from mlflow.models.utils import ModelInputExample, _save_example
from mlflow.protos.databricks_pb2 import RESOURCE_DOES_NOT_EXIST
from mlflow.pytorch import pickle_module as mlflow_pytorch_pickle_module
from mlflow.tracking._model_registry import DEFAULT_AWAIT_MAX_SLEEP_SECONDS
from mlflow.tracking.artifact_utils import _download_artifact_from_uri
from mlflow.utils.autologging_utils import autologging_integration, safe_patch
from mlflow.utils.checkpoint_utils import download_checkpoint_artifact
from mlflow.utils.databricks_utils import (
    is_in_databricks_model_serving_environment,
    is_in_databricks_runtime,
)
from mlflow.utils.docstring_utils import LOG_MODEL_PARAM_DOCS, format_docstring
from mlflow.utils.environment import (
    _CONDA_ENV_FILE_NAME,
    _CONSTRAINTS_FILE_NAME,
    _PYTHON_ENV_FILE_NAME,
    _REQUIREMENTS_FILE_NAME,
    _mlflow_conda_env,
    _process_conda_env,
    _process_pip_requirements,
    _PythonEnv,
    _validate_env_arguments,
)
from mlflow.utils.file_utils import (
    TempDir,
    get_total_file_size,
    write_to,
)
from mlflow.utils.model_utils import (
    _add_code_from_conf_to_system_path,
    _copy_extra_files,
    _get_flavor_configuration,
    _validate_and_copy_code_paths,
    _validate_and_prepare_target_save_path,
)
from mlflow.utils.requirements_utils import _get_pinned_requirement

FLAVOR_NAME = "pytorch"

SERIALIZATION_FORMAT_PICKLE = "pickle"
SERIALIZATION_FORMAT_PT2 = "pt2"
_SERIALIZED_TORCH_MODEL_FILE_NAME = "model.pth"
_EXPORTED_TORCH_MODEL_FILE_NAME = "model.pt2"
_TORCH_STATE_DICT_FILE_NAME = "state_dict.pth"
_PICKLE_MODULE_INFO_FILE_NAME = "pickle_module_info.txt"
_TORCH_CPU_DEVICE_NAME = "cpu"
_TORCH_DEFAULT_GPU_DEVICE_NAME = "cuda"

_logger = logging.getLogger(__name__)

MIN_REQ_VERSION = Version(_ML_PACKAGE_VERSIONS["pytorch-lightning"]["autologging"]["minimum"])
MAX_REQ_VERSION = Version(_ML_PACKAGE_VERSIONS["pytorch-lightning"]["autologging"]["maximum"])


_MODEL_DATA_SUBPATH = "data"


def get_default_pip_requirements():
    """
    Returns:
        A list of default pip requirements for MLflow Models produced by this flavor. Calls to
        :func:`save_model()` and :func:`log_model()` produce a pip environment that, at minimum,
        contains these requirements.
    """
    return list(
        map(
            _get_pinned_requirement,
            [
                "torch",
                # We include CloudPickle in the default environment because
                # it's required by the default pickle module used by `save_model()`
                # and `log_model()`: `mlflow.pytorch.pickle_module`.
                "cloudpickle",
            ],
        )
    )


def get_default_conda_env():
    """
    Returns:
        The default Conda environment as a dictionary for MLflow Models produced by calls to
        :func:`save_model()` and :func:`log_model()`.

    .. code-block:: python
        :caption: Example

        import mlflow

        # Log PyTorch model
        with mlflow.start_run() as run:
            mlflow.pytorch.log_model(model, name="model", signature=signature)

        # Fetch the associated conda environment
        env = mlflow.pytorch.get_default_conda_env()
        print(f"conda env: {env}")

    .. code-block:: text
        :caption: Output

        conda env {'name': 'mlflow-env',
                   'channels': ['conda-forge'],
                   'dependencies': ['python=3.8.15',
                                    {'pip': ['torch==1.5.1',
                                             'mlflow',
                                             'cloudpickle==1.6.0']}]}
    """
    return _mlflow_conda_env(additional_pip_deps=get_default_pip_requirements())


@format_docstring(LOG_MODEL_PARAM_DOCS.format(package_name="torch"))
def log_model(
    pytorch_model,
    artifact_path: str | None = None,
    conda_env=None,
    code_paths=None,
    pickle_module=None,
    registered_model_name=None,
    signature: ModelSignature = None,
    input_example: ModelInputExample = None,
    await_registration_for=DEFAULT_AWAIT_MAX_SLEEP_SECONDS,
    extra_files=None,
    pip_requirements=None,
    extra_pip_requirements=None,
    metadata=None,
    name: str | None = None,
    params: dict[str, Any] | None = None,
    tags: dict[str, Any] | None = None,
    model_type: str | None = None,
    step: int = 0,
    model_id: str | None = None,
    export_model: bool = False,
    serialization_format: Literal["pickle", "pt2"] = SERIALIZATION_FORMAT_PT2,
    **kwargs,
):
    """
    Log a PyTorch model as an MLflow artifact for the current run.

    .. warning:: Log the model with a signature to avoid inference errors.
        If the model is logged without a signature, the MLflow Model Server relies on the
        default inferred data type from NumPy. However, PyTorch often expects different
        defaults, particularly when parsing floats. You must include the signature to ensure
        that the model is logged with the correct data type so that the MLflow model server
        can correctly provide valid input.

    Args:
        pytorch_model: PyTorch model to be saved. Can be either an eager model (subclass of
            ``torch.nn.Module``) or scripted model prepared via ``torch.jit.script`` or
            ``torch.jit.trace``.

            The model accept a single ``torch.FloatTensor`` as input and produce a single output
            tensor.

