Model-Optimizer/modelopt/torch/opt/plugins/transformers.py

# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""ModelOpt plugin for enabling automatic save/restore of ModelOpt state for HuggingFace models."""

import os
import types
from contextlib import contextmanager

import torch
import transformers
from transformers import PreTrainedModel, Trainer, TrainerCallback
from transformers import modeling_utils as tf_modeling_utils

from modelopt.torch.utils import report_memory

from ..conversion import ModeloptStateManager
from .huggingface import (
    _new_save_pretrained,
    _patch_model_init_for_modelopt,
    enable_huggingface_checkpointing,
    register_for_patching,
)

__all__ = ["ModelOptHFTrainer"]


@contextmanager
def _undo_torch_init_override_by_transformers():
    if not hasattr(tf_modeling_utils, "TORCH_INIT_FUNCTIONS"):
        yield
        return
    # transformers override weight initialization during model instantiation for faster loading;
    # this leads to a secondary bug causing fx symbolic tracing to fail (torch does not allow
    # overriding torch.nn.init functions - fx tracing asserts that this does not happen and fails)
    # correct fx symbolic tracing is needed for NAS/Pruned model restoration
    # lets restore the original init functions before modelopt restore so that tracing works during nas restore
    # weight initialization is anyways done, so this wont affect performance
    modelopt_reverted_torch_init_funcs = {}
    for name, init_func in tf_modeling_utils.TORCH_INIT_FUNCTIONS.items():
        torch_init_func = getattr(torch.nn.init, name)
        # Check if the init function has been overridden by transformers
        if id(torch_init_func) != id(init_func):
            modelopt_reverted_torch_init_funcs[name] = torch_init_func
            setattr(torch.nn.init, name, init_func)

    yield

    for name, init_func in modelopt_reverted_torch_init_funcs.items():
        setattr(torch.nn.init, name, init_func)


def _new_from_pretrained(cls, /, pretrained_model_name_or_path, *args, **kwargs):
    """Patch for `cls.from_pretrained` method to restore ModelOpt state."""
    if kwargs.get("tp_plan") is not None or (
        kwargs.get("device_map") == "auto" and os.environ.get("WORLD_SIZE")
    ):
        assert transformers.__version__ >= "4.52.0", (
            "Tensor parallelism with ModelOpt requires transformers >= 4.52.0"
        )

    with _patch_model_init_for_modelopt(
        cls, pretrained_model_name_or_path, extra_context=_undo_torch_init_override_by_transformers
    ):
        model = types.MethodType(cls._modelopt_cache["from_pretrained"].__func__, cls)(
            pretrained_model_name_or_path, *args, **kwargs
        )

    return model


def _new_from_config(cls, /, config, **kwargs):
    """Patch for `cls.from_config` method to restore ModelOpt state."""
    with _patch_model_init_for_modelopt(
        cls, config._name_or_path, extra_context=_undo_torch_init_override_by_transformers
    ):
        model = types.MethodType(cls._modelopt_cache["_from_config"].__func__, cls)(
            config, **kwargs
        )
    return model


def _save_pretrained_with_checks(self, save_directory, *args, **kwargs):
    if getattr(self, "_tp_size", None) is not None and ModeloptStateManager.is_converted(self):
        raise NotImplementedError(
            "ModelOpt does not support saving tensor parallel sharded Huggingface transformer models yet. "
        )
    return _new_save_pretrained(self, save_directory, *args, **kwargs)


# [Fix for huggingface bug] deepspeed zero3 training backend only loads params into the model from
# state_dict, but not buffers. So lets explicitly load the buffers into the model from state_dict.
def _load_params_and_buffers_into_zero3_model(model_to_load, state_dict):
    buffer_names = [name for name, _ in model_to_load.named_buffers()]
    buffer_state_dict = {k: v for k, v in state_dict.items() if k in buffer_names}
    model_to_load.load_state_dict(buffer_state_dict, strict=False)
    return tf_modeling_utils._modelopt_cache["_load_state_dict_into_zero3_model"](
        model_to_load, state_dict
    )


pretrained_model_patch_methods = [
    ("from_pretrained", classmethod(_new_from_pretrained)),
    # We need to patch _from_config of PreTrainedModel; from_config is a private method in _BaseAutoModelClass and
    # patching it is more complex
    ("_from_config", classmethod(_new_from_config)),
    ("save_pretrained", _save_pretrained_with_checks),
]

register_for_patching("transformers", PreTrainedModel, pretrained_model_patch_methods)
register_for_patching(
    "transformers",
    tf_modeling_utils,
    [("_load_state_dict_into_zero3_model", _load_params_and_buffers_into_zero3_model)],
)


def _report_memory(msg):
    if not torch.cuda.is_available():
        return
    if torch.distributed.is_initialized() and torch.distributed.get_rank() == 0:
        report_memory(msg + ":", device=torch.cuda.current_device())
    else:
        for device in range(torch.cuda.device_count()):
            report_memory(f"{msg}, device={device}:", device=device)


class _MemoryReportCallback(TrainerCallback):
    def on_step_end(self, args, state, control, **kwargs):
        if state.global_step == 1:
            _report_memory("Memory usage at training step 1")

    def on_evaluate(self, args, state, control, **kwargs):
        if state.global_step <= 1:
            _report_memory("Memory usage at evaluation")


class ModelOptHFTrainer(Trainer):
    """A drop-in replacement of HuggingFace's Trainer for ModelOpt.

    This class adds extra utilities for ModelOpt checkpointing and memory reporting.
    """

    def __init__(self, *args, **kwargs):
        """Initialize."""
        enable_huggingface_checkpointing()
        super().__init__(*args, **kwargs)
        self.add_callback(_MemoryReportCallback())