full_finetune_distributed.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.

import sys
import time

from functools import partial
from typing import Any, Dict, Optional, Tuple, Union
from warnings import warn

import torch
from omegaconf import DictConfig, ListConfig

from torch import nn
from torch.distributed import init_process_group
from torch.distributed.fsdp import (
    CPUOffload,
    FullOptimStateDictConfig,
    FullStateDictConfig,
    FullyShardedDataParallel as FSDP,
    StateDictType,
)
from torch.optim import Optimizer
from torch.utils.data import DataLoader, DistributedSampler

from torchtune import config, modules, utils
from torchtune.datasets import ConcatDataset
from torchtune.recipe_interfaces import FTRecipeInterface
from torchtune.utils import DummyProfiler, PROFILER_KEY
from torchtune.utils.activations import apply_selective_activation_checkpointing

from tqdm import tqdm


log = utils.get_logger("DEBUG")


class FullFinetuneRecipeDistributed(FTRecipeInterface):
    """
    Full finetuning recipe for dense transformer-based LLMs such as Llama2. This recipe supports
    distributed training and can be run on a single node (1 to 8 GPUs).

    Features:
        - FSDP. Supported using PyTorch's FSDP APIs. DDP is currently not supported. Training on CPU
            is not supported.

        - Activation Checkpointing. This can be controlled using the ``activation_checkpointing``
            flag. Activation checkpointing helps reduce the memory footprint since we no longer keep
            activations in memory and instead recompute them during the backward pass. This is especially
            helpful for larger batch sizes when you're memory constrained. But these savings in memory
            come at the cost of training performance. In most cases training can slow-down quite a bit as
            a result of this activation recomputation.

        - Precision. Full fp32 and bf16 training are supported. Precision is controlled using the ``dtype``
            flag. When ``dtype=bf16``, all activations, gradients and optimizer states are in bfloat16. In
            most cases this should halve the memory footprint of full precision (fp32) training, without
            loss in model quality (will depend on the model, training data and other settings). For
            GPUs which do not support bfloat16, we fall back to fp32. Mixed precision training and fp16
            precision are currently not supported.

        - Gradient Accumulation. You can simulate larger batch sizes by accumulating gradients. This is
            controlled using the ``gradient_accumulation_steps`` flag.

                Total Batch Size = batch_size * number of GPUs * gradient accumulation steps.

            For example: with batch_size=1, nproc_per_node=2 and gradient_accumulation_steps=32 we get a
            total batch size of 64.

            Gradient accumulation is especially useful when you are memory constrained. In this case,
            accumulating gradients might give you better training speed than enabling activation
            checkpointing.

        - Checkpointing. Model weights are checkpointed both at the end of each epoch and at the end of
            training. Optimizer state and recipe state (seed, total_epochs, number of epochs run etc) are
            only saved at the end of a given epoch and used in case of resuming training.

            Resuming training is controlled by the ``resume_from_checkpoint`` flag. Mid-epoch checkpointing is
            currently not supported.

            For more details on the checkpointer, please take a look at
            our checkpointer deepdive (https://pytorch.org/torchtune/main/deep_dives/checkpointer.html).

        - Logging. Terminal, Disk, WandB and TensorBoard are all supported.

    For a full list of example configs for this recipe, run ``tune ls`` on the command line. Each config
    has example commands for how to kick-off training.

    Args:
        cfg (DictConfig): OmegaConf object parsed from yaml file

    Raises:
        ValueError: If ``dtype`` is set to fp16.
    """

    def __init__(self, cfg: DictConfig) -> None:

        self._device = utils.get_device(device=cfg.device)
        self._dtype = utils.get_dtype(cfg.dtype, device=self._device)

        if self._dtype == torch.float16:
            raise ValueError(
                "full fp16 training is not supported with this recipe. Please use bf16 or fp32 instead."
            )

        if (
            cfg.get("fsdp_cpu_offload", False)
            and cfg.optimizer.get("fused", False)
            and not utils.torch_version_ge("2.4.0")
        ):
            raise RuntimeError(
                "Using fused optimizer on CPU is only supported in PyTorch nightly."
            )

