I am attempting Multi-node training of Falcon 7B with Peft using DeepSpeed and Accelerate. During single node training, it takes up 39GB of GPU memory. However, in multi-node training, both machines consume 40GB of memory. Shouldn't it reduce memory usage?

Expected behavior Reduce the memory usage

ds_report output

[2023-07-31 08:56:06,593] [INFO] [real_accelerator.py:133:get_accelerator] Setting ds_accelerator to cuda (auto detect)

DeepSpeed C++/CUDA extension op report

NOTE: Ops not installed will be just-in-time (JIT) compiled at runtime if needed. Op compatibility means that your system meet the required dependencies to JIT install the op.

JIT compiled ops requires ninja ninja .................. [OKAY]

op name ................ installed .. compatible

async_io ............... [YES] ...... [OKAY] cpu_adagrad ............ [NO] ....... [OKAY] cpu_adam ............... [YES] ...... [OKAY] fused_adam ............. [NO] ....... [OKAY] fused_lamb ............. [NO] ....... [OKAY] quantizer .............. [NO] ....... [OKAY] random_ltd ............. [NO] ....... [OKAY] [WARNING] sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.0 [WARNING] using untested triton version (2.1.0+9e3e10c5ed), only 1.0.0 is known to be compatible sparse_attn ............ [NO] ....... [NO] spatial_inference ...... [NO] ....... [OKAY] transformer ............ [NO] ....... [OKAY] stochastic_transformer . [NO] ....... [OKAY] transformer_inference .. [NO] ....... [OKAY]

DeepSpeed general environment info: torch install path ............... ['/anaconda3/envs/venv/lib/python3.10/site-packages/torch'] torch version .................... 2.0.0+cu117 deepspeed install path ........... ['/anaconda3/envs/venv/lib/python3.10/site-packages/deepspeed'] deepspeed info ................... 0.10.0, unknown, unknown torch cuda version ............... 11.7 torch hip version ................ None nvcc version ..................... 12.0 deepspeed wheel compiled w. ...... torch 2.0, cuda 11.7

System info (please complete the following information):

OS: Red Hat Enterprise Linux Server 7.8 (Maipo)
2 machines with 1 GPU in each (A100-80GB)
Interconnects (if applicable) [e.g., two machines connected with 100 Gbps IB]
Python version = 3.10.12

Launcher context

accelerate launch --config_file ds_zero3_multinode.yaml run_clm_no_trainer_lora.py --model_name_or_path "tiiuae/falcon-7b" --dataset_name "train.json" --block_size 2048 --learning_rate 3e-5 --per_device_train_batch_size 1 --per_device_eval_batch_size 1 --num_train_epochs 10 --num_warmup_steps 2000 --checkpointing_steps 1000 --preprocessing_num_workers 8 --with_tracking --output_dir "output/lora_test" --report_to "tensorboard"

The accelerate congiruation file

compute_environment: LOCAL_MACHINE deepspeed_config: deepspeed_config_file: zero_stage3_offload_config.json deepspeed_hostfile: /path/to/hostfile deepspeed_multinode_launcher: pdsh zero3_init_flag: false distributed_type: DEEPSPEED downcast_bf16: 'no' machine_rank: 0 main_process_ip: main_process_port: main_training_function: main num_machines: 2 num_processes: 2 rdzv_backend: static same_network: true tpu_env: [] tpu_use_cluster: false tpu_use_sudo: false use_cpu: false

