nemo/utils/callbacks/cuda_graph.py

# Copyright (c) 2023, NVIDIA CORPORATION.  All rights reserved.
#
# 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.

# CUDAGraphCallback is a full iteration CUDA graph callback designed for
# models with PyTorch Lightning first, this has been tested with Stable
# Diffusion right now.
#
# Prerequisites for this callback:
# 1. Capturable: user has to make sure (almost) all the host & device
#    synchronizations are removed, some of the syncs regarding logging
#    of metrics introduced by PyTorch Lightning itself have been removed
#    by this callback. This ensures the graph can be captured.
# 2. Topology: user has to make sure there's no dynamic control flow
#    within the iteration. Please use APEX alternatives for building
#    blocks that contain dynamic control flow, e.g. gradient clipping.
#    Otherwise the captured graph can run, but may raise silent failure,
#    e.g. NaN loss.
# 3. Parameters: user has to make sure pointers involved in the graph
#    capturing range don't change across iterations. In this case users
#    have to ensure that data is copied to static tensors. Otherwise this
#    can also lead to silent failure.

import os
import time
from dataclasses import dataclass
from types import MethodType
from typing import Any, Dict

import lightning.pytorch as pl
import torch
from lightning.pytorch import LightningModule
from lightning.pytorch.callbacks import Callback
from lightning.pytorch.loops.optimization.automatic import ClosureResult
from lightning.pytorch.trainer.connectors.logger_connector.result import _ResultCollection, _ResultMetric
from lightning.pytorch.utilities import CombinedLoader, rank_zero_info
from lightning.pytorch.utilities.signature_utils import is_param_in_hook_signature
from lightning.pytorch.utilities.types import STEP_OUTPUT
from torch.nn.parallel import DistributedDataParallel

__all__ = ["CUDAGraphCallback"]


def struct_copy_one(src):
    if isinstance(src, tuple):
        return tuple(struct_copy_one(i) for i in src)
    elif isinstance(src, list):
        return list(struct_copy_one(i) for i in src)
    elif isinstance(src, dict):
        return {k: struct_copy_one(src[k]) for k in src}
    elif isinstance(src, torch.Tensor):
        return src.clone().detach().cuda()
    else:
        return src


def struct_copy_two(tgt, src):
    if isinstance(src, tuple):
        raise Exception(f"Unsupported copy for tuple yet: {type(src)}")
    elif isinstance(src, list):
        for i in range(len(src)):
            if isinstance(src[i], (tuple, list, dict, torch.Tensor)):
                struct_copy_two(tgt[i], src[i])
            else:
                tgt[i] = src[i]
    elif isinstance(src, dict):
        for k in src:
            if isinstance(src[k], (tuple, list, dict, torch.Tensor)):
                struct_copy_two(tgt[k], src[k])
            else:
                tgt[k] = src[k]
    elif isinstance(src, torch.Tensor):
        tgt.copy_(src, non_blocking=True)
    else:
        raise Exception(f"Expect top-level as container type but got: {type(src)}")


class StaticBufferLoader:
    """Load data to static buffers."""

    def __init__(self, loader):
        self.loader = loader
        self.stream = torch.cuda.Stream()
        self.static = None

    def __iter__(self):
        for inputs in self.loader:
            if self.static is None:
                with torch.cuda.stream(self.stream):
                    self.static = struct_copy_one(inputs)

            with torch.cuda.stream(self.stream):
                struct_copy_two(self.static, inputs)
            torch.cuda.current_stream().wait_stream(self.stream)
            yield self.static

    def __len__(self):
        return len(self.loader)


def get_lr(lr_scheduler):
    lrs = lr_scheduler.__orig_get_lr__()
    if not hasattr(lr_scheduler, "static_lrs"):
        lr_scheduler.static_lrs = lrs
    for i in range(len(lrs)):
        lr_scheduler.static_lrs[i].copy_(lrs[i])
    return lr_scheduler.static_lrs


def zero_grad(optimizer, *args, **kwargs):
    # We invoke zero_grad before graph capturing.
    if torch.cuda.is_current_stream_capturing():
        rank_zero_info("CUDAGraphCallback: set optimizer.zero_grad as nop during graph capturing.")
    else:
        optimizer.__orig_zero_grad__(*args, **kwargs)