        # logging attributes
        self._output_dir = cfg.output_dir
        self._log_every_n_steps = cfg.get("log_every_n_steps", 1)
        self._log_peak_memory_stats = cfg.get("log_peak_memory_stats", False)

        # _is_rank_zero is used primarily for logging. In the future, the logger
        # should directly take care of this
        _, rank = utils.get_world_size_and_rank()
        self._is_rank_zero = rank == 0

        # Training cfg
        self._resume_from_checkpoint = cfg.resume_from_checkpoint
        self._gradient_accumulation_steps = cfg.gradient_accumulation_steps

        # These are public properties which are updated by the checkpoint loader
        # when ``resume_from_checkpoint`` is `True` or validated in tests
        self.seed = utils.set_seed(seed=cfg.seed)
        self.epochs_run = 0
        self.total_epochs = cfg.epochs
        self.max_steps_per_epoch = cfg.max_steps_per_epoch
        self.global_step = 0

    def load_checkpoint(self, cfg_checkpointer: DictConfig) -> Dict[str, Any]:
        """
        Extract the checkpoint state from file and validate. If resume_from_checkpoint
        is True, this also includes the recipe state.
        """
        self._checkpointer = config.instantiate(
            cfg_checkpointer,
            resume_from_checkpoint=self._resume_from_checkpoint,
        )
        checkpoint_dict = self._checkpointer.load_checkpoint()

        if self._resume_from_checkpoint:
            self._update_recipe_state(checkpoint_dict)
        return checkpoint_dict

    def _update_recipe_state(self, ckpt_dict: Dict[str, Any]) -> None:
        """
        Updates the recipe state from checkpoint.
        """
        try:
            self.epochs_run = ckpt_dict[utils.EPOCHS_KEY]

            # on mismatch, warn the user and prevent the override
            if self.seed != ckpt_dict[utils.SEED_KEY]:
                warn(
                    message=(
                        "Config value for seed does not match the checkpoint value, "
                        f"using the checkpoint value: {ckpt_dict[utils.SEED_KEY]}"
                    )
                )
                self.seed = ckpt_dict[utils.SEED_KEY]
            if self.max_steps_per_epoch != ckpt_dict[utils.MAX_STEPS_KEY]:
                warn(
                    message=(
                        "Config value for max_steps_per_epoch does not match the checkpoint value, "
                        f"using the checkpoint value: {ckpt_dict[utils.MAX_STEPS_KEY]}"
                    )
                )
                self.max_steps_per_epoch = ckpt_dict[utils.MAX_STEPS_KEY]

            # on mismatch, warn the user but allow the override
            if self.total_epochs != ckpt_dict[utils.TOTAL_EPOCHS_KEY]:
                warn(
                    message=(
                        "Config value for total_epochs does not match the checkpoint value, "
                        f"using the config value: {self.total_epochs}"
                    )
                )

        except KeyError as e:
            raise KeyError(
                "Checkpoint does not contain the required keys needed for updating recipe state. "
                "Are you sure you passed in the right recipe checkpoint?"
            ) from e

    def setup(self, cfg: DictConfig) -> None:
        """
        Sets up the recipe state correctly. This includes setting recipe attributes based
        on the ``resume_from_checkpoint`` flag.
        """
        if self._is_rank_zero:
            self._metric_logger = config.instantiate(cfg.metric_logger)

            # log config with parameter override
            self._metric_logger.log_config(cfg)

        ckpt_dict = self.load_checkpoint(cfg.checkpointer)

        # ``_setup_model`` handles initialization and loading the state dict. This method
        # should be called before ``_setup_optimizer`` since transforming the optimizer
        # state dict requires the model
        self._model = self._setup_model(
            cfg_model=cfg.model,
            enable_activation_checkpointing=cfg.enable_activation_checkpointing,
            memory_efficient_fsdp_wrap=cfg.get("memory_efficient_fsdp_wrap", False),
            fsdp_cpu_offload=cfg.get("fsdp_cpu_offload", False),
            model_state_dict=ckpt_dict[utils.MODEL_KEY],
            ac_mode=cfg.get("ac_mode", None),
            ac_option=cfg.get("ac_option", None),
        )

        self._tokenizer = config.instantiate(cfg.tokenizer)

