picotron/train.py

"""Training script for LLaMA model.
torchrun --nproc_per_node 1 --master_addr localhost --master_port 25500 train.py --use_wandb
torchrun --nproc_per_node 2 --master_addr localhost --master_port 25500 train.py --dp_size 2 --use_wandb 
torchrun --nproc_per_node 4 --master_addr localhost --master_port 25500 train.py --tp_size 2 --pp_size 2 --use_wandb 
torchrun --nproc_per_node 4 --master_addr localhost --master_port 25500 train.py --tp_size 2 --pp_size 2 --load_path ckpt/150
torchrun --nproc_per_node 8 --master_addr localhost --master_port 25500 train.py --tp_size 2 --dp_size 2 --pp_size 2 --use_wandb
CUDA_DEVICE_MAX_CONNECTIONS=1 debugpy-run -p 5678 -m torch.distributed.run -- --nproc_per_node=2 --nnodes=1 --rdzv_backend=c10d --rdzv_endpoint=localhost:29400 train.py --dp_size 2
CUDA_DEVICE_MAX_CONNECTIONS=1 torchrun    --nproc_per_node=4    --nnodes=1    --rdzv_backend=c10d    --rdzv_endpoint=localhost:29400    --max_restarts=0    --tee=3    train.py
#VERBOSE=0 torchrun --nproc_per_node 4 --master_addr localhost --master_port 25500 train.py --pp_size 2 --dp_size 2
"""
import os
import inspect
import datetime
import json
import time
import datetime
import argparse
import torch.nn.functional as F
import torch, torch.distributed as dist
from torch.optim import AdamW
from transformers import AutoConfig
from picotron.context_parallel.context_parallel import apply_context_parallel
from picotron.tensor_parallel.tensor_parallel import apply_tensor_parallel
import picotron.process_group_manager as pgm
from picotron.utils import set_all_seed, print, to_readable_format
from picotron.checkpoint import CheckpointManager
from picotron.checkpoint import init_model_with_dematerialized_weights, init_model_with_materialized_weights
from picotron.data import MicroBatchDataLoader
from picotron.process_group_manager import setup_process_group_manager
from picotron.pipeline_parallel.pipeline_parallel import train_step_pipeline_1f1b, train_step_pipeline_afab, PipelineParallel
from picotron.data_parallel.data_parallel import DataParallelBucket
from picotron.model import Llama
import wandb
import lovely_tensors as lt; lt.monkey_patch()

def train_step(model, data_loader, device):
    acc_loss = 0.0
    
    requires_grad_sync = pgm.process_group_manager.cp_dp_world_size > 1
    for i in range(data_loader.grad_acc_steps):
        # get the next batch
        batch = next(data_loader)
        input_ids = batch["input_ids"].to(device)
        target_ids = batch["target_ids"].to(device)

        # disable gradient synchronization for all but the last micro-batch
        if requires_grad_sync:
            model.require_backward_grad_sync = (i == data_loader.grad_acc_steps - 1)

        outputs = model(input_ids=input_ids)

        # compute the loss
        batch_size, seq_len = input_ids.shape
        target_ids = target_ids.reshape(-1)
        outputs = outputs.view(seq_len*batch_size, -1)
        loss = F.cross_entropy(outputs, target_ids, reduction='mean') / data_loader.grad_acc_steps
        
        loss.backward()

        acc_loss += loss.item()

    return acc_loss

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--config", type=str, default="", help="Path to config file")
    args = parser.parse_args()

    with open(args.config, "r") as f:
        config = json.load(f)
    
    os.environ["OMP_NUM_THREADS"] = config["environment"]["OMP_NUM_THREADS"]
    os.environ["TOKENIZERS_PARALLELISM"] = config["environment"]["TOKENIZERS_PARALLELISM"]
    os.environ["FLASH_ATTEN"] = config["environment"]["FLASH_ATTEN"]
    os.environ["DEVICE"] = "cpu" if config["distributed"]["use_cpu"] else "cuda"
    
    dtype = torch.bfloat16 if torch.cuda.is_available() and torch.cuda.is_bf16_supported() and not config["distributed"]["use_cpu"] else torch.float32
    assert (dtype == torch.bfloat16 and os.getenv("FLASH_ATTEN") == "1") or os.getenv("FLASH_ATTEN") != "1", "Kernel operations requires dtype=torch.bfloat16"
    