def to_tensor(self, value, name):
    # Log metrics in PyTorch Lightning often invokes CPU & GPU synchronizations. Here
    # we implement smart metrics to avoid those synchronizations.
    # Refer to: https://github.com/Lightning-AI/pytorch-lightning/blob/2.0.7/src/lightning/pytorch/core/module.py#L615
    value = value.clone().detach() if isinstance(value, torch.Tensor) else torch.tensor(value)
    if not torch.numel(value) == 1:
        raise ValueError(
            f"`self.log({name}, {value})` was called, but the tensor must have a single element."
            f" You can try doing `self.log({name}, {value}.mean())`"
        )
    value = value.squeeze()
    return value


def get_optimizer_step(state):
    def optimizer_step(
        self,
        epoch,
        batch_idx,
        optimizer,
        optimizer_closure=None,
    ) -> None:
        # Not all optimizer supports set_to_none.
        if not hasattr(optimizer, "support_set_to_none"):
            optimizer.support_set_to_none = is_param_in_hook_signature(
                optimizer.zero_grad, "set_to_none", explicit=True
            )
        if optimizer.support_set_to_none:
            zero_grad_kwargs = {"set_to_none": True}
        else:
            zero_grad_kwargs = {}

        if 0 <= state.current_iteration < state.capture_iteration or state.capture_iteration < 0:
            state.stream.wait_stream(torch.cuda.current_stream())
            with torch.cuda.stream(state.stream):
                optimizer.zero_grad(**zero_grad_kwargs)
                self.__orig_optimizer_step__(
                    epoch,
                    batch_idx,
                    optimizer,
                    optimizer_closure=optimizer_closure,
                )
            torch.cuda.current_stream().wait_stream(state.stream)

        if state.current_iteration == state.capture_iteration:
            torch.cuda.synchronize()
            # Sleep for one second to let environment stable
            time.sleep(1)
            rank_zero_info("CUDAGraphCallback: capturing CUDA graph for module %s.", self.__class__.__name__)
            with torch.cuda.graph(state.graph, stream=state.stream, capture_error_mode="global"):
                # PyTorch CUDA graph doc for whole-network capturing mentions:
                #
                #   Sets grads to None before capture, so backward() will create
                #   .grad attributes with allocations from the graph's private pool
                #
                # But it's not necessary, and it can lead to CUDA kernels inside
                # `zero_grad()` being not captured.
                optimizer.zero_grad(**zero_grad_kwargs)
                self.__orig_optimizer_step__(
                    epoch,
                    batch_idx,
                    optimizer,
                    optimizer_closure=optimizer_closure,
                )
            torch.cuda.synchronize()

        # Graph replay and reconstruct missing result
        if state.current_iteration >= state.capture_iteration >= 0:
            state.graph.replay()
            optimizer_closure._result = ClosureResult.from_training_step_output(state.output)

        # If something is not capturable, try to put it there, e.g. `self.log()`.
        if hasattr(self, "non_cuda_graph_capturable"):
            self.non_cuda_graph_capturable()

        state.current_iteration += 1

    return optimizer_step


def get_training_step(state):
    def training_step(self, batch):
        results = self.__orig_training_step__(batch)
        if state.output is None:
            state.output = struct_copy_one(results)

        # Copy results to static buffer to rebuild states required by PL.
        with torch.no_grad():
            struct_copy_two(state.output, results)
        return results

    return training_step


def get_amp_autocast_init(state):
    def amp_autocast_init(self, *args, **kwargs):
        if "cache_enabled" not in kwargs:
            kwargs["cache_enabled"] = False
        if state.current_iteration == 0:
            rank_zero_info("CUDAGraphCallback: disable autocast cache.")
        return self.__orig_init__(*args, **kwargs)

    return amp_autocast_init


def get_ddp_init(state):
    def init(self, *args, **kwargs):
        rank_zero_info("CUDAGraphCallback: init DDP on side stream.")
        with torch.cuda.stream(state.stream):
            self.__orig_init__(*args, **kwargs)

    return init


@dataclass
class CUDAGraphState:
    current_iteration: int = 0
    capture_iteration: int = -1  # -1 to disable
    stream: torch.cuda.Stream = None
    graph: torch.cuda.CUDAGraph = None
    output: Any = None  # static forward output


class CUDAGraphCallback(Callback):
    """Full iteration CUDA graph callback.