            If saving an eager model, any code dependencies of the model's class, including the
            class definition itself, should be included in one of the following locations:

                - The package(s) listed in the model's Conda environment, specified by the
                  ``conda_env`` parameter.
                - One or more of the files specified by the ``code_paths`` parameter.
        artifact_path: Deprecated. Use `name` instead.
        conda_env: {{ conda_env }}
        code_paths: {{ code_paths }}
        pickle_module: The module that PyTorch should use to serialize ("pickle") the specified
            ``pytorch_model``. This is passed as the ``pickle_module`` parameter to
            ``torch.save()``.  By default, this module is also used to deserialize ("unpickle") the
            PyTorch model at load time.
        registered_model_name: If given, create a model version under ``registered_model_name``,
            also create a registered model if one with the given name does not exist.
        signature: {{ signature }}
        input_example: {{ input_example }}
        await_registration_for: Number of seconds to wait for the model version to finish
            being created and is in ``READY`` status. By default, the function waits for five
            minutes.  Specify 0 or None to skip waiting.

        extra_files: {{ extra_files }}
        pip_requirements: {{ pip_requirements }}
        extra_pip_requirements: {{ extra_pip_requirements }}
        metadata: {{ metadata }}
        name: {{ name }}
        params: {{ params }}
        tags: {{ tags }}
        model_type: {{ model_type }}
        step: {{ step }}
        model_id: {{ model_id }}
        export_model: If set to True, save the model as "pt2" format. This argument is deprecated.
            For details, see documentation of `serialization_format` argument.
        serialization_format: The serialization format used to save the PyTorch model.
            Accepted values are "pickle" and "pt2".
            When set to "pickle", the model is serialized using either pickle or cloudpickle,
            depending on the `pickle_module` parameter.
            When set to "pt2", the model is saved using torch.export.save, which exports the model
            as a traced graph. This is a safer serialization format that prevents executing
            arbitrary code during deserialization.
            Note that "pt2" format requires `input_example` (used to trace the model graph by
            virtually executing model.forward) and only supports Numpy array / Tensor or a list
            of Numpy arrays / Tensors as inputs. For details, see
            https://docs.pytorch.org/docs/stable/user_guide/torch_compiler/export/pt2_archive.html.
        kwargs: kwargs to pass to ``torch.save`` method.

    Returns:
        A :py:class:`ModelInfo <mlflow.models.model.ModelInfo>` instance that contains the
        metadata of the logged model.

    .. code-block:: python
        :caption: Example

        import numpy as np
        import torch
        import mlflow
        from mlflow import MlflowClient
        from mlflow.models import infer_signature

        # Define model, loss, and optimizer
        model = nn.Linear(1, 1)
        criterion = torch.nn.MSELoss()
        optimizer = torch.optim.SGD(model.parameters(), lr=0.001)

        # Create training data with relationship y = 2X
        X = torch.arange(1.0, 26.0).reshape(-1, 1)
        y = X * 2

        # Training loop
        epochs = 250
        for epoch in range(epochs):
            # Forward pass: Compute predicted y by passing X to the model
            y_pred = model(X)

            # Compute the loss
            loss = criterion(y_pred, y)

            # Zero gradients, perform a backward pass, and update the weights.
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

        # Create model signature
        signature = infer_signature(X.numpy(), model(X).detach().numpy())

        # Log the model
        with mlflow.start_run() as run:
            mlflow.pytorch.log_model(model, name="model", input_example=X)

            # convert to scripted model and log the model
            scripted_pytorch_model = torch.jit.script(model)
            mlflow.pytorch.log_model(
                scripted_pytorch_model, name="scripted_model", serialization_format="pickle"
            )

        # Fetch the logged model artifacts
        print(f"run_id: {run.info.run_id}")
        for artifact_path in ["model/data", "scripted_model/data"]:
            artifacts = [
                f.path for f in MlflowClient().list_artifacts(run.info.run_id, artifact_path)
            ]
            print(f"artifacts: {artifacts}")

    .. code-block:: text
        :caption: Output

        run_id: 1a1ec9e413ce48e9abf9aec20efd6f71
        artifacts: ['model/data/model.pth',
                    'model/data/pickle_module_info.txt']
        artifacts: ['scripted_model/data/model.pth',
                    'scripted_model/data/pickle_module_info.txt']

    .. figure:: ../_static/images/pytorch_logged_models.png

        PyTorch logged models
    """
    pickle_module = pickle_module or mlflow_pytorch_pickle_module
    return Model.log(
        artifact_path=artifact_path,
        name=name,
        flavor=mlflow.pytorch,
        pytorch_model=pytorch_model,
        conda_env=conda_env,
        code_paths=code_paths,
        pickle_module=pickle_module,
        registered_model_name=registered_model_name,
        signature=signature,
        input_example=input_example,
        await_registration_for=await_registration_for,
        extra_files=extra_files,
        pip_requirements=pip_requirements,
        extra_pip_requirements=extra_pip_requirements,
        metadata=metadata,
        params=params,
        tags=tags,
        model_type=model_type,
        step=step,
        model_id=model_id,
        export_model=export_model,
        serialization_format=serialization_format,
        **kwargs,
    )


@format_docstring(LOG_MODEL_PARAM_DOCS.format(package_name="torch"))
def save_model(
    pytorch_model,
    path,
    conda_env=None,
    mlflow_model=None,
    code_paths=None,
    pickle_module=None,
    signature: ModelSignature = None,
    input_example: ModelInputExample = None,
    extra_files=None,
    pip_requirements=None,
    extra_pip_requirements=None,
    metadata=None,
    export_model: bool = False,
    serialization_format: Literal["pickle", "pt2"] = SERIALIZATION_FORMAT_PT2,
    **kwargs,
):
    """
    Save a PyTorch model to a path on the local file system.

    Args:
        pytorch_model: PyTorch model to be saved. Can be either an eager model (subclass of
            ``torch.nn.Module``) or a scripted model prepared via ``torch.jit.script`` or
            ``torch.jit.trace``.

