Accelerate + deepspeed_stage_zero3 Cannot save model checkpoint

[wangyu-ustc]

System Info

accelerate == 1.2.0
deepspeed == 0.16.2

Information

• The official example scripts
• My own modified scripts

Tasks

• One of the scripts in the examples/ folder of Accelerate or an officially supported no_trainer script in the examples folder of the transformers repo (such as run_no_trainer_glue.py)
• My own task or dataset (give details below)

Reproduction

I copied the example from examples/nlp_example.py and renamed it to main.py in a new folder then I replaced the optimizer and scheduler in the file with DummyOptim and DummyScheduler to make it compatible with deepspeed configs. This is fine, the code is running properly. But once I added the following it started to raise errors:

state_dict = accelerator.get_state_dict()

Yes just this line of reading the state_dict will raise the error:

AssertionErrorAssertionError: : DeepSpeed backend not set, please initialize it using init_process_group()DeepSpeed backend not set, please initialize it using init_process_group()

But How exactly should I initialize it?

Additional Information:

1. My main.py:

# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse

import evaluate
import torch
from datasets import load_dataset
from torch.optim import AdamW
from torch.utils.data import DataLoader
from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed

from accelerate import Accelerator, DistributedType
from accelerate.utils import DummyOptim, DummyScheduler


########################################################################
# This is a fully working simple example to use Accelerate
#
# This example trains a Bert base model on GLUE MRPC
# in any of the following settings (with the same script):
#   - single CPU or single GPU
#   - multi GPUS (using PyTorch distributed mode)
#   - (multi) TPUs
#   - fp16 (mixed-precision) or fp32 (normal precision)
#
# To run it in each of these various modes, follow the instructions
# in the readme for examples:
# https://github.com/huggingface/accelerate/tree/main/examples
#
########################################################################


MAX_GPU_BATCH_SIZE = 16
EVAL_BATCH_SIZE = 32


def get_dataloaders(accelerator: Accelerator, batch_size: int = 16):
    """
    Creates a set of `DataLoader`s for the `glue` dataset,
    using "bert-base-cased" as the tokenizer.

    Args:
        accelerator (`Accelerator`):
            An `Accelerator` object
        batch_size (`int`, *optional*):
            The batch size for the train and validation DataLoaders.
    """
    tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
    datasets = load_dataset("glue", "mrpc")

    def tokenize_function(examples):
        # max_length=None => use the model max length (it's actually the default)
        outputs = tokenizer(examples["sentence1"], examples["sentence2"], truncation=True, max_length=None)
        return outputs

    # Apply the method we just defined to all the examples in all the splits of the dataset
    # starting with the main process first:
    with accelerator.main_process_first():
        tokenized_datasets = datasets.map(
            tokenize_function,
            batched=True,
            remove_columns=["idx", "sentence1", "sentence2"],
        )

    # We also rename the 'label' column to 'labels' which is the expected name for labels by the models of the
    # transformers library
    tokenized_datasets = tokenized_datasets.rename_column("label", "labels")

    def collate_fn(examples):
        # For Torchxla, it's best to pad everything to the same length or training will be very slow.
        max_length = 128 if accelerator.distributed_type == DistributedType.XLA else None
        # When using mixed precision we want round multiples of 8/16
        if accelerator.mixed_precision == "fp8":
            pad_to_multiple_of = 16
        elif accelerator.mixed_precision != "no":
            pad_to_multiple_of = 8
        else:
            pad_to_multiple_of = None

        return tokenizer.pad(
            examples,
            padding="longest",
            max_length=max_length,
            pad_to_multiple_of=pad_to_multiple_of,
            return_tensors="pt",
        )

    # Instantiate dataloaders.
    train_dataloader = DataLoader(
        tokenized_datasets["train"], shuffle=True, collate_fn=collate_fn, batch_size=batch_size, drop_last=True
    )
    eval_dataloader = DataLoader(
        tokenized_datasets["validation"],
        shuffle=False,
        collate_fn=collate_fn,
        batch_size=EVAL_BATCH_SIZE,
        drop_last=(accelerator.mixed_precision == "fp8"),
    )

    return train_dataloader, eval_dataloader


def training_function(config, args):
    # Initialize accelerator
    accelerator = Accelerator(cpu=args.cpu, mixed_precision=args.mixed_precision)
    # Sample hyper-parameters for learning rate, batch size, seed and a few other HPs
    lr = config["lr"]
    num_epochs = int(config["num_epochs"])
    seed = int(config["seed"])
    batch_size = int(config["batch_size"])

    metric = evaluate.load("glue", "mrpc")

    # If the batch size is too big we use gradient accumulation
    gradient_accumulation_steps = 1
    if batch_size > MAX_GPU_BATCH_SIZE and accelerator.distributed_type != DistributedType.XLA:
        gradient_accumulation_steps = batch_size // MAX_GPU_BATCH_SIZE
        batch_size = MAX_GPU_BATCH_SIZE

    set_seed(seed)
    train_dataloader, eval_dataloader = get_dataloaders(accelerator, batch_size)
    # Instantiate the model (we build the model here so that the seed also control new weights initialization)
    model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", return_dict=True)