    Dataloader and LR scheduler are not included in the CUDA graph with this callback.
    """

    def __init__(self, capture_iteration=-1):
        super().__init__()

        # Required by CUDA graph with DDP
        # Ref: https://pytorch.org/docs/stable/notes/cuda.html#usage-with-distributeddataparallel
        if 0 <= capture_iteration <= 11:
            raise Exception("Warmup must run at least 11 DDP-enabled eager iterations before capture.")
        if torch.distributed.is_initialized():
            raise Exception("CUDAGraphCallback should be initialized before process group.")
        os.environ["TORCH_NCCL_ASYNC_ERROR_HANDLING"] = "0"

        self.state = CUDAGraphState(capture_iteration=capture_iteration)

    def setup(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule", stage: str) -> None:
        """Called when fit, validate, test, predict, or tune begins."""
        if self.state.capture_iteration < 0:
            return

        # Hack to avoid CUDA graph issue with AMP, PyTorch Lightning doesn't support
        # changing autocast arguments for now.
        # https://github.com/pytorch/pytorch/blob/v1.13.1/torch/cuda/graphs.py#L234
        torch.autocast.__orig_init__ = torch.autocast.__init__
        torch.autocast.__init__ = get_amp_autocast_init(self.state)

        # Before full-backward capture, DDP must be constructed in a side-stream context.
        # We've merged the change that init DDP on side stream to PyTorch Lightning V2,
        # but not all user defined strategy init DDP on side stream.
        DistributedDataParallel.__orig_init__ = DistributedDataParallel.__init__
        DistributedDataParallel.__init__ = get_ddp_init(self.state)

    def teardown(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule", stage: str) -> None:
        """Called when fit, validate, test, predict, or tune ends."""
        if self.state.capture_iteration < 0:
            return

        torch.autocast.__init__ = torch.autocast.__orig_init__
        del torch.autocast.__orig_init__

        DistributedDataParallel.__init__ = DistributedDataParallel.__orig_init__
        del DistributedDataParallel.__orig_init__

    def on_fit_start(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule") -> None:
        """Called when fit begins."""
        if self.state.capture_iteration < 0:
            return

        if is_param_in_hook_signature(pl_module.training_step, "dataloader_iter", explicit=True):
            raise Exception(
                "Found `dataloader_iter` argument in the `training_step`. This is "
                "not supported by full iteration CUDA graph capturing yet since "
                "dataloader will be within the CUDA graph capturing range.\n"
                "Try to change `dataloader_iter` to `batch` and remove "
                "`next(dataloader_iter)` from `training_step`."
            )

        # Now that CUDA device has been set, we can init stream and graph now
        self.state.stream = torch.cuda.Stream()
        self.state.graph = torch.cuda.CUDAGraph()

    def on_fit_end(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule") -> None:
        """Called when fit ends."""
        if self.state.capture_iteration < 0:
            return

    def on_train_start(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule") -> None:
        """Called when the train begins."""
        if self.state.capture_iteration < 0:
            return

        # Ensure training dataloader loads data to static buffer
        dataloader = trainer.fit_loop._combined_loader._iterables
        assert isinstance(
            dataloader, torch.utils.data.dataloader.DataLoader
        ), f"Expect Dataloader type but got {type(dataloader)}"
        static_loader = StaticBufferLoader(dataloader)
        _mode = trainer.fit_loop._combined_loader._mode
        combined_loader = CombinedLoader(static_loader, mode=_mode)
        trainer.fit_loop.__orig_combined_loader__ = trainer.fit_loop._combined_loader
        trainer.fit_loop._combined_loader = combined_loader
        trainer.fit_loop._data_fetcher.setup(trainer.fit_loop._combined_loader)
        iter(trainer.fit_loop._data_fetcher)

        # Warn if `optimizer.zero_grad()` invoked during graph capturing
        for optimizer in trainer.optimizers:
            assert isinstance(optimizer, torch.optim.Optimizer), f"Expect Optimizer type but got {type(optimizer)}"
            optimizer.__orig_zero_grad__ = optimizer.zero_grad
            optimizer.zero_grad = MethodType(zero_grad, optimizer)

        # Ensure LR scheduler writes to static buffer
        # We don't include LR scheduler in the full CUDA graph for now since
        # its overhead is very small.
        for config in trainer.lr_scheduler_configs:
            assert isinstance(
                config.scheduler, torch.optim.lr_scheduler._LRScheduler
            ), f"Expect _LRScheduler type but got {type(config.scheduler)}"
            config.scheduler.__orig_get_lr__ = config.scheduler.get_lr
            config.scheduler.get_lr = MethodType(get_lr, config.scheduler)