host file machine1 slots=1 machine2 slots=1

Deepspeed conf file : ds_zero3_multinode.yaml

{ "fp16": { "enabled": false, "loss_scale": 1024, "loss_scale_window": 1000, "initial_scale_power": 4, "hysteresis": 2, "min_loss_scale": 1 }, "bf16": { "enabled":true }, "optimizer": { "type": "Adamw", "params": { "lr": "auto", "weight_decay": "auto" } }, "scheduler": { "type": "WarmupLR", "params": { "warmup_min_lr": "auto", "warmup_max_lr": "auto", "warmup_num_steps": "auto" } }, "zero_optimization": { "stage": 3, "offload_optimizer": { "device": "cpu", "pin_memory": true }, "offload_param": { "device": "cpu", "pin_memory": true }, "overlap_comm": true, "contiguous_gradients": true, "sub_group_size": 1e9, "reduce_bucket_size": "auto", "stage3_prefetch_bucket_size": "auto", "stage3_param_persistence_threshold": "auto", "stage3_max_live_parameters": 1e7, "stage3_max_reuse_distance": 1e7, "stage3_gather_16bit_weights_on_model_save": true }, "gradient_accumulation_steps": "auto", "gradient_clipping": "auto", "steps_per_print": 1, "train_batch_size": "auto", "train_micro_batch_size_per_gpu": "auto", "wall_clock_breakdown": true }

code

import argparse import json import logging import math import os import random from itertools import chain from pathlib import Path

import datasets import torch from accelerate import Accelerator, DistributedType from accelerate.logging import get_logger from accelerate.utils import set_seed from datasets import load_dataset from huggingface_hub import Repository, create_repo from torch.utils.data import DataLoader from tqdm.auto import tqdm import psutil import gc import threading import transformers from transformers import ( CONFIG_MAPPING, MODEL_MAPPING, AutoConfig, AutoModelForCausalLM, AutoTokenizer, SchedulerType, default_data_collator, get_scheduler, ) from transformers.utils import check_min_version, get_full_repo_name, send_example_telemetry from transformers.utils.versions import require_version from peft import LoraConfig, TaskType, get_peft_model from peft.utils.other import fsdp_auto_wrap_policy from accelerate.utils import DummyOptim,DummyScheduler from deepspeed.runtime.utils import see_memory_usage

logger = get_logger(name)

MODEL_CONFIG_CLASSES = list(MODEL_MAPPING.keys()) MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)

def b2mb(x): return int(x / 2**20)

def parse_args(): parser = argparse.ArgumentParser(description="Finetune a transformers model on a causal language modeling task") parser.add_argument( "--dataset_name", type=str, default=None, help="The name of the dataset to use (via the datasets library).", ) parser.add_argument( "--dataset_config_name", type=str, default=None, help="The configuration name of the dataset to use (via the datasets library).", ) parser.add_argument( "--train_file", type=str, default=None, help="A csv or a json file containing the training data." ) parser.add_argument( "--validation_file", type=str, default=None, help="A csv or a json file containing the validation data." ) parser.add_argument( "--validation_split_percentage", default=1, help="The percentage of the train set used as validation set in case there's no validation split", ) parser.add_argument( "--model_name_or_path", type=str, help="Path to pretrained model or model identifier from huggingface.co/models.", required=False, ) parser.add_argument( "--config_name", type=str, default=None, help="Pretrained config name or path if not the same as model_name", ) parser.add_argument( "--tokenizer_name", type=str, default=None, help="Pretrained tokenizer name or path if not the same as model_name", ) parser.add_argument( "--use_slow_tokenizer", action="store_true", help="If passed, will use a slow tokenizer (not backed by the 🤗 Tokenizers library).", ) parser.add_argument( "--per_device_train_batch_size", type=int, default=8, help="Batch size (per device) for the training dataloader.", ) parser.add_argument( "--per_device_eval_batch_size", type=int, default=8, help="Batch size (per device) for the evaluation dataloader.", ) parser.add_argument( "--learning_rate", type=float, default=5e-5, help="Initial learning rate (after the potential warmup period) to use.", ) parser.add_argument("--weight_decay", type=float, default=0.0, help="Weight decay to use.") parser.add_argument("--num_train_epochs", type=int, default=3, help="Total number of training epochs to perform.") parser.add_argument( "--max_train_steps", type=int, default=None, help="Total number of training steps to perform. If provided, overrides num_train_epochs.", ) parser.add_argument( "--gradient_accumulation_steps", type=int, default=1, help="Number of updates steps to accumulate before performing a backward/update pass.", ) parser.add_argument( "--lr_scheduler_type", type=SchedulerType, default="linear", help="The scheduler type to use.", choices=["linear", "cosine", "cosine_with_restarts", "polynomial", "constant", "constant_with_warmup"], ) parser.add_argument( "--num_warmup_steps", type=int, default=0, help="Number of steps for the warmup in the lr scheduler." ) parser.add_argument("--output_dir", type=str, default=None, help="Where to store the final model.") parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.") parser.add_argument( "--model_type", type=str, default=None, help="Model type to use if training from scratch.", choices=MODEL_TYPES, ) parser.add_argument( "--block_size", type=int, default=None, help=( "Optional input sequence length after tokenization. The training dataset will be truncated in block of" " this size for training. Default to the model max input length for single sentence inputs (take into" " account special tokens)." ), ) parser.add_argument("--use_group_texts",action="store_true") parser.add_argument( "--preprocessing_num_workers", type=int, default=None, help="The number of processes to use for the preprocessing.", ) parser.add_argument( "--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets" ) parser.add_argument( "--no_keep_linebreaks", action="store_true", help="Do not keep line breaks when using TXT files." ) parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.") parser.add_argument( "--hub_model_id", type=str, help="The name of the repository to keep in sync with the local output_dir." ) parser.add_argument("--hub_token", type=str, help="The token to use to push to the Model Hub.") parser.add_argument( "--checkpointing_steps", type=str, default=None, help="Whether the various states should be saved at the end of every n steps, or 'epoch' for each epoch.", ) parser.add_argument( "--resume_from_checkpoint", type=str, default=None, help="If the training should continue from a checkpoint folder.", ) parser.add_argument( "--with_tracking", action="store_true", help="Whether to enable experiment trackers for logging.", ) parser.add_argument( "--report_to", type=str, default="all", help=( 'The integration to report the results and logs to. Supported platforms are "tensorboard",' ' "wandb", "comet_ml" and "clearml". Use "all" (default) to report to all integrations.' "Only applicable when --with_tracking is passed." ), ) args = parser.parse_args()