        # _setup_optimizer should take in ckpt_dict only if training is resumed from
        # checkpoint. Transforming the opt state dict is handled by this method
        self._optimizer = self._setup_optimizer(
            cfg_optimizer=cfg.optimizer,
            opt_state_dict=ckpt_dict[utils.OPT_KEY]
            if self._resume_from_checkpoint
            else None,
        )

        self._loss_fn = config.instantiate(cfg.loss)

        # sampler and dataloader depend on the tokenizer and loss_fn and should be
        # setup after both of these are initialized
        self._sampler, self._dataloader = self._setup_data(
            cfg_dataset=cfg.dataset,
            shuffle=cfg.shuffle,
            batch_size=cfg.batch_size,
        )

        # Finally update the recipe state which can only be correctly set after all of the
        # other components have been initialized and updated.
        #
        # Number of training steps in each epoch depends on the number of batches produced
        # by the dataloader, the max_steps_per_epoch param set by the user and the
        # gradient_accumulation_steps param. This value is used for logging and tracking
        # training state. The computation should happen after the dataloader has been setup
        self._steps_per_epoch = (
            len(self._dataloader) // self._gradient_accumulation_steps
        )
        if (
            self.max_steps_per_epoch is not None
            and self.max_steps_per_epoch < self._steps_per_epoch
        ):
            self._steps_per_epoch = self.max_steps_per_epoch
        self.global_step = self.epochs_run * self._steps_per_epoch

        # Set up profiler, returns DummyProfiler (nullcontext object with no-op `step` method)
        # if cfg is missing profiler key or if `cfg.profiler.enabled = False`
        self._profiler = self._setup_profiler(cfg.get(PROFILER_KEY, None))

    def _setup_profiler(
        self, cfg_profiler: Optional[DictConfig] = None
    ) -> Union[torch.profiler.profile, DummyProfiler]:
        """
        Parses the `profiler` section of top-level `cfg` and sets up profiler

        Args:
            cfg_profiler (Optional[DictConfig]): ``profiler`` section of the top-level ``cfg`` (the main config passed to
                `recipe.main`). Default None.

        Returns:
            profiler: Union[torch.profiler.profile, DummyProfiler] - DummyProfiler is a nullcontext with no-op methods
            for `start`, `stop`, and `step` that can be used in place of `torch.profiler.profile` if profiler is not enabled such
            that the instrumented training loop does not need to be changed profiling is disabled.

        The profiler config can be provided in configs under the `profiler` key with the following layout:

        .. code-block:: yaml
            profiler:
                enabled: bool

                #Output directory of trace artifacts
                output_dir: str

            #`torch.profiler.ProfilerActivity` types to trace
            cpu: bool
            cuda: bool

                #Trace options
                profile_memory: bool
                with_stack: bool
                record_shapes: bool
                with_flops: bool

            # `torch.profiler.schedule` options:
            # wait_steps -> wait, warmup_steps -> warmup, active_steps -> active, num_cycles -> repeat
            wait_steps: int
            warmup_steps: int
            active_steps: int
            num_cycles: int
        """
        # Missing profiler section in config, assume disabled
        if cfg_profiler is None:
            cfg_profiler = DictConfig({"enabled": False})

        # Check that component is included and set correctly
        if cfg_profiler.get("_component_", None) is None:
            cfg_profiler["_component_"] = "torchtune.utils.setup_torch_profiler"
        else:
            assert (
                cfg_profiler.get("_component_")
                == "torchtune.utils.setup_torch_profiler"
            ), "Only torch profiler supported currently: component must be `torchtune.utils.setup_torch_profiler`"

        profiler, profiler_cfg = config.instantiate(cfg_profiler)

        if self._is_rank_zero:
            log.info(f" Profiler config after instantiation: {profiler_cfg}")

