# Copyright (c) Meta Platforms, Inc. and affiliates. # This software may be used and distributed according to the terms of the Llama 2 Community License Agreement. import gc import logging import os import sys from contextlib import ExitStack from copy import deepcopy from dataclasses import asdict, dataclass from pathlib import Path from timeit import default_timer as timer from typing import Any, Dict, Type, TypeVar import torch import torch.distributed import torch.nn.functional import torch.nn.functional as F import wandb import xformers.profiler from omegaconf import OmegaConf from torch.distributed._tensor import DTensor from torch.distributed.checkpoint.stateful import Stateful from torch.optim import lr_scheduler from bytelatent.args import TrainArgs from bytelatent.checkpoint import CheckpointManager, load_from_checkpoint from bytelatent.data.data_types import DataLoaderState from bytelatent.distributed import ( check_model_value_range, clean_env, dist_mean_dict, get_device_mesh, get_is_master, get_world_size, init_signal_handler, parallelize_model, requeue_slurm_job, setup_env, setup_torch_distributed, ) from bytelatent.logger import init_logger from bytelatent.metrics import GPUMemoryMonitor, MetricLogger, get_num_params from bytelatent.model.blt import ByteLatentTransformer from bytelatent.optim import build_optimizer from bytelatent.probe import AutoProbeD from bytelatent.profiling import maybe_run_profiler from bytelatent.stool import StoolArgs, launch_job from bytelatent.transformer import ( build_fsdp_grouping_plan, get_no_recompute_ops, get_num_flop_per_token, tp_parallelize, ) logger = logging.getLogger() T = TypeVar("T") def flatten_dict(d, parent_key="", sep="_"): items = [] for k, v in d.items(): new_key = f"{parent_key}{sep}{k}" if parent_key else k if isinstance(v, dict): items.extend(flatten_dict(v, new_key, sep=sep).items()) else: items.append((new_key, v)) return dict(items) def dataclass_from_dict(cls: Type[T], data: dict, strict: bool = True) -> T: """ Converts a dictionary to a dataclass instance, recursively for nested structures. """ base = OmegaConf.structured(cls()) OmegaConf.set_struct(base, strict) override = OmegaConf.create(data) return OmegaConf.to_object(OmegaConf.merge(base, override)) @dataclass class TrainState(Stateful): step: int # Nb of steps taken by the optimizer acc_step: int # Nb of accumulation steps done since last optimizer step scheduler: lr_scheduler.LambdaLR data_loader_state: DataLoaderState scale: float = 1.0 def state_dict(self) -> Dict[str, Any]: return { "step": self.step, "acc_step": self.acc_step, "data_loader_state": self.data_loader_state.dict(), "scheduler": self.scheduler.state_dict(), } def load_state_dict(self, state_dict): self.step = state_dict["step"] self.acc_step = state_dict["acc_step"] self.data_loader_state = DataLoaderState(**state_dict["data_loader_state"]) self.scheduler.load_state_dict(state_dict["scheduler"]) def validate_train_args(args: TrainArgs, output_size: int): if args.model.vocab_size < 0: logger.info(f"Setting model output size to {args.model.vocab_size}") args.model.vocab_size = output_size assert args.dump_dir, "Dump dir not set" if args.checkpoint.path is None: logger.info(f"Setting checkpoint path to {args.checkpoint.path}") args.checkpoint.path = str(Path(args.dump_dir) / "checkpoints") for source in args.data.sources: data_path = os.path.join(args.data.root_dir, source) assert os.path.exists(data_path), f"{data_path} doesn't exist" if ( args.distributed.dp_replicate * args.distributed.dp_shard * args.distributed.tp_size != get_world_size() ): assert get_world_size() % args.distributed.dp_shard == 0 args.distributed.dp_replicate = get_world_size() // args.distributed.dp_shard assert args.distributed.dp_replicate % args.distributed.tp_size == 0 args.distributed.dp_replicate = ( args.distributed.dp_replicate // args.distributed.tp_size ) logger.warning( f"Setting Data Parallel size to {args.distributed.dp_replicate * args.distributed.dp_shard}" ) assert ( args.distributed.dp_replicate * args.distributed.dp_shard * args.distributed.tp_size == get_world_size() ) if args.distributed.fsdp_type == "no_shard": assert ( args.distributed.dp_shard == 1 and args.distributed.dp_replicate == get_world_size() ) args.model.max_seqlen = args.data.seq_len if args.distributed.tp_size == 1: logger.warning( "Tensor parallelism has not been tested for a while, use at your own risk" ) assert ( args.probe_freq != args.profiling.mem_steps ), "Don't profile during probe step" assert ( args.probe_freq != args.profiling.profile_steps ), "Don't profile during probe step" if args.logging.wandb is not None: args.logging.wandb.name = args.name if args.probe_freq is not None: assert ( args.distributed.tp_size == 1 ), "Probing not supported with tensor parallelism" assert ( args.distributed.selective_activation_checkpointing is False ), "Probing not supported with selective activation checkpointing" preemption_flag = dict(flag=False) def set_preemption_flag(signum, frame): logger.warning("Signal handler called with signal " + str(signum)) logger.warning("Preemption ! checkpointing asap and exiting.") preemption_flag["flag"] = True def every_n_steps(train_state, freq, acc_step=None, acc_freq=None): test = train_state.step % freq == 0 if acc_step is not None: test = test and (train_state.acc_step == acc_step) elif acc_freq is not None: test = test and ((train_state.acc_step % acc_freq) == 0) return test def compute_loss(p, y, mask, scale): tok_loss = scale * F.cross_entropy( p.flatten(0, 1), y.flatten(0, 1), reduction="none" ) if mask is None: loss = tok_loss.mean() else: mask = mask.flatten(0, 1) tok_loss = tok_loss * mask loss = tok_loss.sum() / (mask.sum() + 1e-6) return loss, tok_loss def train(args: TrainArgs): with ExitStack() as context_stack: tokenizer = args.data.tokenizer_args.build() validate_train_args( args, tokenizer.n_words, ) if get_is_master(): os.makedirs(args.dump_dir, exist_ok=True) args.dump_to_yaml_file(Path(args.dump_dir) / "config.yaml") init_logger(Path(args.dump_dir) / "train.log") init_signal_handler(set_preemption_flag) # For handling preemption signals. setup_env(args.env) setup_torch_distributed(args.distributed) world_mesh = get_device_mesh(args.distributed) logger.info(f"Starting job: {args.name}") # build dataloader # need dp world size and rank dp_mesh = world_mesh["dp_replicate"] dp_degree = dp_mesh.size() dp_rank = dp_mesh.get_local_rank() if args.distributed.dp_shard > 1: dp_rank = dp_rank * dp_degree + world_mesh["dp_shard"].get_local_rank() dp_degree *= world_mesh["dp_shard"].size() logger.info(f"Running on dp rank : {dp_rank}") logger.info(f"Running on dp size : {dp_degree}") torch.manual_seed(args.seed) logger.info("Building model") # Initializing Model in meta device allows us to initialize models much bigger than 1 gpu's memory with torch.device("meta"): model = ByteLatentTransformer(args.model) logger.info("Model is built !") model_param_count = get_num_params(model) model = parallelize_model( model, world_mesh, args.model, args.distributed, fsdp_grouping_plan=build_fsdp_grouping_plan(args.model), tp_parallelize=tp_parallelize, no_recompute_ops=get_no_recompute_ops(), ) # Once we shard the model on different gpus we can actually initialize the model # First we create empty tensors of the correct shapes model = model.to_empty(device="cuda") # Then we init the model. Please make sure this function initializes *ALL* parameters # and buffers, otherwise you will have random values in the unitialized tensors # which will silently fail (give nan gradients for example) if args.checkpoint.init_ckpt_path: logger.info(f"Loading initial model from {args.checkpoint.init_ckpt_path}") load_from_checkpoint( args.checkpoint.init_ckpt_path, model, model_key="model" ) # Put model_key="" if its directly the model checkpoint model.rope_embeddings.reset_parameters() # For RoPe initialization since it's a buffer it might not be loaded else: with torch.random.fork_rng(devices=[torch.cuda.current_device()]): torch.manual_seed(args.model.seed) model.init_weights() check_model_value_range(model, range=10.0, std=1.0) # log model size logger.info(f"Model size: {model_param_count:,} total parameters") gpu_memory_monitor = GPUMemoryMonitor("cuda") logger.info( f"GPU capacity: {gpu_memory_monitor.device_name} ({gpu_memory_monitor.device_index}) " f"with {gpu_memory_monitor.device_capacity_gib:.2f}GiB memory" ) logger.info(f"GPU memory usage: {gpu_memory_monitor}") # build optimizer after apply parallelisms to the model optimizer, scheduler = build_optimizer(model, args.optim, args.steps) data_loader = args.data.build_from_rank(dp_rank, dp_degree) data_loader_state = data_loader.get_state() train_state = TrainState( step=0, acc_step=0, data_loader_state=data_loader_state, scheduler=scheduler, scale=1.0, ) checkpoint = CheckpointManager.instantiate_and_make_dir(args.checkpoint) checkpoint.load(model, optimizer, train_state, world_mesh) # Either load from latest checkpoint or start from scratch if args.probe_freq is not None: if get_is_master(): os.makedirs(Path(args.dump_dir) / "probe", exist_ok=True) torch.distributed.barrier() probe = AutoProbeD( model, ( Path(args.dump_dir) / "probe" / f"probe.