        # Use smart metrics to avoid syncs
        LightningModule.__orig_to_tensor__ = LightningModule._LightningModule__to_tensor
        LightningModule._LightningModule__to_tensor = to_tensor

        # Save model outputs to static buffer for PL states reconstruct
        pl_module.__orig_training_step__ = pl_module.training_step
        training_step = get_training_step(self.state)
        pl_module.training_step = MethodType(training_step, pl_module)

        # Capture CUDA graph from model forward propagation to optimizer step
        pl_module.__orig_optimizer_step__ = pl_module.optimizer_step
        optimizer_step = get_optimizer_step(self.state)
        pl_module.optimizer_step = MethodType(optimizer_step, pl_module)

    def on_train_end(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule") -> None:
        """Called when the train ends."""
        if self.state.capture_iteration < 0:
            return

        trainer.fit_loop._combined_loader = trainer.fit_loop.__orig_combined_loader__
        trainer.fit_loop._data_fetcher.setup(trainer.fit_loop._combined_loader)
        iter(trainer.fit_loop._data_fetcher)
        del trainer.fit_loop.__orig_combined_loader__

        for optimizer in trainer.optimizers:
            optimizer.zero_grad = optimizer.__orig_zero_grad__
            del optimizer.__orig_zero_grad__

        for config in trainer.lr_scheduler_configs:
            config.scheduler.get_lr = config.scheduler.__orig_get_lr__
            del config.scheduler.__orig_get_lr__

        LightningModule._LightningModule__to_tensor = LightningModule.__orig_to_tensor__
        del LightningModule.__orig_to_tensor__

        pl_module.training_step = pl_module.__orig_training_step__
        del pl_module.__orig_training_step__

        pl_module.optimizer_step = pl_module.__orig_optimizer_step__
        del pl_module.__orig_optimizer_step__

    def on_train_epoch_start(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule") -> None:
        """Called when the train epoch begins."""
        pass

    def on_train_epoch_end(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule") -> None:
        """Called when the train epoch ends.

        To access all batch outputs at the end of the epoch, either:

        1. Implement `training_epoch_end` in the `LightningModule` and access outputs via the module OR
        2. Cache data across train batch hooks inside the callback implementation to post-process in this hook.
        """
        pass

    def on_train_batch_start(
        self, trainer: "pl.Trainer", pl_module: "pl.LightningModule", batch: Any, batch_idx: int
    ) -> None:
        """Called when the train batch begins."""
        pass

    def on_train_batch_end(
        self, trainer: "pl.Trainer", pl_module: "pl.LightningModule", outputs: STEP_OUTPUT, batch: Any, batch_idx: int
    ) -> None:
        """Called when the train batch ends.

        Note:
            The value ``outputs["loss"]`` here will be the normalized value w.r.t ``accumulate_grad_batches`` of the
            loss returned from ``training_step``.
        """
        pass

    def on_save_checkpoint(
        self, trainer: "pl.Trainer", pl_module: "pl.LightningModule", checkpoint: Dict[str, Any]
    ) -> None:
        r"""
        Called when saving a checkpoint to give you a chance to store anything else you might want to save.

        Args:
            trainer: the current :class:`~lightning.pytorch.trainer.Trainer` instance.
            pl_module: the current :class:`~lightning.pytorch.core.module.LightningModule` instance.
            checkpoint: the checkpoint dictionary that will be saved.
        """
        # Since we've add bound method to optimizer and lr_scheduler, it can lead to more
        # CUDA tensors passed to consumer process unexpectedly.
        if "optimizer_states" in checkpoint:
            for optimizer_state in checkpoint["optimizer_states"]:
                for k in list(optimizer_state.keys()):
                    v = optimizer_state[k]
                    if isinstance(v, MethodType) and hasattr(v, "__self__"):
                        del optimizer_state[k]
        if "lr_schedulers" in checkpoint:
            for lr_scheduler in checkpoint["lr_schedulers"]:
                for k in list(lr_scheduler.keys()):
                    v = lr_scheduler[k]
                    if isinstance(v, MethodType) and hasattr(v, "__self__"):
                        del lr_scheduler[k]