        return profiler

    def _setup_model(
        self,
        cfg_model: DictConfig,
        enable_activation_checkpointing: bool,
        memory_efficient_fsdp_wrap: bool,
        fsdp_cpu_offload: bool,
        model_state_dict: Dict[str, Any],
        ac_mode: Optional[str] = None,
        ac_option: Optional[int] = None,
    ) -> nn.Module:
        """
        Model initialization has some important considerations:
            a. To minimize GPU peak memory, we load the model on CPU with the right
               dtype. To ensure that we don't instantiate ``world_size`` number of models,
               we initialize on meta_device for all ranks other than rank 0.
            b. Rank 0 is also responsible for calling ``load_state_dict`` and loading the
               model weights from checkpoint.
            c. While wrapping the model with FSDP, we set ``sync_module_states``
               to TRUE and broadcast module params and buffers from rank 0.
            d. The ``device_id`` param ensures that the FSDP initialization happens on
               the correct device.
        """
        if self._is_rank_zero:
            log.info("FSDP is enabled. Instantiating Model on CPU for Rank 0 ...")
            init_start = time.perf_counter()

            with utils.set_default_dtype(self._dtype):
                model = config.instantiate(cfg_model)

            log.info(
                f"Model instantiation took {time.perf_counter() - init_start:.2f} secs"
            )

            # Load both the model weights. This should happen only on Rank 0
            model.load_state_dict(model_state_dict)

        else:
            # For non-zero ranks, load the model on meta device
            with utils.set_default_dtype(self._dtype), torch.device("meta"):
                model = config.instantiate(cfg_model)

        if self._dtype == torch.bfloat16:
            model = model.to(torch.bfloat16)

        # We currently have two versions of activation checkpointing in this recipe
        # for testing and BC purposes. ``enable_activation_checkpointing`` controls
        # the older version of AC and this behavior is unchanged
        # ac_mode and ac_option together control selective AC. This is only enabled
        # when these are set AND ``enable_activation_checkpointing`` is set to False
        # We'll clean this up as soon as testing of AC is complete
        ac_mode = ac_mode
        ac_option = ac_option

        if (not enable_activation_checkpointing) and (ac_mode is not None):
            apply_selective_activation_checkpointing(
                model,
                ac_mode,
                ac_option,
            )

        # Wrap the model with FSDP. This will ensure that the model is sharded
        # across all available GPUs.
        model = FSDP(
            module=model,
            auto_wrap_policy=utils.get_full_finetune_fsdp_wrap_policy(
                memory_efficient_fsdp_wrap=memory_efficient_fsdp_wrap,
                modules_to_wrap={modules.TransformerDecoderLayer},
            ),
            cpu_offload=CPUOffload(offload_params=fsdp_cpu_offload),
            sharding_strategy=torch.distributed.fsdp.ShardingStrategy.FULL_SHARD,
            device_id=self._device,
            # this recipe does not currently support mixed precision training
            mixed_precision=None,
            # Ensure we broadcast params and buffers from rank 0
            sync_module_states=True,
            # Initialize empty modules on all non-zero ranks
            param_init_fn=(
                lambda module: module.to_empty(
                    device=torch.device("cuda"), recurse=False
                )
                if not self._is_rank_zero
                else None
            ),
        )

        # Ensure no params and buffers are on meta device
        utils.validate_no_params_on_meta_device(model)

        # original activation checkpointing (full) - flip the condition above
        if enable_activation_checkpointing and ac_mode is None:
            utils.set_activation_checkpointing(
                model, auto_wrap_policy={modules.TransformerDecoderLayer}
            )

        if self._is_rank_zero:
            memory_stats = utils.get_memory_stats(device=self._device)
            utils.log_memory_stats(memory_stats)