            To save an eager model, any code dependencies of the model's class, including the class
            definition itself, should be included in one of the following locations:

                - The package(s) listed in the model's Conda environment, specified by the
                  ``conda_env`` parameter.
                - One or more of the files specified by the ``code_paths`` parameter.

        path: Local path where the model is to be saved.
        conda_env: {{ conda_env }}
        mlflow_model: :py:mod:`mlflow.models.Model` this flavor is being added to.
        code_paths: {{ code_paths }}
        pickle_module: The module that PyTorch should use to serialize ("pickle") the specified
            ``pytorch_model``. This is passed as the ``pickle_module`` parameter to
            ``torch.save()``. By default, this module is also used to deserialize ("unpickle") the
            model at loading time.
        signature: {{ signature }}
        input_example: {{ input_example }}

        extra_files: {{ extra_files }}
        pip_requirements: {{ pip_requirements }}
        extra_pip_requirements: {{ extra_pip_requirements }}
        metadata:{{ metadata }}
        export_model: If set to True, save the model as "pt2" format. This argument is deprecated.
            For details, see documentation of `serialization_format` argument.
        serialization_format: The serialization format used to save the PyTorch model.
            Accepted values are "pickle" and "pt2".
            When set to "pickle", the model is serialized using either pickle or cloudpickle,
            depending on the `pickle_module` parameter.
            When set to "pt2", the model is saved using torch.export.save, which exports the model
            as a traced graph. This is a safer serialization format that prevents executing
            arbitrary code during deserialization.
            Note that "pt2" format requires `input_example` (used to trace the model graph by
            virtually executing model.forward) and only supports Numpy array / Tensor or a list
            of Numpy arrays / Tensors as inputs. For details, see
            https://docs.pytorch.org/docs/stable/user_guide/torch_compiler/export/pt2_archive.html.
        kwargs: kwargs to pass to ``torch.save`` method.

    .. code-block:: python
        :caption: Example

        import os
        import mlflow
        import torch


        model = nn.Linear(1, 1)

        # Save PyTorch models to current working directory
        with mlflow.start_run() as run:
            mlflow.pytorch.save_model(model, "model", input_example=torch.ones(1, 1))

            # Convert to a scripted model and save it
            scripted_pytorch_model = torch.jit.script(model)
            mlflow.pytorch.save_model(
                scripted_pytorch_model, "scripted_model", serialization_format="pickle"
            )

        # Load each saved model for inference
        for model_path in ["model", "scripted_model"]:
            model_uri = f"{os.getcwd()}/{model_path}"
            loaded_model = mlflow.pytorch.load_model(model_uri)
            print(f"Loaded {model_path}:")
            for x in [6.0, 8.0, 12.0, 30.0]:
                X = torch.Tensor([[x]])
                y_pred = loaded_model(X)
                print(f"predict X: {x}, y_pred: {y_pred.data.item():.2f}")
            print("--")

    .. code-block:: text
        :caption: Output

        Loaded model:
        predict X: 6.0, y_pred: 11.90
        predict X: 8.0, y_pred: 15.92
        predict X: 12.0, y_pred: 23.96
        predict X: 30.0, y_pred: 60.13
        --
        Loaded scripted_model:
        predict X: 6.0, y_pred: 11.90
        predict X: 8.0, y_pred: 15.92
        predict X: 12.0, y_pred: 23.96
        predict X: 30.0, y_pred: 60.13

    """
    import torch
    from torch.export import Dim as ExportDim

    _validate_env_arguments(conda_env, pip_requirements, extra_pip_requirements)

    if export_model:
        if serialization_format == SERIALIZATION_FORMAT_PICKLE:
            raise MlflowException.invalid_parameter_value(
                "Cannot set both `export_model=True` and `serialization_format='pickle'`. "
                "Please use only the `serialization_format` argument."
            )
        warnings.warn(
            "`export_model` argument is deprecated. "
            "Please set `serialization_format` argument instead.",
            FutureWarning,
            stacklevel=2,
        )
        serialization_format = "pt2"

    if serialization_format not in [SERIALIZATION_FORMAT_PICKLE, SERIALIZATION_FORMAT_PT2]:
        raise MlflowException.invalid_parameter_value(
            "The `serialization_format` param value must be one of 'pickle' or 'pt2'."
        )

    pickle_module = pickle_module or mlflow_pytorch_pickle_module

    if not isinstance(pytorch_model, torch.nn.Module):
        raise TypeError("Argument 'pytorch_model' should be a torch.nn.Module")
    path = os.path.abspath(path)
    _validate_and_prepare_target_save_path(path)

    if mlflow_model is None:
        mlflow_model = Model()

    try:
        saved_example = _save_example(mlflow_model, input_example, path)
    except MlflowException:
        # `_save_example` does not support tensor / list of tensors / list of numpy array as input.
        saved_example = None

    if signature is None and saved_example is not None:
        wrapped_model = _PyTorchWrapper(pytorch_model, device="cpu")
        signature = _infer_signature_from_input_example(saved_example, wrapped_model)
    elif signature is False:
        signature = None

    if signature is not None:
        mlflow_model.signature = signature
    if metadata is not None:
        mlflow_model.metadata = metadata

    code_dir_subpath = _validate_and_copy_code_paths(code_paths, path)

    model_data_subpath = _MODEL_DATA_SUBPATH
    model_data_path = os.path.join(path, model_data_subpath)
    os.makedirs(model_data_path)

    # Save pytorch model
    if serialization_format == SERIALIZATION_FORMAT_PT2:
        if Version(torch.__version__) < Version("2.4"):
            raise MlflowException(
                "If `serialization_format` is set to 'pt2', `torch` package version must be >= 2.4."
            )

        if isinstance(pytorch_model, torch.jit.ScriptModule):
            raise MlflowException(
                "The 'pt2' serialization format is not supported for torch.jit.ScriptModule "
                "models, please set `serialization_format` to 'pickle'."
            )

        if isinstance(input_example, (np.ndarray, torch.Tensor)):
            input_example = (input_example,)
        if input_example is None or not all(
            isinstance(value, (np.ndarray, torch.Tensor)) for value in input_example
        ):
            raise MlflowException(
                "If `serialization_format` is set to 'pt2', then input_example is required. "
                "It must be a numpy array or torch tensor, or a tuple/list of numpy arrays "
                "or torch tensors. This is because 'pt2' is a traced-graph format: "
                "PyTorch traces the model graph by virtually executing model.forward with "
                "the provided example input."
            )

        if not (
            signature is not None
            and signature.inputs is not None
            and len(signature.inputs) >= 1
            and signature.inputs.is_tensor_spec()
        ):
            raise MlflowException(
                "Unsupported signature type for the selected serialization format. "
                "If the `serialization_format` argument is set to 'pt2', the input signature "
                "must be specified using `TensorSpec`. Please update the model signature or "
                "set `serialization_format` to 'pickle'."
            )