    local_rank = int(os.environ["LOCAL_RANK"])
    global_rank = int(os.environ["RANK"])
    world_size = int(os.environ["WORLD_SIZE"])

    backend = "gloo" if config["distributed"]["use_cpu"] else "nccl"
    
    assert config["training"]["seq_length"] % config["distributed"]["cp_size"] == 0, "seq_length must be divisible by cp_size for Context Parallelism"
    assert world_size == config["distributed"]["tp_size"] * config["distributed"]["pp_size"] * config["distributed"]["dp_size"] * config["distributed"]["cp_size"], "world_size must be equal to tp_size * pp_size * dp_size * cp_size"

    if backend == "nccl":
        torch.cuda.set_device(local_rank)
        device = torch.device("cuda", local_rank)
    else:
        device = torch.device("cpu")

    dist.init_process_group(rank=global_rank, world_size=world_size, backend=backend, init_method=f"env://", timeout=datetime.timedelta(minutes=3))
    setup_process_group_manager(
        tp_size=config["distributed"]["tp_size"],
        cp_size=config["distributed"]["cp_size"],
        pp_size=config["distributed"]["pp_size"],
        dp_size=config["distributed"]["dp_size"]
    )
    is_wandb_rank = pgm.process_group_manager.tp_rank == 0 and pgm.process_group_manager.dp_rank == 0 and pgm.process_group_manager.cp_rank == 0 and pgm.process_group_manager.pp_is_last_stage

    set_all_seed(config["training"]["seed"])
    
    start_time = time.time()
    data_loader = MicroBatchDataLoader(
        micro_batch_size=config["training"]["micro_batch_size"],
        seq_length=config["training"]["seq_length"],
        dataset_name=config["dataset"]["name"],
        tokenizer_name=config["model"]["name"],
        grad_acc_steps=config["training"]["gradient_accumulation_steps"],
        num_workers=config["dataset"]["num_workers"],
        num_proc=config["dataset"]["num_proc"],
        num_samples=config["training"]["num_samples"]
    )

    dist.barrier()

    print("init dataloader time:", time.time()-start_time, is_print_rank=is_wandb_rank)
    tokens_per_step = data_loader.global_batch_size * config["training"]["seq_length"]
    
    if pgm.process_group_manager.global_rank == 0:
        print("Tokens per step:", to_readable_format(tokens_per_step), is_print_rank=is_wandb_rank)

    if is_wandb_rank and config["logging"]["use_wandb"]:
        wandb.init(
            project="picotron",
            name=f"{config['logging']['run_name']}_{tokens_per_step}_{pgm.process_group_manager}",
            config={
                "tensor_parallel_size": pgm.process_group_manager.tp_size,
                "context_parallel_size": pgm.process_group_manager.cp_size,
                "pipeline_parallel_size": pgm.process_group_manager.pp_size,
                "data_parallel_size": pgm.process_group_manager.dp_size,
                "model": config["model"]["name"],
                "dataset": config["dataset"]["name"],
                "max_tokens": config["training"]["max_tokens"],
                "learning_rate": config["training"]["learning_rate"],
                "seed": config["training"]["seed"],
                "micro_batch_size": data_loader.micro_batch_size,
                "global_batch_size": data_loader.global_batch_size,
                "gradient_accumulation": data_loader.grad_acc_steps,
            },
        )

    model_config = AutoConfig.from_pretrained(config["model"]["name"])
    model_config.num_hidden_layers = config["model"]["num_hidden_layers"]
    model_config.num_attention_heads = config["model"]["num_attention_heads"]
    model_config.num_key_value_heads = config["model"]["num_key_value_heads"]
    model_config.max_position_embeddings = config["training"]["seq_length"]

    start_time = time.time()

    with init_model_with_dematerialized_weights():
        model = Llama(config=model_config)

        if pgm.process_group_manager.tp_world_size > 1:
            model = apply_tensor_parallel(model)

        if pgm.process_group_manager.pp_world_size > 1:
            model = PipelineParallel(model, model_config)

    model = init_model_with_materialized_weights(model, model_config, hf_hub_checkpoint_path=config["checkpoint"]["hf_hub_checkpoint_path"])

