train.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 contextlib
import gc
import os
import time

from dataclasses import dataclass, field
from datetime import timedelta
from io import BytesIO
from timeit import default_timer as timer
from typing import Any, Dict, List

import numpy as np

import torch
import torch.nn.functional as F
from torch.distributed import destroy_process_group
from torch.distributed.checkpoint.stateful import Stateful
from torch.distributed.elastic.multiprocessing.errors import record
from torch.distributed.tensor.parallel import loss_parallel

from torchtitan.checkpoint import CheckpointManager
from torchtitan.config_manager import JobConfig
from torchtitan.datasets import build_hf_data_loader, create_tokenizer
from torchtitan.float8_linear import (
    maybe_build_fp8_linear,
    maybe_precompute_fp8_dynamic_scale_for_fsdp,
)
from torchtitan.logging_utils import init_logger, logger
from torchtitan.lr_scheduling import get_lr_schedulers
from torchtitan.metrics import build_gpu_memory_monitor, build_metric_logger
from torchtitan.models import model_name_to_cls, model_name_to_tokenizer, models_config
from torchtitan.parallelisms import (
    models_parallelize_fns,
    models_pipelining_fns,
    ParallelDims,
)
from torchtitan.parallelisms.pipelining_utils import build_pipeline_schedule
from torchtitan.profiling import maybe_enable_memory_snapshot, maybe_enable_profiling
from torchtitan.utils import (
    Color,
    dist_max,
    dist_mean,
    get_metrics_rank,
    get_num_flop_per_token,
    get_num_params,
    get_peak_flops,
    init_distributed,
    NoColor,
    set_pg_timeouts,
)


@dataclass
class TrainState(Stateful):
    step: int = 0
    global_avg_losses: List[float] = field(default_factory=list)
    global_max_losses: List[float] = field(default_factory=list)
    log_steps: List[int] = field(default_factory=list)

    def state_dict(self) -> Dict[str, Any]:
        # Only checkpoint global_avg_losses and global_max_losses per log frequency
        # to avoid sync overhead in every iteration.
        global_avg_losses_bytes = BytesIO()
        torch.save(self.global_avg_losses, global_avg_losses_bytes)
        global_max_losses_bytes = BytesIO()
        torch.save(self.global_max_losses, global_max_losses_bytes)
        log_steps_bytes = BytesIO()
        torch.save(self.log_steps, log_steps_bytes)
        return {
            "step": torch.tensor(self.step, dtype=torch.int32),
            "global_avg_losses": global_avg_losses_bytes,
            "global_max_losses": global_max_losses_bytes,
            "log_steps": log_steps_bytes,
        }

    def load_state_dict(self, state_dict) -> None:
        self.step = state_dict["step"].item()
        state_dict["global_avg_losses"].seek(0)
        self.global_avg_losses = torch.load(
            state_dict["global_avg_losses"], weights_only=False
        )
        state_dict["global_max_losses"].seek(0)
        self.global_max_losses = torch.load(
            state_dict["global_max_losses"], weights_only=False
        )
        state_dict["log_steps"].seek(0)
        self.log_steps = torch.load(state_dict["log_steps"], weights_only=False)


def build_optimizers(model_parts, job_config: JobConfig):
    """Wrap one optimizer per model part in an OptimizersContainer which provides a single
    step() and zero_grad() method for all the child optimizers.
    """

    def _build_optimizer(model):
        name = job_config.optimizer.name
        lr = job_config.optimizer.lr
        fused = job_config.optimizer.fused

        # Common parameters for both optimizers
        optimizer_kwargs = {
            "lr": lr,
            "betas": (0.9, 0.95),
            "weight_decay": 0.1,
            "fused": fused,
            "foreach": not fused,
        }
        if name == "Adam":
            # TODO: make the optimizer options configurable by toml/cmd args
            optimizer = torch.optim.Adam(model.parameters(), **optimizer_kwargs)
        elif name == "AdamW":
            optimizer = torch.optim.AdamW(model.parameters(), **optimizer_kwargs)
        else:
            raise NotImplementedError(f"Optimizer {name} not added.")

        return optimizer

    class OptimizersContainer:
        """Util for calling step/zero_grad on multiple optimizers needed for virtual pipeline stages"""

        def __init__(self, optimizers):
            self.optimizers = optimizers

        def step(self):
            for optimizer in self.optimizers:
                optimizer.step()

        def zero_grad(self):
            for optimizer in self.optimizers:
                optimizer.zero_grad()

    return OptimizersContainer([_build_optimizer(model) for model in model_parts])


def get_train_context(enable_loss_parallel: bool, enable_compiled_autograd: bool):
    @contextlib.contextmanager
    def context():
        with contextlib.ExitStack() as stack:
            if enable_loss_parallel:
                stack.enter_context(loss_parallel())
            if enable_compiled_autograd:
                stack.enter_context(
                    torch._dynamo.utils.maybe_enable_compiled_autograd(True)
                )

            yield

    return context


# Enable debug tracing on failure: https://pytorch.org/docs/stable/elastic/errors.html
@record
def main(job_config: JobConfig):
    init_logger()
    logger.info(f"Starting job: {job_config.job.description}")