        tensor_spec_list = signature.inputs.inputs

        dynamic_shapes = []

        for tensor_spec in tensor_spec_list:
            try:
                dynamic_dim = tensor_spec.shape.index(-1)
                dynamic_shape = {dynamic_dim: ExportDim("dynamic_dim")}
            except ValueError:
                dynamic_shape = None
            dynamic_shapes.append(dynamic_shape)

        exported_prog = torch.export.export(
            pytorch_model,
            tuple(torch.from_numpy(v) if isinstance(v, np.ndarray) else v for v in input_example),
            dynamic_shapes=dynamic_shapes,
        )
        model_path = os.path.join(model_data_path, _EXPORTED_TORCH_MODEL_FILE_NAME)
        torch.export.save(exported_prog, model_path)
    else:
        if not is_in_databricks_runtime():
            _logger.warning(
                "Saving pytorch model by Pickle or CloudPickle format requires exercising "
                "caution because these formats rely on Python's object serialization mechanism, "
                "which can execute arbitrary code during deserialization. "
                "The recommended safe alternative is to set `serialization_format` to 'pt2' to "
                "save the PyTorch model using the safe graph model format."
            )
        # Persist the pickle module name as a file in the model's `data` directory. This is
        # necessary
        # because the `data` directory is the only available parameter to `_load_pyfunc`, and it
        # does not contain the MLmodel configuration; therefore, it is not sufficient to place
        # the module name in the MLmodel
        #
        # TODO: Stop persisting this information to the filesystem once we have a mechanism for
        # supplying the MLmodel configuration to `mlflow.pytorch._load_pyfunc`
        pickle_module_path = os.path.join(model_data_path, _PICKLE_MODULE_INFO_FILE_NAME)
        with open(pickle_module_path, "w") as f:
            f.write(pickle_module.__name__)
        model_path = os.path.join(model_data_path, _SERIALIZED_TORCH_MODEL_FILE_NAME)
        if isinstance(pytorch_model, torch.jit.ScriptModule):
            torch.jit.ScriptModule.save(pytorch_model, model_path)
        else:
            torch.save(pytorch_model, model_path, pickle_module=pickle_module, **kwargs)

    extra_files_config = _copy_extra_files(extra_files, path)

    mlflow_model.add_flavor(
        FLAVOR_NAME,
        model_data=model_data_subpath,
        pytorch_version=str(torch.__version__),
        code=code_dir_subpath,
        **extra_files_config,
    )
    # For the case that `input_example` is a tensor or list of numpy arrays / tensors,
    # PyFunc model is not supported yet.
    if not (
        serialization_format == SERIALIZATION_FORMAT_PT2
        and (len(input_example) > 1 or any(isinstance(x, torch.Tensor) for x in input_example))
    ):
        pyfunc.add_to_model(
            mlflow_model,
            loader_module="mlflow.pytorch",
            data=model_data_subpath,
            pickle_module_name=pickle_module.__name__,
            code=code_dir_subpath,
            conda_env=_CONDA_ENV_FILE_NAME,
            python_env=_PYTHON_ENV_FILE_NAME,
            model_config={"device": None},
        )
    if size := get_total_file_size(path):
        mlflow_model.model_size_bytes = size
    mlflow_model.save(os.path.join(path, MLMODEL_FILE_NAME))

    if conda_env is None:
        if pip_requirements is None:
            default_reqs = get_default_pip_requirements()
            # To ensure `_load_pyfunc` can successfully load the model during the dependency
            # inference, `mlflow_model.save` must be called beforehand to save an MLmodel file.
            inferred_reqs = mlflow.models.infer_pip_requirements(
                model_data_path,
                FLAVOR_NAME,
                fallback=default_reqs,
            )
            default_reqs = sorted(set(inferred_reqs).union(default_reqs))
        else:
            default_reqs = None
        conda_env, pip_requirements, pip_constraints = _process_pip_requirements(
            default_reqs,
            pip_requirements,
            extra_pip_requirements,
        )
    else:
        conda_env, pip_requirements, pip_constraints = _process_conda_env(conda_env)

    with open(os.path.join(path, _CONDA_ENV_FILE_NAME), "w") as f:
        yaml.safe_dump(conda_env, stream=f, default_flow_style=False)

    # Save `constraints.txt` if necessary
    if pip_constraints:
        write_to(os.path.join(path, _CONSTRAINTS_FILE_NAME), "\n".join(pip_constraints))

    # Save `requirements.txt`
    write_to(os.path.join(path, _REQUIREMENTS_FILE_NAME), "\n".join(pip_requirements))

    _PythonEnv.current().to_yaml(os.path.join(path, _PYTHON_ENV_FILE_NAME))


def _load_by_pickle_check(is_loading_state_dict: bool):
    if (
        not MLFLOW_ALLOW_PICKLE_DESERIALIZATION.get()
        and not is_in_databricks_runtime()
        and not is_in_databricks_model_serving_environment()
    ):
        if is_loading_state_dict:
            raise MlflowException(
                "Deserializing model using pickle is disallowed, but this state dict is saved "
                "in pickle format. You can set environment variable "
                "'MLFLOW_ALLOW_PICKLE_DESERIALIZATION' to 'true' to allow deserializing state "
                "dict using pickle."
            )
        raise MlflowException(
            "Deserializing model using pickle is disallowed, but this model is saved "
            "in pickle format. The recommended way is to save the model with "
            "serialization_format='pt2' like "
            "`mlflow.pytorch.save_model(model, path, serialization_format='pt2')`, "
            "or you can set environment variable "
            "'MLFLOW_ALLOW_PICKLE_DESERIALIZATION' to 'true' to allow deserializing model "
            "using pickle."
        )


def _load_model(path, device=None, **kwargs):
    """
    Args:
        path: The path to a serialized PyTorch model.
        device: If specified, load the model on the specified device.
        kwargs: Additional kwargs to pass to the PyTorch ``torch.load`` function.
    """
    import torch

    if os.path.isdir(path):
        # `path` is a directory containing a serialized PyTorch model and a text file containing
        # information about the pickle module that should be used by PyTorch to load it