    #TODO: load existing checkpoint here to continue pre-training

    if pgm.process_group_manager.cp_world_size > 1:
        model = apply_context_parallel(model)

    model.to(dtype).to(device)
        
    if pgm.process_group_manager.cp_dp_world_size > 1:
        model = DataParallelBucket(model)
    
    print("init model parallel time:", time.time()-start_time, is_print_rank=is_wandb_rank)
    
    model.train()
    
    tensor_shapes = (data_loader.micro_batch_size, data_loader.seq_length_per_gpu, model_config.hidden_size)
    
    extra_args = dict()
    if config["model"]["use_fused_adam"]:
        fused_available = 'fused' in inspect.signature(torch.optim.AdamW).parameters
        use_fused = fused_available and device == 'cuda'
        extra_args = dict(fused=True) if use_fused else dict()

    optimizer = AdamW(model.parameters(), lr=config["training"]["learning_rate"], **extra_args)
    
    checkpoint_manager = CheckpointManager()

    trained_tokens, step = 0, 0
    if config["checkpoint"]["load_path"]:
        step, trained_tokens = checkpoint_manager.load_checkpoint(model, optimizer, config["checkpoint"]["load_path"])
    
    dist.barrier()
        
    def _all_reduce_loss_across_dp_cp_ranks(loss, device):
        reduced_loss = torch.tensor([loss if loss is not None else 0.0], dtype=torch.float32, device=device)
        if pgm.process_group_manager.pp_is_last_stage:
            dist.all_reduce(reduced_loss, op=dist.ReduceOp.AVG, group=pgm.process_group_manager.cp_dp_group)
        return reduced_loss.item()
    
    while config["training"]["max_tokens"] is None or trained_tokens < config["training"]["max_tokens"]:
        step_start_time = time.time()
        optimizer.zero_grad()
        
        if pgm.process_group_manager.pp_world_size > 1:
            if config["distributed"]["pp_engine"] == "afab":
                loss = train_step_pipeline_afab(model, data_loader, tensor_shapes, device, dtype)
            elif config["distributed"]["pp_engine"] == "1f1b":
                loss = train_step_pipeline_1f1b(model, data_loader, tensor_shapes, device, dtype)
            else:
                raise ValueError(f"Invalid pipeline parallel engine: {config['distributed']['pp_engine']}")
        else:
            loss = train_step(model, data_loader, device)
            
        loss = _all_reduce_loss_across_dp_cp_ranks(loss, device)
        
        optimizer.step()
        trained_tokens += tokens_per_step
        step += 1
        
        if hasattr(model, 'reset'):
            model.reset()

        step_duration = time.time() - step_start_time
        
        if is_wandb_rank:
            print(f"[rank {pgm.process_group_manager.global_rank}] Step: {step}, Loss: {loss:.4f}, "
                f"Global batch size: {to_readable_format(tokens_per_step)}, "
                f"Tokens/s: {to_readable_format(tokens_per_step / step_duration)}, "
                f"Tokens/s/GPU: {to_readable_format(tokens_per_step / step_duration / world_size)}, "
                f"Tokens: {to_readable_format(trained_tokens)}{('/' + to_readable_format(config['training']['max_tokens'])) if config['training']['max_tokens'] else ''}, "
                f"Memory usage: {torch.cuda.memory_reserved() / 1e9:.2f}GB"
            , is_print_rank=is_wandb_rank)
        
            if config["logging"]["use_wandb"]:
                wandb.log({
                    "loss": loss,
                    "tokens_per_step": tokens_per_step,
                    "tokens_per_second": tokens_per_step / step_duration,
                    "memory_usage": torch.cuda.memory_reserved() / 1e9,
                    "trained_tokens": trained_tokens
                })
        
        if step % config["checkpoint"]["save_frequency"] == 0:
            checkpoint_manager.save_checkpoint(model, optimizer, step, trained_tokens, config["checkpoint"]["save_dir"]+f"/{step}")
        
        if step >= config["training"]["total_train_steps"]:
            break
    
    if is_wandb_rank and config["logging"]["use_wandb"]:
        wandb.finish()

    dist.destroy_process_group()