Sanity checks

  if args.dataset_name is None and args.train_file is None and args.validation_file is None:
      raise ValueError("Need either a dataset name or a training/validation file.")
  else:
      if args.train_file is not None:
          extension = args.train_file.split(".")[-1]
          assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, json or txt file."
      if args.validation_file is not None:
          extension = args.validation_file.split(".")[-1]
          assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, json or txt file."

  if args.push_to_hub:
      assert args.output_dir is not None, "Need an `output_dir` to create a repo when `--push_to_hub` is passed."

  return args

def main(): args = parse_args()

  # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
  # information sent is the one passed as arguments along with your Python/PyTorch versions.
  send_example_telemetry("run_clm_no_trainer", args)

  # Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
  # If we're using tracking, we also need to initialize it here and it will by default pick up all supported trackers
  # in the environment
  accelerator_log_kwargs = {}

  if args.with_tracking:
      accelerator_log_kwargs["log_with"] = args.report_to
      # accelerator_log_kwargs["logging_dir"] = args.output_dir

  accelerator = Accelerator(gradient_accumulation_steps=args.gradient_accumulation_steps, **accelerator_log_kwargs)

  # Make one log on every process with the configuration for debugging.
  logging.basicConfig(
      format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
      datefmt="%m/%d/%Y %H:%M:%S",
      level=logging.INFO,
  )
  logger.info(accelerator.state, main_process_only=False)
  if accelerator.is_local_main_process:
      datasets.utils.logging.set_verbosity_warning()
      transformers.utils.logging.set_verbosity_info()
  else:
      datasets.utils.logging.set_verbosity_error()
      transformers.utils.logging.set_verbosity_error()

  # If passed along, set the training seed now.
  if args.seed is not None:
      set_seed(args.seed)

  # Handle the repository creation
  if accelerator.is_main_process:
      if args.push_to_hub:
          if args.hub_model_id is None:
              repo_name = get_full_repo_name(Path(args.output_dir).name, token=args.hub_token)
          else:
              repo_name = args.hub_model_id
          create_repo(repo_name, exist_ok=True, token=args.hub_token)
          repo = Repository(args.output_dir, clone_from=repo_name, token=args.hub_token)

          with open(os.path.join(args.output_dir, ".gitignore"), "w+") as gitignore:
              if "step_*" not in gitignore:
                  gitignore.write("step_*\n")
              if "epoch_*" not in gitignore:
                  gitignore.write("epoch_*\n")
      elif args.output_dir is not None:
          os.makedirs(args.output_dir, exist_ok=True)
  accelerator.wait_for_everyone()

  # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
  # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
  # (the dataset will be downloaded automatically from the datasets Hub).
  #
  # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
  # 'text' is found. You can easily tweak this behavior (see below).
  #
  # In distributed training, the load_dataset function guarantee that only one local process can concurrently
  # download the dataset.
  if args.dataset_name is not None:
      # Downloading and loading a dataset from the hub.
      raw_datasets = load_dataset("json",data_files=args.dataset_name)
      if "validation" not in raw_datasets.keys():
          raw_datasets["validation"] = load_dataset(
              "json",
              data_files=args.dataset_name,
              split=f"train[:{args.validation_split_percentage}%]",
          )
          raw_datasets["train"] = load_dataset(
              "json",
              data_files=args.dataset_name,
              split=f"train[{args.validation_split_percentage}%:]",
          )
  else:
      data_files = {}
      dataset_args = {}
      if args.train_file is not None:
          data_files["train"] = args.train_file
      if args.validation_file is not None:
          data_files["validation"] = args.validation_file
      extension = args.train_file.split(".")[-1]
      if extension == "txt":
          extension = "text"
          dataset_args["keep_linebreaks"] = not args.no_keep_linebreaks
      raw_datasets = load_dataset(extension, data_files=data_files, **dataset_args)
      # If no validation data is there, validation_split_percentage will be used to divide the dataset.
      if "validation" not in raw_datasets.keys():
          raw_datasets["validation"] = load_dataset(
              extension,
              data_files=data_files,
              split=f"train[:{args.validation_split_percentage}%]",
              **dataset_args,
          )
          raw_datasets["train"] = load_dataset(
              extension,
              data_files=data_files,
              split=f"train[{args.validation_split_percentage}%:]",
              **dataset_args,
          )

  # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
  # https://huggingface.co/docs/datasets/loading_datasets.html.

  # Load pretrained model and tokenizer
  #
  # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
  # download model & vocab.
  if args.config_name:
      config = AutoConfig.from_pretrained(args.config_name,trust_remote_code=True)
  elif args.model_name_or_path:
      config = AutoConfig.from_pretrained(args.model_name_or_path,trust_remote_code=True)
  else:
      config = CONFIG_MAPPING[args.model_type]()
      logger.warning("You are instantiating a new config instance from scratch.")

  if args.tokenizer_name:
      tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, use_fast=not args.use_slow_tokenizer,trust_remote_code=True,cache_dir = "../cache",)
  elif args.model_name_or_path:
      tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=not args.use_slow_tokenizer,trust_remote_code=True,cache_dir = "../cache",)
  else:
      raise ValueError(
          "You are instantiating a new tokenizer from scratch. This is not supported by this script."
          "You can do it from another script, save it, and load it from here, using --tokenizer_name."
      )
  tokenizer.pad_token = tokenizer.eos_token
  if args.model_name_or_path:
      model = AutoModelForCausalLM.from_pretrained(
          args.model_name_or_path,
          from_tf=bool(".ckpt" in args.model_name_or_path),
          config=config,
          trust_remote_code=True,
          cache_dir = "../cache",
      )
  else:
      logger.info("Training new model from scratch")
      model = AutoModelForCausalLM.from_config(config)