        model_path = os.path.join(path, "model.pth")
        if os.path.exists(model_path):
            is_exported_model = False
            pickle_module_path = os.path.join(path, _PICKLE_MODULE_INFO_FILE_NAME)
            with open(pickle_module_path) as f:
                pickle_module_name = f.read()
            if "pickle_module" in kwargs and kwargs["pickle_module"].__name__ != pickle_module_name:
                _logger.warning(
                    "Attempting to load the PyTorch model with a pickle module, '%s', that does not"
                    " match the pickle module that was used to save the model: '%s'.",
                    kwargs["pickle_module"].__name__,
                    pickle_module_name,
                )
            else:
                try:
                    kwargs["pickle_module"] = importlib.import_module(pickle_module_name)
                except ImportError as exc:
                    raise MlflowException(
                        message=(
                            "Failed to import the pickle module that was used to save the PyTorch"
                            f" model. Pickle module name: `{pickle_module_name}`"
                        ),
                        error_code=RESOURCE_DOES_NOT_EXIST,
                    ) from exc
        else:
            is_exported_model = True
            model_path = os.path.join(path, "model.pt2")
            kwargs = {}
    else:
        is_exported_model = False
        model_path = path

    if Version(torch.__version__) >= Version("1.5.0"):
        if is_exported_model:
            if Version(torch.__version__) < Version("2.4"):
                raise MlflowException(
                    "The model is exported by `torch.export` API. To load the model, "
                    "`torch` package version must be >= 2.4"
                )
            pytorch_model = torch.export.load(model_path, **kwargs).module()
        else:
            _load_by_pickle_check(False)
            pytorch_model = torch.load(model_path, **kwargs)
    else:
        try:
            # load the model as an eager model.
            _load_by_pickle_check(False)
            pytorch_model = torch.load(model_path, **kwargs)
        except Exception:
            # If fails, assume the model as a scripted model
            # `torch.jit.load` does not accept `pickle_module`.
            kwargs.pop("pickle_module", None)
            pytorch_model = torch.jit.load(model_path, **kwargs)

    if not is_exported_model:
        pytorch_model.eval()
    if device:
        if is_exported_model:
            target_device_type = torch.device(device).type
            # If the model is loaded from an exported model (pt2 format),
            # the model weights / buffers can't be moved across devices.
            # so we do device check instead.
            for tensor in itertools.chain(
                pytorch_model.parameters(),
                pytorch_model.buffers(),
            ):
                if tensor.device.type != target_device_type:
                    raise MlflowException(
                        "The saved model is exported by `torch.export` API, the original model "
                        f"contains weights / buffers on '{tensor.device.type}' device, it can't "
                        f"be loaded on '{target_device_type}' device. To address this issue, "
                        f"You should save the model in the following way: "
                        f"`mlflow.pytorch.save_model("
                        f"model.to('{target_device_type}'), path=..., serialization_format='pt2')`"
                    )
        else:
            pytorch_model.to(device=device)
    return pytorch_model


def load_model(model_uri, dst_path=None, **kwargs):
    """
    Load a PyTorch model from a local file or a run.

    Args:
        model_uri: The location, in URI format, of the MLflow model, for example:

            - ``/Users/me/path/to/local/model``
            - ``relative/path/to/local/model``
            - ``s3://my_bucket/path/to/model``
            - ``runs:/<mlflow_run_id>/run-relative/path/to/model``
            - ``models:/<model_name>/<model_version>``
            - ``models:/<model_name>/<stage>``

            For more information about supported URI schemes, see `Referencing Artifacts \
            <https://www.mlflow.org/docs/latest/concepts.html#artifact-locations>`_.
        dst_path: The local filesystem path to which to download the model artifact.
            This directory must already exist. If unspecified, a local output path will be created.
        kwargs: kwargs to pass to ``torch.load`` method.

    Returns:
        A PyTorch model.

    .. code-block:: python
        :caption: Example

        import torch
        import mlflow.pytorch


        model = nn.Linear(1, 1)

        # Log the model
        with mlflow.start_run() as run:
            mlflow.pytorch.log_model(model, name="model")

        # Inference after loading the logged model
        model_uri = f"runs:/{run.info.run_id}/model"
        loaded_model = mlflow.pytorch.load_model(model_uri)
        for x in [4.0, 6.0, 30.0]:
            X = torch.Tensor([[x]])
            y_pred = loaded_model(X)
            print(f"predict X: {x}, y_pred: {y_pred.data.item():.2f}")

    .. code-block:: text
        :caption: Output

        predict X: 4.0, y_pred: 7.57
        predict X: 6.0, y_pred: 11.64
        predict X: 30.0, y_pred: 60.48
    """
    import torch

    local_model_path = _download_artifact_from_uri(artifact_uri=model_uri, output_path=dst_path)
    pytorch_conf = _get_flavor_configuration(model_path=local_model_path, flavor_name=FLAVOR_NAME)
    _add_code_from_conf_to_system_path(local_model_path, pytorch_conf)

    if torch.__version__ != pytorch_conf["pytorch_version"]:
        _logger.warning(
            "Stored model version '%s' does not match installed PyTorch version '%s'",
            pytorch_conf["pytorch_version"],
            torch.__version__,
        )
    torch_model_artifacts_path = os.path.join(local_model_path, pytorch_conf["model_data"])
    return _load_model(path=torch_model_artifacts_path, **kwargs)


def _is_forecasting_model(model) -> bool:
    try:
        from pytorch_forecasting.models import BaseModel
    except ImportError:
        return False

    return isinstance(model, BaseModel)


def _load_pyfunc(path, model_config=None, weights_only=False):
    """
    Load PyFunc implementation. Called by ``pyfunc.load_model``.

