How to Train LLM? - From Data Parallel To Fully Sharded Data Parallel

# State initialization
            if len(state) == 0:
                # note(crcrpar): Deliberately host `step` on CPU if both capturable and fused are off.
                # This is because kernel launches are costly on CUDA and XLA.
                state["step"] = (
                    torch.zeros((), dtype=_get_scalar_dtype(is_fused=group["fused"]), device=p.device)
                    if group["capturable"] or group["fused"]
                    else torch.tensor(0.0, dtype=_get_scalar_dtype())
                )

# Creates model and optimizer in default precision
model = Net().cuda()
optimizer = optim.SGD(model.parameters(), ...)

# Creates a GradScaler once at the beginning of training.
scaler = GradScaler()

for epoch in epochs:
    for input, target in data:
        optimizer.zero_grad()

        # Runs the forward pass with autocasting.
        with autocast(device_type='cuda', dtype=torch.float16):
            output = model(input)
            loss = loss_fn(output, target)

        # Scales loss. Calls backward() on scaled loss to create scaled gradients.
        # Backward passes under autocast are not recommended.
        # Backward ops run in the same dtype autocast chose for corresponding forward ops.
        scaler.scale(loss).backward()

        # scaler.step() first unscales the gradients of the optimizer's assigned params.
        # If these gradients do not contain infs or NaNs, optimizer.step() is then called,
        # otherwise, optimizer.step() is skipped.
        scaler.step(optimizer)

        # Updates the scale for next iteration.
        scaler.update()

import torch.nn as nn

def data_parallel(module, inputs, labels, device_ids,
                  output_device):
  # (1) [Scatter] data를 device들에 scatter
  inputs = nn.parallel.scatter(inputs, device_ids)

  # (2) [Replicate] model weight을 device_ids들에 복제
  replicas = nn.parallel.replicate(module, device_ids)

  # (3) [Forward] 각 device 에 복제된 model이 각 device의 data를 forward
  logit = nn.parallel.parallel_apply(replicas, inputs)

  # (4) [Gather] 모델의 logit을 output_device(하나의 device) 로 모음
  logit = nn.parallel.gather(outputs, output_device)

  return logits

import torch
import torch.distributed as dist

dist.init_process_group("nccl")
rank = dist.get_rank()
torch.cuda.set_device(rank)

tensor = torch.ones(2,2).to(torch.cuda.current_device()) * rank
# rank == 0 => [[0, 0], [0, 0]]
# rank == 1 => [[1, 1], [1, 1]]
# rank == 2 => [[2, 2], [2, 2]]
# rank == 3 => [[3, 3], [3, 3]]

dist.all_reduce(tensor, op=torch.distributed.ReduceOp.SUM)

print(f"rank {rank}: {tensor}\n)
# result
rank 1: tensor([[6., 6.],
        [6., 6.]], device='cuda:1')

rank 2: tensor([[6., 6.],
        [6., 6.]], device='cuda:2')
rank 0: tensor([[6., 6.],
        [6., 6.]], device='cuda:0')


rank 3: tensor([[6., 6.],
        [6., 6.]], device='cuda:3')

import torch
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel
from torch.optim import AdamW
from torch.utils.data import DataLoader, DistributedSampler
from transformers import BertForSequenceClassification, BertTokenizer
from datasets import load_dataset

# 1. initialize process group
dist.init_process_group("nccl")
rank = dist.get_rank()
torch.cuda.set_device(rank)
device = torch.cuda.current_device()
world_size = dist.get_world_size()

# 2. create dataset
datasets = load_dataset("multi_nli").data["train"]
datasets = [
    {
        "premise": str(p),
        "hypothesis": str(h),
        "labels": l.as_py(),
    }
    for p, h, l in zip(datasets[2], datasets[5], datasets[9])
]