        # synchronize before training begins
        torch.distributed.barrier()

        return model

    def _setup_optimizer(
        self, cfg_optimizer: DictConfig, opt_state_dict: Optional[Dict[str, Any]] = None
    ) -> Optimizer:
        """
        Set up the optimizer. This method also handles transforing the state dict
        for FSDP.
        """
        optimizer = config.instantiate(cfg_optimizer, self._model.parameters())

        if opt_state_dict:
            opt_state_dict = FSDP.optim_state_dict_to_load(
                self._model, optimizer, opt_state_dict
            )
            optimizer.load_state_dict(opt_state_dict)

        if self._is_rank_zero:
            log.info("Optimizer is initialized.")
        return optimizer

    def _setup_data(
        self,
        cfg_dataset: DictConfig,
        shuffle: bool,
        batch_size: int,
    ) -> Tuple[DistributedSampler, DataLoader]:
        """
        All data related setup happens here. Currently this recipe only supports the
        DistributedSamplers with Map-style Datasets which fit into memory. Other samplers,
        iterable datasets and streaming datasets are not supported.
        """
        world_size, rank = utils.get_world_size_and_rank()

        if isinstance(cfg_dataset, ListConfig):
            datasets = [
                config.instantiate(single_cfg_dataset, self._tokenizer)
                for single_cfg_dataset in cfg_dataset
            ]
            ds = ConcatDataset(datasets=datasets)
            packed = False
        else:
            ds = config.instantiate(cfg_dataset, self._tokenizer)
            packed = cfg_dataset.get("packed", False)

        sampler = DistributedSampler(
            ds,
            num_replicas=world_size,
            rank=rank,
            shuffle=shuffle,
            seed=0,
        )
        dataloader = DataLoader(
            dataset=ds,
            batch_size=batch_size,
            sampler=sampler,
            collate_fn=partial(
                utils.padded_collate,
                padding_idx=self._tokenizer.pad_id,
                ignore_idx=self._loss_fn.ignore_index,
            )
            if not packed
            else None,
        )

        if self._is_rank_zero:
            log.info("Dataset and Sampler are initialized.")

        return sampler, dataloader

    def save_checkpoint(self, epoch: int) -> None:
        """
        Save state dict to file. The recipe save_checkpoint method is responsible for
        correctly creating the checkpoint dict and passing to the checkpointer.
        """
        checkpoint_dict = {}

        # To prevent GPU memory from spiking during checkpoint save,
        # we consolidate the full model and optim state dicts on CPU for rank 0
        with FSDP.state_dict_type(
            self._model,
            StateDictType.FULL_STATE_DICT,
            FullStateDictConfig(offload_to_cpu=True, rank0_only=True),
            FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=True),
        ):
            cpu_state_dict = self._model.state_dict()
            opt_state_dict = FSDP.optim_state_dict(self._model, self._optimizer)

        # Now that we have the model and opt state dict, create the actual checkpoint dict
        # to be sent to the checkpointer and ultimately written to file
        if self._is_rank_zero:

            checkpoint_dict.update({utils.MODEL_KEY: cpu_state_dict})

            # if training is in-progress, checkpoint the optimizer state as well
            if epoch + 1 < self.total_epochs:
                checkpoint_dict.update(
                    {
                        utils.OPT_KEY: opt_state_dict,
                        utils.SEED_KEY: self.seed,
                        utils.EPOCHS_KEY: self.epochs_run,
                        utils.TOTAL_EPOCHS_KEY: self.total_epochs,
                        utils.MAX_STEPS_KEY: self.max_steps_per_epoch,
                    }
                )

            self._checkpointer.save_checkpoint(
                checkpoint_dict,
                epoch=epoch,
                intermediate_checkpoint=(epoch + 1 < self.total_epochs),
            )

    def train(self) -> None:
        """
        The core training loop. Supports training on subsets of the dataset using the
        ``max_steps_per_epoch``.
        """
        # clean up before training begins
        utils.cleanup_before_training()

        _, rank = utils.get_world_size_and_rank()

        # zero out the gradients before starting training
        self._optimizer.zero_grad()

        # Initialize tokens count and running loss (for grad accumulation)
        t0 = time.perf_counter()
        running_loss = 0
        num_tokens = 0

        self._profiler.start()
        # self.epochs_run should be non-zero when we're resuming from a checkpoint
        for curr_epoch in range(self.epochs_run, self.total_epochs):