    Args:
        path: Local filesystem path to the MLflow Model with the ``pytorch`` flavor.
    """
    import torch

    device = model_config.get("device", None) if model_config else None
    # if CUDA is available, we use the default CUDA device.
    # To force inference to the CPU when the GPU is available, please set
    # MLFLOW_DEFAULT_PREDICTION_DEVICE to "cpu"
    # If a specific non-default device is passed in, we continue to respect that.
    if device is None:
        if MLFLOW_DEFAULT_PREDICTION_DEVICE.get():
            device = MLFLOW_DEFAULT_PREDICTION_DEVICE.get()
        elif torch.cuda.is_available():
            device = _TORCH_DEFAULT_GPU_DEVICE_NAME
        else:
            device = _TORCH_CPU_DEVICE_NAME

    # in pytorch >= 2.6.0, the `weights_only` kwarg default has been changed from
    # `False` to `True`. this can cause pickle deserialization errors when loading
    # models, unless the model classes have been explicitly marked as safe using
    # `torch.serialization.add_safe_globals()`
    if Version(torch.__version__) >= Version("2.6.0"):
        return _PyTorchWrapper(
            _load_model(path, device=device, weights_only=weights_only), device=device
        )

    return _PyTorchWrapper(_load_model(path, device=device), device=device)


class _PyTorchWrapper:
    """
    Wrapper class that creates a predict function such that
    predict(data: pd.DataFrame) -> model's output as pd.DataFrame (pandas DataFrame)
    """

    def __init__(self, pytorch_model, device):
        self.pytorch_model = pytorch_model
        self.device = device
        self._is_forecasting_model = _is_forecasting_model(self.pytorch_model)

    def get_raw_model(self):
        """
        Returns the underlying model.
        """
        return self.pytorch_model

    def predict(self, data, params: dict[str, Any] | None = None):
        """
        Args:
            data: Model input data.
            params: Additional parameters to pass to the model for inference.

        Returns:
            Model predictions.
        """
        import torch

        if params and "device" in params:
            raise ValueError(
                "device' can no longer be specified as an inference parameter. "
                "It must be specified at load time. "
                "Please specify the device at load time, for example: "
                "`mlflow.pyfunc.load_model(model_uri, model_config={'device': 'cuda'})`."
            )

        if isinstance(data, pd.DataFrame):
            inp_data = data if self._is_forecasting_model else data.to_numpy(dtype=np.float32)
        elif isinstance(data, np.ndarray):
            if self._is_forecasting_model:
                raise TypeError(
                    "The pytorch forecasting model does not support numpy.ndarray input data, "
                    "please provide pandas.DataFrame input data."
                )
            inp_data = data
        elif isinstance(data, (list, dict)):
            raise TypeError(
                "The PyTorch flavor does not support List or Dict input types. "
                "Please use a pandas.DataFrame or a numpy.ndarray"
            )
        else:
            raise TypeError("Input data should be pandas.DataFrame or numpy.ndarray")

        device = self.device
        with torch.no_grad():
            if self._is_forecasting_model:
                # forecasting model `predict` method supports
                # dataframe input.
                preds = self.pytorch_model.predict(inp_data)
            else:
                input_tensor = torch.from_numpy(inp_data).to(device)
                preds = self.pytorch_model(input_tensor, **(params or {}))
            # if the predictions happened on a remote device, copy them back to
            # the host CPU for processing
            if device != _TORCH_CPU_DEVICE_NAME:
                preds = preds.to(_TORCH_CPU_DEVICE_NAME)
            if not isinstance(preds, torch.Tensor):
                raise TypeError(
                    "Expected PyTorch model to output a single output tensor, "
                    f"but got output of type '{type(preds)}'"
                )
            if isinstance(data, pd.DataFrame) and not self._is_forecasting_model:
                predicted = pd.DataFrame(preds.numpy())
                predicted.index = data.index
            else:
                predicted = preds.numpy()
            return predicted


def log_state_dict(state_dict, artifact_path, **kwargs):
    """
    Log a state_dict as an MLflow artifact for the current run.

    .. warning::
        This function just logs a state_dict as an artifact and doesn't generate
        an :ref:`MLflow Model <models>`.

    Args:
        state_dict: state_dict to be saved.
        artifact_path: Run-relative artifact path.
        kwargs: kwargs to pass to ``torch.save``.

    .. code-block:: python
        :caption: Example

        # Log a model as a state_dict
        with mlflow.start_run():
            state_dict = model.state_dict()
            mlflow.pytorch.log_state_dict(state_dict, artifact_path="model")

        # Log a checkpoint as a state_dict
        with mlflow.start_run():
            state_dict = {
                "model": model.state_dict(),
                "optimizer": optimizer.state_dict(),
                "epoch": epoch,
                "loss": loss,
            }
            mlflow.pytorch.log_state_dict(state_dict, artifact_path="checkpoint")
    """

    with TempDir() as tmp:
        local_path = tmp.path()
        save_state_dict(state_dict=state_dict, path=local_path, **kwargs)
        mlflow.log_artifacts(local_path, artifact_path)


def save_state_dict(state_dict, path, **kwargs):
    """
    Save a state_dict to a path on the local file system

    Args:
        state_dict: state_dict to be saved.
        path: Local path where the state_dict is to be saved.
        kwargs: kwargs to pass to ``torch.save``.
    """
    import torch

    # The object type check here aims to prevent a scenario where a user accidentally passees
    # a model instead of a state_dict and `torch.save` (which accepts both model and state_dict)
    # successfully completes, leaving the user unaware of the mistake.
    if not isinstance(state_dict, dict):
        raise TypeError(
            "Invalid object type for `state_dict`: {}. Must be an instance of `dict`".format(
                type(state_dict)
            )
        )

    os.makedirs(path, exist_ok=True)
    state_dict_path = os.path.join(path, _TORCH_STATE_DICT_FILE_NAME)
    torch.save(state_dict, state_dict_path, **kwargs)


def load_state_dict(state_dict_uri, **kwargs):
    """
    Load a state_dict from a local file or a run.

    Args:
        state_dict_uri: The location, in URI format, of the state_dict, for example:

            - ``/Users/me/path/to/local/state_dict``
            - ``relative/path/to/local/state_dict``
            - ``s3://my_bucket/path/to/state_dict``
            - ``runs:/<mlflow_run_id>/run-relative/path/to/state_dict``

            For more information about supported URI schemes, see `Referencing Artifacts \
            <https://www.mlflow.org/docs/latest/concepts.html#artifact-locations>`_.

        kwargs: kwargs to pass to ``torch.load``.