{dp_rank}.jsonl" if (dp_rank % 128 == 0) else None ), ) probe_mod = model._orig_mod if args.distributed.compile else model gc.disable() # train loop model.train() metric_logger = context_stack.enter_context( MetricLogger(Path(args.dump_dir) / "metrics.jsonl", args) ) data_loader = train_state.data_loader_state.build() batch_iterator = data_loader.create_iter() torch_profiler = context_stack.enter_context( maybe_run_profiler(args.dump_dir, model, args.profiling) ) nwords_since_last_log = 0 time_last_log = timer() gc.collect() while train_state.step < args.steps: # We constrain train_state.acc_step to be in range 0 to args.grad_acc_steps - 1 train_state.acc_step += 1 train_state.acc_step = train_state.acc_step % args.grad_acc_steps # get batch curr_lr = float(optimizer.param_groups[0]["lr"]) data_load_start = timer() batch = next(batch_iterator) batch_x = torch.from_numpy( batch.x, ).cuda() batch_y = torch.from_numpy(batch.y).cuda() batch_patch_lengths = torch.from_numpy(batch.patch_lengths).cuda() mask = None if batch.mask is None else torch.from_numpy(batch.mask).cuda() if args.model.encoder_enable_byte_ngrams and batch.ngram_ids is None: raise ValueError( "Cannot enable byte ngrams and have batch.ngram_ids be None" ) ngram_ids = ( None if batch.ngram_ids is None else torch.from_numpy(batch.ngram_ids).cuda() ) if every_n_steps(train_state, args.gc_collect_freq, acc_step=0): logger.info("garbage collection") # we do garbage collection manually otherwise different processes # run the GC at different times so they slow down the whole pipeline gc.collect() data_load_time = round(timer() - data_load_start, 4) nwords_since_last_log += batch_x.numel() bsz, seqlen = batch_y.shape # forward start_timer = torch.cuda.Event(enable_timing=True) end_timer = torch.cuda.Event(enable_timing=True) start_timer.record() # This is an automatic probe that will compute statistics # of all linears' inputs, weights and outputs # along with attention logits and entropy # both in forward and backward pass tok_loss = None if (args.probe_freq is not None) and every_n_steps( train_state, args.probe_freq, acc_step=1 % args.grad_acc_steps ): # Here we do a fake forward and backward pass on a smaller # batch size to avoid OOM # This assumes the model has no stateful layers (batch norm..) assert ( next(probe_mod.parameters()).grad is None ), "Can't probe model if grads are not reset" with probe: probe.metadata = { "it": train_state.step, "global_step": train_state.step, "loop": "lingua", } # Non compiled model uses roughly 2x memory in our exps # So we divide bsz by 2 or seqlen by 2 probe_bsz = max(1, bsz // 2) probe_seq = seqlen if (bsz // 2 >= 1) else (seqlen // 2) probe_loss = probe_mod( batch_x[:probe_bsz, :probe_seq], batch_y[:probe_bsz, :probe_seq], ) probe_loss.backward() # We zero grads to cancel this fake step optimizer.zero_grad() assert ( next(probe_mod.parameters()).grad is None ), "Probe model shouldn't have grads at this point" pred = model( batch_x, patch_lengths=batch_patch_lengths, ngram_ids=ngram_ids ) loss, _ = compute_loss(pred, batch_y, mask, train_state.scale) # We scale loss with grad_acc_steps so the gradient is the same # regardless of grad_acc_steps loss = loss / args.grad_acc_steps # backward on scaled loss to create scaled gradients loss.backward() # For logging we undo that scaling loss = loss.detach() * args.grad_acc_steps grad_norm = torch.nn.utils.clip_grad_norm_( model.parameters(), max_norm=args.optim.clip, foreach=True ) grad_norm = ( grad_norm.full_tensor() if isinstance(grad_norm, DTensor) else grad_norm ).item() # optimizer step if train_state.acc_step == 0: optimizer.step() scheduler.step() optimizer.zero_grad() train_state.step += 1 # updates the scale for next iteration # training iteration complete end_timer.record() torch.cuda.synchronize() curr_iter_time = round(start_timer.elapsed_time(end_timer) * 1e-3, 4) # if profiler is active if torch_profiler: xformers.profiler.step() # log metrics if every_n_steps( train_state, args.logging.freq, acc_step=None if args.logging.acc_freq else 0, acc_freq=args.logging.acc_freq, ): time_delta = timer() - time_last_log wps = nwords_since_last_log / (time_delta * args.distributed.tp_size) gpu_mem_stats = gpu_memory_monitor.get_peak_stats() total_acc_steps = ( args.grad_acc_steps * train_state.step + train_state.acc_step ) tokens_per_gpu = ( total_acc_steps * args.data.batch_size * args.data.seq_len ) total_tokens = dp_degree * tokens_per_gpu # This is an estimate and the correct values may change # if you change the architecture # Use xformer's analyze profile trace to get actual measurement FLOPS = ( get_num_flop_per_token( model_param_count - args.model.vocab_size * args.model.dim, args.model.n_layers, args.model.dim, args.data.seq_len, ) * wps ) metrics = flatten_dict( { "global_step": train_state.step, "acc_step": train_state.acc_step, "speed": { "wps": wps, "FLOPS": FLOPS, "curr_iter_time": curr_iter_time, "data_load_time": data_load_time, }, "optim": { "grad_norm": grad_norm, "lr": curr_lr, "total_tokens": total_tokens, }, "memory": gpu_mem_stats._asdict(), }, sep="/", ) to_sync = {} to_sync["loss/out"] = loss.item() metrics.update(dist_mean_dict(to_sync)) if get_is_master(): metric_logger.log(metrics) gpu_memory_monitor.reset_peak_stats() nwords_since_last_log = 0 time_last_log = timer() logger.info( f"step: {train_state.step}" f" acc: {train_state.acc_step}" f" loss: {round(loss.item(),4):>7}" f" grad: {grad_norm:.2e}" f" flops: {FLOPS:.2e}" f" wps: {wps:.2e}" f" iter: {curr_iter_time:>7}" f" data: {data_load_time:>5}" f" lr: {curr_lr:.2e}" f" mem: {gpu_mem_stats.max_active_pct:.0f}%" f" pow: {gpu_mem_stats.power_draw/1000} W" ) saved = False if every_n_steps( train_state, args.checkpoint.dump.every, acc_step=0 ) or every_n_steps(train_state, args.checkpoint.eval.every, acc_step=0): saved = checkpoint.save( model, optimizer, train_state, args, device_mesh=world_mesh, ) if args.eval is not None and every_n_steps( train_state, args.checkpoint.eval.every, acc_step=0 ): from apps.main.eval import EVAL_FOLDER_NAME, EvalArgs, launch_eval eval_args = dataclass_from_dict(EvalArgs, args.eval) eval_args.global_step = train_state.step eval_args.ckpt_dir = str(checkpoint.existing_saves[-1]) eval_args.dump_dir = str( os.path.join( args.dump_dir, "evals", EVAL_FOLDER_NAME.format(train_state.step), ) ) eval_args.metric_log_dir = args.dump_dir if args.async_eval_gpus is None: launch_eval(eval_args) elif get_is_master(): if wandb.run is not None and args.logging.wandb is not None: eval_args.wandb = deepcopy(args.logging.wandb) assert args.async_eval_gpus > 0 logger.info(f"Launching evals on {args.async_eval_gpus} gpus") with clean_env(): launch_job( StoolArgs( asdict(eval_args), script="apps.main.eval", copy_code=False, nodes=args.async_eval_gpus // 8, qos="lowest", ) ) if preemption_flag["flag"]: if not saved: checkpoint.save( model, optimizer, train_state, args, device_mesh=world_mesh, ) requeue_slurm_job() sys.exit(0) if not saved: checkpoint.save( model, optimizer, train_state, args, device_mesh=world_mesh, ) gc.collect() def main(): """ The command line interface here uses OmegaConf https://omegaconf.readthedocs.io/en/2.3_branch/usage.html#from-command-line-arguments This accepts arguments as a dot list So if the dataclass looks like @dataclass class DummyArgs: name: str model: LMTransformerArgsgs @dataclass class LMTransformerArgsgs: dim: int Then you can pass model.dim=32 to change values in LMTransformerArgsgs or just name=tictac for top level attributes. The behavior here is as follows: 1. We instantiate TrainArgs with its default values 2. We override those default values with the ones in the provided config file 3. We override the result with the additional arguments provided through command line For example, if the config is the following model: dim: 128 n_layers: 4 and you call train.py with train.py model.dim=64 Then the final TrainArgs will have model: dim: 64 n_layers: 4 Plus all the default values in TrainArgs dataclass. """ cli_args = OmegaConf.from_cli() file_cfg = OmegaConf.load(cli_args.config) # We remove 'config' attribute from config as the underlying DataClass does not have it del cli_args.config default_cfg = OmegaConf.create(TrainArgs().model_dump()) cfg = OmegaConf.merge(default_cfg, file_cfg, cli_args) cfg = OmegaConf.to_container(cfg, resolve=True, throw_on_missing=True) train_args = TrainArgs.model_validate(cfg) if train_args.debug_dynamo: import torch._dynamo torch._dynamo.config.suppress_errors = True train(train_args) if __name__ == "__main__": main()