  # Preprocessing the datasets.
  # First we tokenize all the texts.
  column_names = raw_datasets["train"].column_names
  #text_column_name = "text" if "text" in column_names else column_names[0]

  def generate_prompt(data_point):
      data_point["full_prompt"]= f"""Below is an instruction that describes a task, paired with context and input. Write a response that appropriately completes the request.  # noqa: E501
  ### Instruction:
  {data_point["instruction"]}
  ### Context:
  {data_point["context"]}
  ### Input:
  {data_point["input"]}
  ### Response:
  {data_point["output"]}"""
      return data_point

  raw_datasets =raw_datasets.map(generate_prompt)

  def tokenize_function(examples):
      result = tokenizer(examples["full_prompt"],
                       padding='max_length' if not args.use_group_texts else False,
                       truncation=True if not args.use_group_texts else False,
                       max_length = args.block_size)
      result["labels"] = result["input_ids"].copy()
      return result

  with accelerator.main_process_first():
      tokenized_datasets = raw_datasets.map(
          tokenize_function,
          batched=True,
          num_proc=args.preprocessing_num_workers,
          remove_columns=column_names,
          load_from_cache_file=not args.overwrite_cache,
          desc="Running tokenizer on dataset",
      )

  if args.block_size is None:
      block_size = tokenizer.model_max_length
      # if block_size > 1024:
      #     logger.warning(
      #         "The chosen tokenizer supports a `model_max_length` that is longer than the default `block_size` value"
      #         " of 1024. If you would like to use a longer `block_size` up to `tokenizer.model_max_length` you can"
      #         " override this default with `--block_size xxx`."
      #     )
      # block_size = 1024
  else:
      if args.block_size > tokenizer.model_max_length:
          logger.warning(
              f"The block_size passed ({args.block_size}) is larger than the maximum length for the model"
              f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}."
          )
      block_size = min(args.block_size, tokenizer.model_max_length)

  # Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
  def group_texts(examples):
      # Concatenate all texts.
      concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
      total_length = len(concatenated_examples[list(examples.keys())[0]])
      # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
      # customize this part to your needs.
      if total_length >= block_size:
          total_length = (total_length // block_size) * block_size
      # Split by chunks of max_len.
      result = {
          k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
          for k, t in concatenated_examples.items()
      }
      result["labels"] = result["input_ids"].copy()
      return result

  # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder
  # for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value might be slower
  # to preprocess.
  #
  # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
  # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map

  with accelerator.main_process_first():
      if args.use_group_texts:
          lm_datasets = tokenized_datasets.map(
              group_texts,
              batched=True,
              num_proc=args.preprocessing_num_workers,
              load_from_cache_file=not args.overwrite_cache,
              desc=f"Grouping texts in chunks of {block_size}",
          )
      else:
          lm_datasets = tokenized_datasets
  train_dataset = lm_datasets["train"]
  eval_dataset = lm_datasets["validation"]
  train_dataset =train_dataset.remove_columns(["token_type_ids"])
  eval_dataset =eval_dataset.remove_columns(["token_type_ids"])
  # Log a few random samples from the training set:
  for index in random.sample(range(len(train_dataset)), 3):
      logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")

  # DataLoaders creation:
  train_dataloader = DataLoader(
      train_dataset, shuffle=True, collate_fn=default_data_collator, batch_size=args.per_device_train_batch_size
  )

  eval_dataloader = DataLoader(
      eval_dataset, collate_fn=default_data_collator, batch_size=args.per_device_eval_batch_size
  )
  LORA_TARGET_MODULES = [
      "query_key_value",
      "dense"
  ]
  peft_config = LoraConfig(
      task_type=TaskType.CAUSAL_LM, inference_mode=False, r=8, lora_alpha=32, lora_dropout=0.1,target_modules=LORA_TARGET_MODULES,bias="none"
  )
  model = get_peft_model(model, peft_config)

  # Optimizer
  # Split weights in two groups, one with weight decay and the other not.
  # no_decay = ["bias", "layer_norm.weight"]
  # optimizer_grouped_parameters = [
  #     {
  #         "params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
  #         "weight_decay": args.weight_decay,
  #     },
  #     {
  #         "params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)],
  #         "weight_decay": 0.0,
  #     },
  # ]

  # Creates Dummy Optimizer if `optimizer` was spcified in the config file else creates Adam Optimizer
  optimizer_cls = (
      torch.optim.AdamW
      if accelerator.state.deepspeed_plugin is None
         or "optimizer" not in accelerator.state.deepspeed_plugin.deepspeed_config
      else DummyOptim
  )
  optimizer = optimizer_cls(model.parameters(), lr=args.learning_rate)

