Skip to content

Latest commit

 

History

History
88 lines (51 loc) · 2.92 KB

File metadata and controls

88 lines (51 loc) · 2.92 KB

TAGLINE

distributed PyTorch training launcher

TLDR

Launch a training script with default configuration

accelerate launch [train.py]

Configure accelerate for your hardware

accelerate config

Launch with specific GPU configuration

accelerate launch --num_processes [4] --gpu_ids [0,1,2,3] [train.py]

Launch training on multiple machines

accelerate launch --num_machines [2] --machine_rank [0] --main_process_ip [192.168.1.1] [train.py]

SYNOPSIS

accelerate command [options] [script] [script_args]

DESCRIPTION

accelerate is a Hugging Face library that enables PyTorch code to run on any distributed configuration with minimal code changes. It handles the complexity of distributed training across multiple GPUs, TPUs, and machines while keeping your training code simple.

The tool abstracts away the boilerplate needed for mixed precision training, gradient accumulation, and multi-device parallelism. It automatically detects available hardware and configures the training environment appropriately.

PARAMETERS

config

Run the configuration wizard to set up your environment

launch

Launch a training script with the configured settings

--num_processes n

Total number of processes to launch

--gpu_ids ids

Comma-separated GPU IDs to use

--mixed_precision type

Enable mixed precision: no, fp16, bf16

--num_machines n

Number of machines for distributed training

--machine_rank n

Rank of the current machine (0-indexed)

--main_process_ip ip

IP address of the main machine

--main_process_port port

Port for the main machine (default: 29500)

--use_deepspeed

Enable DeepSpeed for training

--use_fsdp

Enable Fully Sharded Data Parallel

test

Test your accelerate configuration

env

Print environment information

CONFIGURATION

Running accelerate config creates a YAML configuration file at ~/.cache/huggingface/accelerate/default_config.yaml. This file stores settings for compute environment type, distributed training backend, number of processes, mixed precision mode, and DeepSpeed/FSDP options. The configuration can also be specified per-project by placing an accelerate_config.yaml in the project directory or by passing --config_file to the launch command.

CAVEATS

Requires PyTorch to be installed. Configuration should match your actual hardware; mismatches can cause silent failures or crashes. DeepSpeed and FSDP have additional dependencies. Some features require specific GPU architectures (e.g., bf16 requires Ampere or newer).

HISTORY

accelerate was developed by Hugging Face and first released in 2021. It was created to simplify distributed training and mixed precision workflows, reducing the barrier to training large models on diverse hardware configurations.

SEE ALSO

python(1), torchrun(1), deepspeed(1)