    Returns:
        A state_dict

    .. code-block:: python
        :caption: Example

        with mlflow.start_run():
            artifact_path = "model"
            mlflow.pytorch.log_state_dict(model.state_dict(), artifact_path)
            state_dict_uri = mlflow.get_artifact_uri(artifact_path)

        state_dict = mlflow.pytorch.load_state_dict(state_dict_uri)
    """
    import torch

    _load_by_pickle_check(True)
    local_path = _download_artifact_from_uri(artifact_uri=state_dict_uri)
    state_dict_path = os.path.join(local_path, _TORCH_STATE_DICT_FILE_NAME)
    return torch.load(state_dict_path, **kwargs)


@autologging_integration(FLAVOR_NAME)
def autolog(
    log_every_n_epoch=1,
    log_every_n_step=None,
    log_models=True,
    log_datasets=True,
    disable=False,
    exclusive=False,
    disable_for_unsupported_versions=False,
    silent=False,
    registered_model_name=None,
    extra_tags=None,
    checkpoint=True,
    checkpoint_monitor="val_loss",
    checkpoint_mode="min",
    checkpoint_save_best_only=True,
    checkpoint_save_weights_only=False,
    checkpoint_save_freq="epoch",
    log_model_signatures=True,
):
    """
    Enables (or disables) and configures autologging from `PyTorch Lightning
    <https://pytorch-lightning.readthedocs.io/en/latest>`_ to MLflow.

    Autologging is performed when you call the `fit` method of
    `pytorch_lightning.Trainer() \
    <https://pytorch-lightning.readthedocs.io/en/latest/trainer.html#>`_.

    Explore the complete `PyTorch MNIST \
    <https://github.com/mlflow/mlflow/tree/master/examples/pytorch/MNIST>`_ for
    an expansive example with implementation of additional lightening steps.

    **Note**: Full autologging is only supported for PyTorch Lightning models,
    i.e., models that subclass
    `pytorch_lightning.LightningModule \
    <https://pytorch-lightning.readthedocs.io/en/latest/lightning_module.html>`_.
    Autologging support for vanilla PyTorch (ie models that only subclass
    `torch.nn.Module <https://pytorch.org/docs/stable/generated/torch.nn.Module.html>`_)
    only autologs calls to
    `torch.utils.tensorboard.SummaryWriter <https://pytorch.org/docs/stable/tensorboard.html>`_'s
    ``add_scalar`` and ``add_hparams`` methods to mlflow. In this case, there's also
    no notion of an "epoch".

    Args:
        log_every_n_epoch: If specified, logs metrics once every `n` epochs. By default, metrics
            are logged after every epoch.
        log_every_n_step: If specified, logs batch metrics once every `n` training step.
            By default, metrics are not logged for steps. Note that setting this to 1 can cause
            performance issues and is not recommended. Metrics are logged against Lightning's global
            step number, and when multiple optimizers are used it is assumed that all optimizers
            are stepped in each training step.
        log_models: If ``True``, trained models are logged as MLflow model artifacts.
            If ``False``, trained models are not logged.
        log_datasets: If ``True``, dataset information is logged to MLflow Tracking.
            If ``False``, dataset information is not logged.
        disable: If ``True``, disables the PyTorch Lightning autologging integration.
            If ``False``, enables the PyTorch Lightning autologging integration.
        exclusive: If ``True``, autologged content is not logged to user-created fluent runs.
            If ``False``, autologged content is logged to the active fluent run, which may be
            user-created.
        disable_for_unsupported_versions: If ``True``, disable autologging for versions of
            pytorch and pytorch-lightning that have not been tested against this version
            of the MLflow client or are incompatible.
        silent: If ``True``, suppress all event logs and warnings from MLflow during PyTorch
            Lightning autologging. If ``False``, show all events and warnings during PyTorch
            Lightning autologging.
        registered_model_name: If given, each time a model is trained, it is registered as a
            new model version of the registered model with this name. The registered model is
            created if it does not already exist.
        extra_tags: A dictionary of extra tags to set on each managed run created by autologging.
        checkpoint: Enable automatic model checkpointing, this feature only supports
            pytorch-lightning >= 1.6.0.
        checkpoint_monitor: In automatic model checkpointing, the metric name to monitor if
            you set `model_checkpoint_save_best_only` to True.
        checkpoint_mode: one of {"min", "max"}. In automatic model checkpointing,
            if save_best_only=True, the decision to overwrite the current save file is made based on
            either the maximization or the minimization of the monitored quantity.
        checkpoint_save_best_only: If True, automatic model checkpointing only saves when
            the model is considered the "best" model according to the quantity
            monitored and previous checkpoint model is overwritten.
        checkpoint_save_weights_only: In automatic model checkpointing, if True, then
            only the model's weights will be saved. Otherwise, the optimizer states,
            lr-scheduler states, etc are added in the checkpoint too.
        checkpoint_save_freq: `"epoch"` or integer. When using `"epoch"`, the callback
            saves the model after each epoch. When using integer, the callback
            saves the model at end of this many batches. Note that if the saving isn't aligned to
            epochs, the monitored metric may potentially be less reliable (it
            could reflect as little as 1 batch, since the metrics get reset
            every epoch). Defaults to `"epoch"`.
        log_model_signatures: Whether to log model signature when `log_model` is True.