  # Scheduler and math around the number of training steps.
  overrode_max_train_steps = False
  num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
  if args.max_train_steps is None:
      args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
      overrode_max_train_steps = True

      # Creates Dummy Scheduler if `scheduler` was spcified in the config file else creates `args.lr_scheduler_type` Scheduler
  if (
          accelerator.state.deepspeed_plugin is None
          or "scheduler" not in accelerator.state.deepspeed_plugin.deepspeed_config
  ):
      lr_scheduler = get_scheduler(
          name=args.lr_scheduler_type,
          optimizer=optimizer,
          num_warmup_steps=args.num_warmup_steps * args.gradient_accumulation_steps,
          num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
      )
  else:
      lr_scheduler = DummyScheduler(
          optimizer, total_num_steps=args.max_train_steps, warmup_num_steps=args.num_warmup_steps
      )

  is_ds_zero_3 = False
  if getattr(accelerator.state, "deepspeed_plugin", None):
      is_ds_zero_3 = accelerator.state.deepspeed_plugin.zero_stage == 3

  # Prepare everything with our `accelerator`.
  model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = accelerator.prepare(
      model, optimizer, train_dataloader, eval_dataloader, lr_scheduler
  )
  accelerator.print(model.print_trainable_parameters())
  # On TPU, the tie weights in our model have been disconnected, so we need to restore the ties.
  if accelerator.distributed_type == DistributedType.TPU:
      model.tie_weights()

  # We need to recalculate our total training steps as the size of the training dataloader may have changed.
  num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
  if overrode_max_train_steps:
      args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
  # Afterwards we recalculate our number of training epochs
  args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)

  # Figure out how many steps we should save the Accelerator states
  checkpointing_steps = args.checkpointing_steps
  if checkpointing_steps is not None and checkpointing_steps.isdigit():
      checkpointing_steps = int(checkpointing_steps)

  # We need to initialize the trackers we use, and also store our configuration.
  # The trackers initializes automatically on the main process.
  if args.with_tracking:
      experiment_config = vars(args)
      # TensorBoard cannot log Enums, need the raw value
      experiment_config["lr_scheduler_type"] = experiment_config["lr_scheduler_type"].value
      accelerator.init_trackers("clm_no_trainer", experiment_config)

  # Train!
  total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps

  logger.info("***** Running training *****")
  logger.info(f"  Num examples = {len(train_dataset)}")
  logger.info(f"  Num Epochs = {args.num_train_epochs}")
  logger.info(f"  Instantaneous batch size per device = {args.per_device_train_batch_size}")
  logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
  logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
  logger.info(f"  Total optimization steps = {args.max_train_steps}")
  # Only show the progress bar once on each machine.
  progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process)
  completed_steps = 0
  starting_epoch = 0

  # Potentially load in the weights and states from a previous save
  if args.resume_from_checkpoint:
      if args.resume_from_checkpoint is not None or args.resume_from_checkpoint != "":
          accelerator.print(f"Resumed from checkpoint: {args.resume_from_checkpoint}")
          accelerator.load_state(args.resume_from_checkpoint)
          path = os.path.basename(args.resume_from_checkpoint)
      else:
          # Get the most recent checkpoint
          dirs = [f.name for f in os.scandir(os.getcwd()) if f.is_dir()]
          dirs.sort(key=os.path.getctime)
          path = dirs[-1]  # Sorts folders by date modified, most recent checkpoint is the last
      # Extract `epoch_{i}` or `step_{i}`
      training_difference = os.path.splitext(path)[0]

      if "epoch" in training_difference:
          starting_epoch = int(training_difference.replace("epoch_", "")) + 1
          resume_step = None
      else:
          # need to multiply `gradient_accumulation_steps` to reflect real steps
          resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
          starting_epoch = resume_step // len(train_dataloader)
          resume_step -= starting_epoch * len(train_dataloader)