    # used for colorful printing
    color = Color if job_config.metrics.enable_color_printing else NoColor

    # take control of garbage collection to avoid stragglers
    _gc_freq = job_config.training.gc_freq
    gc.disable()
    gc.collect(1)

    # init distributed
    world_size = int(os.environ["WORLD_SIZE"])
    parallel_dims = ParallelDims(
        dp=job_config.training.data_parallel_degree,
        tp=job_config.training.tensor_parallel_degree,
        pp=job_config.experimental.pipeline_parallel_degree,
        world_size=world_size,
        enable_loss_parallel=job_config.training.enable_loss_parallel,
        dp_type=job_config.training.data_parallel_type,
    )
    device = torch.device(f"cuda:{int(os.environ['LOCAL_RANK'])}")
    torch.cuda.set_device(device)
    init_distributed(job_config)

    # build meshes
    world_mesh = parallel_dims.build_mesh(device_type="cuda")
    if parallel_dims.dp_enabled:
        dp_mesh = world_mesh["dp"]
        dp_degree = dp_mesh.size()
        dp_rank = dp_mesh.get_local_rank()
    else:
        dp_degree, dp_rank = 1, 0

    if parallel_dims.pp_enabled:
        pp_mesh = world_mesh["pp"]

    model_name = job_config.model.name

    # build tokenizer
    tokenizer_type = model_name_to_tokenizer[model_name]
    tokenizer = create_tokenizer(tokenizer_type, job_config.model.tokenizer_path)

    # build dataloader
    data_loader = build_hf_data_loader(
        job_config.training.dataset,
        job_config.training.dataset_path,
        tokenizer,
        job_config.training.batch_size,
        job_config.training.seq_len,
        dp_degree,
        dp_rank,
    )

    train_context = get_train_context(
        parallel_dims.loss_parallel_enabled,
        job_config.experimental.enable_compiled_autograd,
    )

    # loss fn can be shared by pipeline-parallel or non-pp execution
    def loss_fn(pred, labels):
        return F.cross_entropy(pred.flatten(0, 1), labels.flatten(0, 1))

    # build model (using meta init)
    model_cls = model_name_to_cls[model_name]
    model_config = models_config[model_name][job_config.model.flavor]
    # set the model configs from training inputs:
    # 1. norm type to decide which norm layer to use
    # 2. vocab size from tokenizer
    # 3. max_seq_len base on inputs
    model_config.norm_type = job_config.model.norm_type
    model_config.vocab_size = tokenizer.n_words
    model_config.max_seq_len = job_config.training.seq_len

    logger.info(f"Building {model_name} {job_config.model.flavor} with {model_config}")
    with torch.device("meta"):
        whole_model = model_cls.from_model_args(model_config)

    # swap to Float8Linear base on fp8 config
    maybe_build_fp8_linear(whole_model, job_config, parallel_dims.dp_enabled)

    # log model size
    model_param_count = get_num_params(whole_model)
    num_flop_per_token = get_num_flop_per_token(
        get_num_params(whole_model, exclude_embedding=True),
        model_config,
        job_config.training.seq_len,
    )
    logger.info(
        f"{color.blue}Model {model_name} {job_config.model.flavor} "
        f"{color.red}size: {model_param_count:,} total parameters{color.reset}"
    )

    # initialize GPU memory monitor before applying parallelisms to the model
    gpu_memory_monitor = build_gpu_memory_monitor()
    # obtain the peak flops of bf16 type for MFU calculation
    gpu_peak_flops = get_peak_flops(gpu_memory_monitor.device_name)

    if parallel_dims.pp_enabled:
        stages, model_parts = models_pipelining_fns[model_name](
            whole_model, world_mesh, parallel_dims, job_config, device, model_config
        )
    else:
        # In 1D/2D cases or PP with simple schedules, model_parts is just one item
        # for PP with looped schedules, each item is one stage-model-chunk
        # we iterate all model_parts for applying SPMD parallelism, compilation, optimizer, and checkpointing
        model_parts = [whole_model]