    .. code-block:: python
        :test:
        :caption: Example

        import os

        import lightning as L
        import torch
        from torch.nn import functional as F
        from torch.utils.data import DataLoader, Subset
        from torchmetrics import Accuracy
        from torchvision import transforms
        from torchvision.datasets import MNIST

        import mlflow.pytorch
        from mlflow import MlflowClient


        class MNISTModel(L.LightningModule):
            def __init__(self):
                super().__init__()
                self.l1 = torch.nn.Linear(28 * 28, 10)
                self.accuracy = Accuracy("multiclass", num_classes=10)

            def forward(self, x):
                return torch.relu(self.l1(x.view(x.size(0), -1)))

            def training_step(self, batch, batch_nb):
                x, y = batch
                logits = self(x)
                loss = F.cross_entropy(logits, y)
                pred = logits.argmax(dim=1)
                acc = self.accuracy(pred, y)

                # PyTorch `self.log` will be automatically captured by MLflow.
                self.log("train_loss", loss, on_epoch=True)
                self.log("acc", acc, on_epoch=True)
                return loss

            def configure_optimizers(self):
                return torch.optim.Adam(self.parameters(), lr=0.02)


        def print_auto_logged_info(r):
            tags = {k: v for k, v in r.data.tags.items() if not k.startswith("mlflow.")}
            artifacts = [f.path for f in MlflowClient().list_artifacts(r.info.run_id, "model")]
            print(f"run_id: {r.info.run_id}")
            print(f"artifacts: {artifacts}")
            print(f"params: {r.data.params}")
            print(f"metrics: {r.data.metrics}")
            print(f"tags: {tags}")


        # Initialize our model.
        mnist_model = MNISTModel()

        # Load MNIST dataset.
        train_ds = MNIST(
            os.getcwd(), train=True, download=True, transform=transforms.ToTensor()
        )
        # Only take a subset of the data for faster training.
        indices = torch.arange(32)
        train_ds = Subset(train_ds, indices)
        train_loader = DataLoader(train_ds, batch_size=8)

        # Initialize a trainer.
        trainer = L.Trainer(max_epochs=3)

        # Auto log all MLflow entities
        mlflow.pytorch.autolog()

        # Train the model.
        with mlflow.start_run() as run:
            trainer.fit(mnist_model, train_loader)

        # Fetch the auto logged parameters and metrics.
        print_auto_logged_info(mlflow.get_run(run_id=run.info.run_id))
    """
    try:
        import pytorch_lightning as pl
    except ImportError:
        pass
    else:
        from mlflow.pytorch._lightning_autolog import patched_fit

        safe_patch(
            FLAVOR_NAME, pl.Trainer, "fit", patched_fit, manage_run=True, extra_tags=extra_tags
        )

    try:
        import lightning as L
    except ImportError:
        pass
    else:
        from mlflow.pytorch._lightning_autolog import patched_fit

        safe_patch(
            FLAVOR_NAME, L.Trainer, "fit", patched_fit, manage_run=True, extra_tags=extra_tags
        )

    try:
        import torch.utils.tensorboard.writer
    except ImportError:
        pass
    else:
        from mlflow.pytorch._pytorch_autolog import (
            flush_metrics_queue,
            patched_add_event,
            patched_add_hparams,
            patched_add_summary,
        )

        safe_patch(
            FLAVOR_NAME,
            torch.utils.tensorboard.writer.FileWriter,
            "add_event",
            partial(patched_add_event, mlflow_log_every_n_step=log_every_n_step),
            manage_run=True,
            extra_tags=extra_tags,
        )
        safe_patch(
            FLAVOR_NAME,
            torch.utils.tensorboard.writer.FileWriter,
            "add_summary",
            patched_add_summary,
            manage_run=True,
            extra_tags=extra_tags,
        )
        safe_patch(
            FLAVOR_NAME,
            torch.utils.tensorboard.SummaryWriter,
            "add_hparams",
            patched_add_hparams,
            manage_run=True,
            extra_tags=extra_tags,
        )

        atexit.register(flush_metrics_queue)


if autolog.__doc__ is not None:
    autolog.__doc__ = autolog.__doc__.replace("MIN_REQ_VERSION", str(MIN_REQ_VERSION)).replace(
        "MAX_REQ_VERSION", str(MAX_REQ_VERSION)
    )


def load_checkpoint(model_class, run_id=None, epoch=None, global_step=None, kwargs=None):
    """
    If you enable "checkpoint" in autologging, during pytorch-lightning model
    training execution, checkpointed models are logged as MLflow artifacts.
    Using this API, you can load the checkpointed model.

    If you want to load the latest checkpoint, set both `epoch` and `global_step` to None.
    If "checkpoint_save_freq" is set to "epoch" in autologging,
    you can set `epoch` param to the epoch of the checkpoint to load specific epoch checkpoint.
    If "checkpoint_save_freq" is set to an integer in autologging,
    you can set `global_step` param to the global step of the checkpoint to load specific
    global step checkpoint.
    `epoch` param and `global_step` can't be set together.

    Args:
        model_class: The class of the training model, the class should inherit
            'pytorch_lightning.LightningModule'.
        run_id: The id of the run which model is logged to. If not provided,
            current active run is used.
        epoch: The epoch of the checkpoint to be loaded, if you set
            "checkpoint_save_freq" to "epoch".
        global_step: The global step of the checkpoint to be loaded, if
            you set "checkpoint_save_freq" to an integer.
        kwargs: Any extra kwargs needed to init the model.

    Returns:
        The instance of a pytorch-lightning model restored from the specified checkpoint.

    .. code-block:: python
        :caption: Example

        import mlflow

        mlflow.pytorch.autolog(checkpoint=True)

        model = MyLightningModuleNet()  # A custom-pytorch lightning model
        train_loader = create_train_dataset_loader()
        trainer = Trainer()

        with mlflow.start_run() as run:
            trainer.fit(model, train_loader)

        run_id = run.info.run_id

        # load latest checkpoint model
        latest_checkpoint_model = mlflow.pytorch.load_checkpoint(MyLightningModuleNet, run_id)

        # load history checkpoint model logged in second epoch
        checkpoint_model = mlflow.pytorch.load_checkpoint(MyLightningModuleNet, run_id, epoch=2)
    """
    with TempDir() as tmp_dir:
        downloaded_checkpoint_filepath = download_checkpoint_artifact(
            run_id=run_id, epoch=epoch, global_step=global_step, dst_path=tmp_dir.path()
        )
        return model_class.load_from_checkpoint(downloaded_checkpoint_filepath, **(kwargs or {}))


__all__ = [
    "autolog",
    "load_model",
    "save_model",
    "log_model",
    "get_default_pip_requirements",
    "get_default_conda_env",
    "load_checkpoint",
]

try:
    from mlflow.pytorch._lightning_autolog import MlflowModelCheckpointCallback  # noqa: F401

    __all__.append("MlflowModelCheckpointCallback")
except ImportError:
    # Swallow exception if pytorch-lightning is not installed.
    pass