# 3. create DistributedSampler
# DistributedSampler 는 데이터를 쪼개서 다른 프로세스로 전송하기 위한 모듈입니다.

sampler = DistributedSampler(
    datasets,
    num_replicas=world_size,
    rank=rank,
    shuffle=True,
)

data_loader = DataLoader(
    datasets,
    batch_size=32,
    num_workers=4,
    sampler=sampler,
    shuffle=False,
    pin_memory=True,
)

# 4. create model and tokenizer
model_name = "bert-base-cased"
tokenizer = BertTokenizer.from_pretrained(model_name)
model = BertForSequenceClassification.from_pretrained(model_name, num_labels=3).cuda()

# 5. make distributed data parallel module
model = DistributedDataParallel(model, device_ids=[device], output_device=device)

# 6. create optimizer
optimizer = AdamW(model.parameters(), lr=3e-5)

# 7. start training
for i, data in enumerate(data_loader):
    optimizer.zero_grad()
    tokens = tokenizer(
        data["premise"],
        data["hypothesis"],
        padding=True,
        truncation=True,
        max_length=512,
        return_tensors="pt",
    )

    loss = model(
        input_ids=tokens.input_ids.cuda(),
        attention_mask=tokens.attention_mask.cuda(),
        labels=data["labels"],
    ).loss

    loss.backward()
    optimizer.step()

    if i % 10 == 0 and rank == 0:
        print(f"step: {i}, loss: {loss}")
    if i == 300:
        break

# 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 os

import dataclasses
import fire
import random
import torch
import torch.optim as optim
from peft import get_peft_model, prepare_model_for_kbit_training
from torch.distributed.fsdp import (
    FullyShardedDataParallel as FSDP,
    ShardingStrategy
)

from torch.distributed.fsdp.fully_sharded_data_parallel import CPUOffload
from torch.optim.lr_scheduler import StepLR
from transformers import (
    AutoTokenizer,
    LlamaForCausalLM,
    LlamaConfig,
)
from transformers.models.llama.modeling_llama import LlamaDecoderLayer

from llama_recipes.configs import fsdp_config as FSDP_CONFIG
from llama_recipes.configs import train_config as TRAIN_CONFIG
from llama_recipes.data.concatenator import ConcatDataset
from llama_recipes.policies import AnyPrecisionAdamW, apply_fsdp_checkpointing

from llama_recipes.utils import fsdp_auto_wrap_policy
from llama_recipes.utils.config_utils import (
    update_config,
    generate_peft_config,
    generate_dataset_config,
    get_dataloader_kwargs,
)
from llama_recipes.utils.dataset_utils import get_preprocessed_dataset

from llama_recipes.utils.fsdp_utils import hsdp_device_mesh
from llama_recipes.utils.train_utils import (
    train,
    freeze_transformer_layers,
    setup,
    setup_environ_flags,
    clear_gpu_cache,
    print_model_size,
    get_policies,
)
from accelerate.utils import is_xpu_available

def setup_wandb(train_config, fsdp_config, **kwargs):
    try:
        import wandb
    except ImportError:
        raise ImportError(
            "You are trying to use wandb which is not currently installed. "
            "Please install it using pip install wandb"
        )
    from llama_recipes.configs import wandb_config as WANDB_CONFIG
    wandb_config = WANDB_CONFIG()
    update_config(wandb_config, **kwargs)
    init_dict = dataclasses.asdict(wandb_config)
    run = wandb.init(**init_dict)
    run.config.update(train_config)
    run.config.update(fsdp_config, allow_val_change=True)
    return run


def main(**kwargs):
    # Update the configuration for the training and sharding process
    train_config, fsdp_config = TRAIN_CONFIG(), FSDP_CONFIG()
    update_config((train_config, fsdp_config), **kwargs)
    # Set the seeds for reproducibility
    if is_xpu_available():
        torch.xpu.manual_seed(train_config.seed)
    torch.manual_seed(train_config.seed)
    random.seed(train_config.seed)

    if train_config.enable_fsdp:
        setup()
        # torchrun specific
        local_rank = int(os.environ["LOCAL_RANK"])
        rank = int(os.environ["RANK"])
        world_size = int(os.environ["WORLD_SIZE"])

    if torch.distributed.is_initialized():
        if is_xpu_available():
            torch.xpu.set_device(local_rank)
        elif torch.cuda.is_available():
            torch.cuda.set_device(local_rank)
        clear_gpu_cache(local_rank)
        setup_environ_flags(rank)

    wandb_run = None

    if train_config.use_wandb:
        if not train_config.enable_fsdp or rank==0:
            wandb_run = setup_wandb(train_config, fsdp_config, **kwargs)