            # Update the sampler to ensure data is correctly shuffled across epochs
            # in case shuffle is True
            self._sampler.set_epoch(curr_epoch)

            pbar = tqdm(total=self._steps_per_epoch, disable=not (rank == 0))
            for idx, batch in enumerate(self._dataloader):
                if (
                    self.max_steps_per_epoch is not None
                    and (idx // self._gradient_accumulation_steps)
                    == self.max_steps_per_epoch
                ):
                    break

                # Both are shape [b, s]
                tokens, labels = batch["tokens"], batch["labels"]
                # Get the attention mask and position ids from the dataset if they
                # exist. Currently, only sample packing in PackedDataset returns these
                mask = batch.get("mask", None)  # shape [b, s, s]
                input_pos = batch.get("input_pos", None)  # shape [b, s]

                tokens = tokens.to(self._device)
                num_tokens += tokens.numel()
                labels = labels.to(self._device)
                mask = mask.to(self._device) if mask is not None else None
                input_pos = (
                    input_pos.to(self._device) if input_pos is not None else None
                )

                logits = self._model(tokens, mask=mask, input_pos=input_pos)
                # Shift so that tokens < n predict n
                logits = logits[..., :-1, :].contiguous()
                labels = labels[..., 1:].contiguous()
                logits = logits.transpose(1, 2)
                # Compute loss
                loss = self._loss_fn(logits, labels)
                # free logits otherwise it peaks backward memory
                del logits

                loss = loss / self._gradient_accumulation_steps
                running_loss += loss
                loss.backward()

                # Step with optimizer
                if (idx + 1) % self._gradient_accumulation_steps == 0:
                    self._optimizer.step()
                    self._optimizer.zero_grad(set_to_none=True)

                    # Update the number of steps when the weights are updated
                    self.global_step += 1

                    loss_to_log = running_loss.item()
                    pbar.update(1)
                    pbar.set_description(
                        f"{curr_epoch + 1}|{self.global_step}|Loss: {loss_to_log}"
                    )

                    # Log per-step metrics
                    if (
                        self.global_step % self._log_every_n_steps == 0
                        and self._is_rank_zero
                    ):
                        time_per_step = time.perf_counter() - t0
                        log_dict = {
                            "loss": loss_to_log,
                            "lr": self._optimizer.param_groups[0]["lr"],
                            "tokens_per_second_per_gpu": num_tokens / time_per_step,
                        }
                        if self._log_peak_memory_stats:
                            log_dict.update(utils.get_memory_stats(device=self._device))
                        self._metric_logger.log_dict(
                            log_dict,
                            step=self.global_step,
                        )

                    # Reset running stats for the next step
                    running_loss = 0
                    num_tokens = 0
                    t0 = time.perf_counter()

                    # Step profiler
                    # Note that this is called within gradient accumulation block, hence
                    # will include multiple forward / backward passes if gradient accumulation > 1
                    self._profiler.step()

            self.epochs_run += 1
            self.save_checkpoint(epoch=curr_epoch)

        self._profiler.stop()

    def cleanup(self) -> None:
        if self._is_rank_zero:
            self._metric_logger.close()
        torch.distributed.destroy_process_group()


@config.parse
def recipe_main(cfg: DictConfig) -> None:
    """
    Entry point for the recipe.

    Configurable parameters are read in the following order:
        - Parameters specified in config (see available configs through ``tune ls``)
        - Overwritten by arguments from the command-line
    """
    if not utils.is_distributed():
        raise RuntimeError(
            "Distributed finetune recipe should be run via a distributed launcher."
            "If using tune CLI, please specify --nnodes 1 and --nproc_per_node [num_gpus]"
        )

    init_process_group(backend="gloo" if cfg.device == "cpu" else "nccl")
    if cfg.get("fsdp_cpu_offload", False):
        # Utilize all available CPU cores for intra-op parallelism. This provides ~2x
        # speed up when benchmarking fused AdamW on CPU
        utils.set_torch_num_threads()

    config.log_config(recipe_name="FullFinetuneRecipeDistributed", cfg=cfg)

    recipe = FullFinetuneRecipeDistributed(cfg=cfg)
    recipe.setup(cfg=cfg)
    recipe.train()
    recipe.cleanup()


if __name__ == "__main__":
    sys.exit(recipe_main())