    # apply PT-D DP/TP parallelisms and activation checkpointing
    model_parts = [
        models_parallelize_fns[model_name](m, world_mesh, parallel_dims, job_config)
        for m in model_parts
    ]

    init_device = "cpu" if job_config.checkpoint.create_seed_checkpoint else "cuda"
    for model in model_parts:
        model.to_empty(device=init_device)

    if parallel_dims.pp_enabled:
        pp_schedule = build_pipeline_schedule(
            job_config, parallel_dims, stages, loss_fn
        )
    else:
        # If PP is enabled, we can't rely on init_weights, because some layers are missing.
        # In the future, we may make init_weights handle missing layers, but also have to consider RNG seed propagation.
        # allocate sharded model on GPU and initialize weights via DTensor
        whole_model.init_weights()

    gpu_mem_stats = gpu_memory_monitor.get_peak_stats()
    logger.info(
        f"GPU memory usage for model: "
        f"{gpu_mem_stats.max_reserved_gib:.2f}GiB"
        f"({gpu_mem_stats.max_reserved_pct:.2f}%)"
    )

    # build optimizer after applying parallelisms to the model
    optimizers = build_optimizers(model_parts, job_config)
    lr_schedulers = get_lr_schedulers(optimizers.optimizers, job_config)

    metric_logger = build_metric_logger(
        job_config, metrics_log_rank=get_metrics_rank(world_mesh, parallel_dims)
    )

    train_state = TrainState()

    # train loop
    for model in model_parts:
        model.train()

    # load initial checkpoint
    checkpoint = CheckpointManager(
        model_parts=model_parts,
        optimizers=optimizers.optimizers,
        lr_schedulers=lr_schedulers.schedulers,
        dataloader=data_loader,
        states={"train_state": train_state},
        job_config=job_config,
    )

    if job_config.checkpoint.create_seed_checkpoint:
        assert (
            world_size == 1
        ), "Must create seed-checkpoint using one gpu, to disable sharding"
        checkpoint.save(curr_step=0, force=True)
        logger.info("Created seed checkpoint")
        return

    checkpoint_loaded = checkpoint.load()

    if parallel_dims.pp_enabled and not checkpoint_loaded:
        raise RuntimeError(
            "Pipeline Parallelism requires meta-initialization and loading seed checkpoint. "
            "Please run `./create_seed_checkpoint.sh` and rerun training with `--checkpoint.enable_checkpoint`"
        )

    # plot losses loaded from checkpoint (if any) to TensorBoard
    # NOTE: Loss info after the last log step before checkpoint saving will not be ploted.
    #       This can be avoided by setting checkpoint.interval to be a multiple of metrics.log_freq
    if train_state.step > 0:
        for idx, step in enumerate(train_state.log_steps):
            metrics = {
                "loss_metrics/global_avg_loss": train_state.global_avg_losses[idx],
                "loss_metrics/global_max_loss": train_state.global_max_losses[idx],
            }
            metric_logger.log(metrics, step=step)

    data_iterator = iter(data_loader)

    checkpoint.reset()

    # variables used to keep info for metrics logging
    losses_since_last_log: List[float] = []
    ntokens_since_last_log = 0
    data_loading_times: List[float] = []
    time_last_log = timer()
    gpu_memory_monitor.reset_peak_stats()

    # train loop
    logger.info(f"Training starts at step {train_state.step + 1}")
    with maybe_enable_profiling(
        job_config, global_step=train_state.step
    ) as torch_profiler, maybe_enable_memory_snapshot(
        job_config, global_step=train_state.step
    ) as memory_profiler:
        while train_state.step < job_config.training.steps:
            train_state.step += 1
            if train_state.step > 1 and train_state.step % _gc_freq == 0:
                gc.collect(1)

            # get batch
            data_load_start = timer()
            batch = next(data_iterator)
            input_ids, labels = batch
            ntokens_since_last_log += labels.numel()
            data_loading_times.append(timer() - data_load_start)

            input_ids = input_ids.cuda()
            labels = labels.cuda()
            optimizers.zero_grad()

            if parallel_dims.pp_enabled:
                # pipeline parallel forward / backward inside step() call
                is_last_stage = pp_mesh.get_local_rank() == pp_mesh.size() - 1

                with train_context():
                    if pp_mesh.get_local_rank() == 0:
                        pp_schedule.step(input_ids)
                    elif is_last_stage:
                        losses = []
                        pp_schedule.step(target=labels, losses=losses)
                    else:
                        pp_schedule.step()

                # accumulate losses across pipeline microbatches
                loss = (
                    torch.mean(torch.stack(losses))
                    if is_last_stage
                    else torch.Tensor([-1.0])
                )
            else:
                # Non-PP forward / backward
                with train_context():
                    pred = model(input_ids)
                    loss = loss_fn(pred, labels)
                    # pred.shape=(bs, seq_len, vocab_size)
                    # need to free to before bwd to avoid peaking memory
                    del pred
                    loss.backward()