    # Load the pre-trained model and setup its configuration
    use_cache = False if train_config.enable_fsdp else None
    if train_config.enable_fsdp and train_config.low_cpu_fsdp:
        """
        for FSDP, we can save cpu memory by loading pretrained model on rank0 only.
        this avoids cpu oom when loading large models like llama 70B, in which case
        model alone would consume 2+TB cpu mem (70 * 4 * 8). This will add some comms
        overhead and currently requires latest nightly.
        """
        if rank == 0:
            model = LlamaForCausalLM.from_pretrained(
                train_config.model_name,
                load_in_8bit=True if train_config.quantization else None,
                device_map="auto" if train_config.quantization else None,
                use_cache=use_cache,
                attn_implementation="sdpa" if train_config.use_fast_kernels else None,
            )
        else:
            llama_config = LlamaConfig.from_pretrained(train_config.model_name)
            llama_config.use_cache = use_cache
            with torch.device("meta"):
                model = LlamaForCausalLM(llama_config)

    else:
        model = LlamaForCausalLM.from_pretrained(
            train_config.model_name,
            load_in_8bit=True if train_config.quantization else None,
            device_map="auto" if train_config.quantization else None,
            use_cache=use_cache,
            attn_implementation="sdpa" if train_config.use_fast_kernels else None,
        )

    # Load the tokenizer and add special tokens
    tokenizer = AutoTokenizer.from_pretrained(train_config.model_name if train_config.tokenizer_name is None else train_config.tokenizer_name)
    tokenizer.pad_token_id = tokenizer.eos_token_id

    # If there is a mismatch between tokenizer vocab size and embedding matrix,
    # throw a warning and then expand the embedding matrix
    if len(tokenizer) > model.get_input_embeddings().weight.shape[0]:
        print("WARNING: Resizing the embedding matrix to match the tokenizer vocab size.")
        model.resize_token_embeddings(len(tokenizer))

    print_model_size(model, train_config, rank if train_config.enable_fsdp else 0)

    # Prepare the model for int8 training if quantization is enabled
    if train_config.quantization:
        model = prepare_model_for_kbit_training(model)

    # Convert the model to bfloat16 if fsdp and pure_bf16 is enabled
    if train_config.enable_fsdp and fsdp_config.pure_bf16:
        model.to(torch.bfloat16)

    if train_config.use_peft:
        peft_config = generate_peft_config(train_config, kwargs)
        model = get_peft_model(model, peft_config)
        model.print_trainable_parameters()
        if wandb_run:
            wandb_run.config.update(peft_config)


    hsdp_device_mesh = None
    if fsdp_config.hsdp and fsdp_config.sharding_strategy == ShardingStrategy.HYBRID_SHARD:
        hsdp_device_mesh = hsdp_device_mesh(replica_group_size=fsdp_config.replica_group_size, sharding_group_size=fsdp_config.sharding_group_size)
        print("HSDP device mesh is ready")