  # update the progress_bar if load from checkpoint
  progress_bar.update(starting_epoch * num_update_steps_per_epoch)
  completed_steps = starting_epoch * num_update_steps_per_epoch
  with accelerator.main_process_first():
      torch.cuda.empty_cache()
  for epoch in range(starting_epoch, args.num_train_epochs):
      model.train()
      if args.with_tracking:
          total_loss = 0
      for step, batch in enumerate(train_dataloader):
          # We need to skip steps until we reach the resumed step
          if args.resume_from_checkpoint and epoch == starting_epoch:
              if resume_step is not None and step < resume_step:
                  if step % args.gradient_accumulation_steps == 0:
                      progress_bar.update(1)
                      completed_steps += 1
                  continue
          # with accelerator.accumulate(model):
          see_memory_usage(f'before forward {model.global_steps}', force=True)
          outputs = model(**batch)
          loss = outputs.loss
          # We keep track of the loss at each epoch
          if args.with_tracking:
              total_loss += loss.detach().cpu().float()
          see_memory_usage(f'before backward {model.global_steps}', force=True)
          accelerator.backward(loss)
          see_memory_usage(f'before optimizer {model.global_steps}', force=True)
          optimizer.step()
          lr_scheduler.step()
          optimizer.zero_grad()
          see_memory_usage(f'after optimizer {model.global_steps}', force=True)
          # Checks if the accelerator has performed an optimization step behind the scenes
          # if accelerator.sync_gradients:
          progress_bar.update(1)
          completed_steps += 1
          if isinstance(checkpointing_steps, int):
              if completed_steps % checkpointing_steps == 0:
                  output_dir = f"step_{completed_steps }"
                  if args.output_dir is not None:
                      output_dir = os.path.join(args.output_dir, output_dir)
                  accelerator.save_state(output_dir)
          if completed_steps >= args.max_train_steps:
              break

      model.eval()
      losses = []
      for step, batch in enumerate(eval_dataloader):
          with torch.no_grad():
              outputs = model(**batch)

          loss = outputs.loss
          losses.append(accelerator.gather_for_metrics(loss.repeat(args.per_device_eval_batch_size)))

      losses = torch.cat(losses)
      try:
          eval_loss = torch.mean(losses)
          perplexity = math.exp(eval_loss)
      except OverflowError:
          perplexity = float("inf")

      logger.info(f"epoch {epoch}: perplexity: {perplexity} eval_loss: {eval_loss}")

      if args.with_tracking:
          accelerator.log(
              {
                  "perplexity": perplexity,
                  "eval_loss": eval_loss,
                  "train_loss": total_loss.item() / len(train_dataloader),
                  "epoch": epoch,
                  "step": completed_steps,
              },
              step=completed_steps,
          )

      if args.push_to_hub and epoch < args.num_train_epochs - 1:
          accelerator.wait_for_everyone()
          unwrapped_model = accelerator.unwrap_model(model)
          unwrapped_model.save_pretrained(
              args.output_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save
          )
          if accelerator.is_main_process:
              tokenizer.save_pretrained(args.output_dir)
              repo.push_to_hub(
                  commit_message=f"Training in progress epoch {epoch}", blocking=False, auto_lfs_prune=True
              )

      if args.checkpointing_steps == "epoch":
          output_dir = f"epoch_{epoch}"
          if args.output_dir is not None:
              output_dir = os.path.join(args.output_dir, output_dir)
          accelerator.save_state(output_dir)

  if args.with_tracking:
      accelerator.end_training()

  if args.output_dir is not None:
      accelerator.wait_for_everyone()
      unwrapped_model = accelerator.unwrap_model(model)
      unwrapped_model.save_pretrained(
          args.output_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save
      )
      if accelerator.is_main_process:
          tokenizer.save_pretrained(args.output_dir)
          if args.push_to_hub:
              repo.push_to_hub(commit_message="End of training", auto_lfs_prune=True)

          with open(os.path.join(args.output_dir, "all_results.json"), "w") as f:
              json.dump({"perplexity": perplexity}, f)

if name == "main": main()

Log log.txt

microsoft / DeepSpeed

Deepspeed Multinode training doesnt reduce memory usage #4066