            # clip gradients
            for model in model_parts:
                torch.nn.utils.clip_grad_norm_(
                    model.parameters(), job_config.training.max_norm, foreach=True
                )

            # optimizer step
            checkpoint.wait_for_staging()
            optimizers.step()
            lr_schedulers.step()

            # when fp8 config is on,
            # calculate float8 dynamic amax/scale for all-parameter for FSDP2
            # it issues a single all-reduce for all parameters at once for better performance
            maybe_precompute_fp8_dynamic_scale_for_fsdp(model, job_config)

            losses_since_last_log.append(loss)

            # log metrics
            if (
                train_state.step == 1
                or train_state.step % job_config.metrics.log_freq == 0
            ):
                losses = [loss.item() for loss in losses_since_last_log]
                avg_loss, max_loss = (
                    np.mean(losses),
                    np.max(losses),
                )
                if parallel_dims.dp_enabled:
                    global_avg_loss, global_max_loss = (
                        dist_mean(avg_loss, dp_mesh).item(),
                        dist_max(max_loss, dp_mesh).item(),
                    )
                else:
                    global_avg_loss, global_max_loss = avg_loss, max_loss

                train_state.log_steps.append(train_state.step)
                train_state.global_avg_losses.append(global_avg_loss)
                train_state.global_max_losses.append(global_max_loss)

                time_delta = timer() - time_last_log

                # tokens per second, abbr. as wps by convention
                wps = ntokens_since_last_log / (
                    time_delta * parallel_dims.model_parallel_size
                )
                # model FLOPS utilization
                # For its definition and calculation, please refer to the PaLM paper:
                # https://arxiv.org/abs/2204.02311
                mfu = 100 * num_flop_per_token * wps / gpu_peak_flops

                time_end_to_end = time_delta / job_config.metrics.log_freq
                time_data_loading = np.mean(data_loading_times)
                time_data_loading_pct = 100 * np.sum(data_loading_times) / time_delta

                gpu_mem_stats = gpu_memory_monitor.get_peak_stats()

                metrics = {
                    "loss_metrics/global_avg_loss": global_avg_loss,
                    "loss_metrics/global_max_loss": global_max_loss,
                    "wps": wps,
                    "mfu(%)": mfu,
                    "time_metrics/end_to_end(s)": time_end_to_end,
                    "time_metrics/data_loading(s)": time_data_loading,
                    "time_metrics/data_loading(%)": time_data_loading_pct,
                    "memory/max_active(GiB)": gpu_mem_stats.max_active_gib,
                    "memory/max_active(%)": gpu_mem_stats.max_active_pct,
                    "memory/max_reserved(GiB)": gpu_mem_stats.max_reserved_gib,
                    "memory/max_reserved(%)": gpu_mem_stats.max_reserved_pct,
                    "memory/num_alloc_retries": gpu_mem_stats.num_alloc_retries,
                    "memory/num_ooms": gpu_mem_stats.num_ooms,
                }
                metric_logger.log(metrics, step=train_state.step)

                logger.info(
                    f"{color.cyan}step: {train_state.step:2}  "
                    f"{color.green}loss: {global_avg_loss:7.4f}  "
                    f"{color.yellow}memory: {gpu_mem_stats.max_reserved_gib:5.2f}GiB"
                    f"({gpu_mem_stats.max_reserved_pct:.2f}%)  "
                    f"{color.blue}wps: {round(wps):,}  "
                    f"{color.magenta}mfu: {mfu:.2f}%{color.reset}"
                )

                losses_since_last_log.clear()
                ntokens_since_last_log = 0
                data_loading_times.clear()
                time_last_log = timer()
                gpu_memory_monitor.reset_peak_stats()

            checkpoint.save(
                train_state.step, force=(train_state.step == job_config.training.steps)
            )

            # signals the profiler that the next profiling step has started
            if torch_profiler:
                torch_profiler.step()

            if memory_profiler:
                memory_profiler.step()

            # Reduce timeout after first train step for faster signal (assumes lazy init, compile are finished)
            if train_state.step == 1:
                set_pg_timeouts(
                    timeout=timedelta(seconds=job_config.comm.train_timeout_seconds),
                    world_mesh=world_mesh,
                )

    if torch.distributed.get_rank() == 0:
        logger.info("Sleeping 2 seconds for other ranks to complete")
        time.sleep(2)

    metric_logger.close()
    logger.info("Training completed")


if __name__ == "__main__":
    config = JobConfig()
    config.parse_args()
    main(config)
    destroy_process_group()