    #setting up FSDP if enable_fsdp is enabled
    if train_config.enable_fsdp:
        if not train_config.use_peft and train_config.freeze_layers:

            freeze_transformer_layers(train_config.num_freeze_layers)

        mixed_precision_policy, wrapping_policy = get_policies(fsdp_config, rank)
        my_auto_wrapping_policy = fsdp_auto_wrap_policy(model, LlamaDecoderLayer)

        device_id = 0
        if is_xpu_available():
            device_id = torch.xpu.current_device()
        elif torch.cuda.is_available():
            device_id = torch.cuda.current_device()

        model = FSDP(
            model,
            auto_wrap_policy= my_auto_wrapping_policy if train_config.use_peft else wrapping_policy,
            cpu_offload=CPUOffload(offload_params=True) if fsdp_config.fsdp_cpu_offload else None,
            mixed_precision=mixed_precision_policy if not fsdp_config.pure_bf16 else None,
            sharding_strategy=fsdp_config.sharding_strategy,
            device_mesh=hsdp_device_mesh,
            device_id=device_id,
            limit_all_gathers=True,
            sync_module_states=train_config.low_cpu_fsdp,
            param_init_fn=lambda module: module.to_empty(device=torch.device("cuda"), recurse=False)
            if train_config.low_cpu_fsdp and rank != 0 else None,
        )
        if fsdp_config.fsdp_activation_checkpointing:
            apply_fsdp_checkpointing(model)
    elif not train_config.quantization and not train_config.enable_fsdp:
        if is_xpu_available():
            model.to("xpu:0")
        elif torch.cuda.is_available():
            model.to("cuda")

    dataset_config = generate_dataset_config(train_config, kwargs)

     # Load and preprocess the dataset for training and validation
    dataset_train = get_preprocessed_dataset(
        tokenizer,
        dataset_config,
        split="train",
    )

    if not train_config.enable_fsdp or rank == 0:
        print(f"--> Training Set Length = {len(dataset_train)}")

    dataset_val = get_preprocessed_dataset(
        tokenizer,
        dataset_config,
        split="test",
    )
    if not train_config.enable_fsdp or rank == 0:
            print(f"--> Validation Set Length = {len(dataset_val)}")

    if train_config.batching_strategy == "packing":
        dataset_train = ConcatDataset(dataset_train, chunk_size=train_config.context_length)

    train_dl_kwargs = get_dataloader_kwargs(train_config, dataset_train, tokenizer, "train")

    # Create DataLoaders for the training and validation dataset
    train_dataloader = torch.utils.data.DataLoader(
        dataset_train,
        num_workers=train_config.num_workers_dataloader,
        pin_memory=True,
        **train_dl_kwargs,
    )

    eval_dataloader = None
    if train_config.run_validation:
        if train_config.batching_strategy == "packing":
            dataset_val = ConcatDataset(dataset_val, chunk_size=train_config.context_length)

        val_dl_kwargs = get_dataloader_kwargs(train_config, dataset_val, tokenizer, "val")

        eval_dataloader = torch.utils.data.DataLoader(
            dataset_val,
            num_workers=train_config.num_workers_dataloader,
            pin_memory=True,
            **val_dl_kwargs,
        )

    # Initialize the optimizer and learning rate scheduler
    if fsdp_config.pure_bf16 and fsdp_config.optimizer == "anyprecision":
        optimizer = AnyPrecisionAdamW(
            model.parameters(),
            lr=train_config.lr,
            momentum_dtype=torch.bfloat16,
            variance_dtype=torch.bfloat16,
            use_kahan_summation=False,
            weight_decay=train_config.weight_decay,
        )
    else:
        optimizer = optim.AdamW(
            model.parameters(),
            lr=train_config.lr,
            weight_decay=train_config.weight_decay,
        )
    scheduler = StepLR(optimizer, step_size=1, gamma=train_config.gamma)

    # Start the training process
    results = train(
        model,
        train_dataloader,
        eval_dataloader,
        tokenizer,
        optimizer,
        scheduler,
        train_config.gradient_accumulation_steps,
        train_config,
        fsdp_config if train_config.enable_fsdp else None,
        local_rank if train_config.enable_fsdp else None,
        rank if train_config.enable_fsdp else None,
        wandb_run,
    )
    if not train_config.enable_fsdp or rank==0:
        [print(f'Key: {k}, Value: {v}') for k, v in results.items()]
        if train_config.use_wandb:
            for k,v in results.items():
                wandb_run.summary[k] = v

if __name__ == "__main__":
    fire.Fire(main)