    # We could avoid this line since the accelerator is set with `device_placement=True` (default value).
    # Note that if you are placing tensors on devices manually, this line absolutely needs to be before the optimizer
    # creation otherwise training will not work on TPU (`accelerate` will kindly throw an error to make us aware of that).
    model = model.to(accelerator.device)
    # Instantiate optimizer
    optimizer = DummyOptim(params=model.parameters(), lr=lr)

    # Instantiate scheduler
    # lr_scheduler = get_linear_schedule_with_warmup(
    lr_scheduler = DummyScheduler(
        optimizer=optimizer,
        # num_warmup_steps=100,
        # num_training_steps=(len(train_dataloader) * num_epochs) // gradient_accumulation_steps,
        num_training_steps=args.max_train_steps,
        total_num_steps=args.max_train_steps,
    )

    # Prepare everything
    # There is no specific order to remember, we just need to unpack the objects in the same order we gave them to the
    # prepare method.

    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = accelerator.prepare(
        model, optimizer, train_dataloader, eval_dataloader, lr_scheduler
    )

    # Now we train the model
    for epoch in range(num_epochs):
        model.train()
        for step, batch in enumerate(train_dataloader):
            # We could avoid this line since we set the accelerator with `device_placement=True`.
            batch.to(accelerator.device)
            outputs = model(**batch)
            loss = outputs.loss
            loss = loss / gradient_accumulation_steps
            accelerator.backward(loss)
            if step % gradient_accumulation_steps == 0:
                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad()

        model.eval()
        for step, batch in enumerate(eval_dataloader):
            # We could avoid this line since we set the accelerator with `device_placement=True`.
            batch.to(accelerator.device)
            with torch.no_grad():
                outputs = model(**batch)
            predictions = outputs.logits.argmax(dim=-1)
            predictions, references = accelerator.gather_for_metrics((predictions, batch["labels"]))
            metric.add_batch(
                predictions=predictions,
                references=references,
            )

        eval_metric = metric.compute()
        # Use accelerator.print to print only on the main process.
        accelerator.print(f"epoch {epoch}:", eval_metric)
    accelerator.end_training()

    # accelerator.wait_for_everyone()
    state_dict = accelerator.get_state_dict(model)
    accelerator.unwrap_model(model).save_pretrained(
        f"try",
        is_main_process=accelerator.is_main_process,
        save_function=accelerator.save,
        state_dict=state_dict,
    )


def main():
    parser = argparse.ArgumentParser(description="Simple example of training script.")
    parser.add_argument(
        "--mixed_precision",
        type=str,
        default=None,
        choices=["no", "fp16", "bf16", "fp8"],
        help="Whether to use mixed precision. Choose"
        "between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >= 1.10."
        "and an Nvidia Ampere GPU.",
    )
    parser.add_argument(
        "--max_train_steps",
        type=int,
        default=100,
    )
    parser.add_argument("--cpu", action="store_true", help="If passed, will train on the CPU.")
    args = parser.parse_args()
    config = {"lr": 2e-5, "num_epochs": 1, "seed": 42, "batch_size": 16}
    training_function(config, args)


if __name__ == "__main__":
    main()

2. Put the following file into the same folder
   deepspeed_zero3.json

{
    "bf16": {
        "enabled": true,
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },
    "optimizer": {
        "type": "AdamW",
        "params": {
            "lr": "auto",
            "weight_decay": "auto",
            "torch_adam": true,
            "adam_w_mode": true
        }
    },
    "scheduler": {
        "type": "WarmupDecayLR",
        "params": {
            "warmup_min_lr": "auto",
            "warmup_max_lr": "auto",
            "warmup_num_steps": "auto",
            "total_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_param_persistence_threshold": "auto",
        "stage3_max_live_parameters": 1e9,
        "stage3_max_reuse_distance": 1e9,
        "stage3_gather_16bit_weights_on_model_save": true
      },
    "gradient_accumulation_steps": 1,
    "gradient_clipping": "auto",
    "steps_per_print": 2000,
    "train_batch_size": "auto",
    "train_micro_batch_size_per_gpu": "auto",
    "wall_clock_breakdown": false
}

3. Put the following content into config.yaml:

compute_environment: LOCAL_MACHINE
debug: true
deepspeed_config:
  deepspeed_config_file: deepspeed_zero3.json
  zero3_init_flag: true
distributed_type: DEEPSPEED
downcast_bf16: 'no'
dynamo_config:
  dynamo_backend: INDUCTOR
  dynamo_mode: default
  dynamo_use_dynamic: true
  dynamo_use_fullgraph: true
enable_cpu_affinity: false
machine_rank: 0
main_training_function: main
num_machines: 1
num_processes: 4
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false

Then run accelerate launch --config_file config.yaml main.py, it will raise the error AssertionError: DeepSpeed backend not set, please initialize it using init_process_group()
I really do not know how to initialize deepspeed in this setting (I thought accelerate should help me initialize but it doesn't seem to be doing that).

Expected behavior

I would expect to get the state_dict and save it.