accelerate 1.12.0

Accelerate

Stars: 9491, Watchers: 9491, Forks: 1284, Open Issues: 87

The huggingface/accelerate repo was created 5 years ago and the last code push was 9 hours ago.
The project is extremely popular with a mindblowing 9491 github stars!

How to Install accelerate

You can install accelerate using pip

pip install accelerate

or add it to a project with poetry

poetry add accelerate

Package Details

Author

The HuggingFace team

License

Apache

Homepage

https://github.com/huggingface/accelerate

PyPi:

https://pypi.org/project/accelerate/

GitHub Repo:

https://github.com/huggingface/accelerate

Classifiers

• Scientific/Engineering/Artificial Intelligence

Related Packages

Errors

A list of common accelerate errors.

Code Examples

Here are some accelerate code examples and snippets.

GitHub Issues

The accelerate package has 87 open issues on GitHub

• save_function arg in save_pretrained is dropped in transformers 5 RCs
• Fix KMP_AFFINITY incorrectly set for non-CPU training
• Severe GPU training performance degradation due to incorrectly set KMP_AFFINITY
• fix bug when both cpu_ram_efficient_loading and cpu_offload are enabled
• FSDP2: Activation-checkpoint modules directly matching auto-wrap policy, not those whose parent module matches policy
• Fix typo in mult-node gpu example script
• FSDP 2 + cpu_ram_efficient_loading=True fails for models with buffers (e.g. DS 3.1)
• activation_checkpointing=True uses a lot more memory than gradient_checkpointing=True with FSDP 2.0 (FSDP 1.0 is fine)
• Fail to load_state immediately after save_state with FSDP plugins.
• How to specify different number of process per node
• Support stateful_dataloader for IterableDatasetShard
• enhance new codes to XPU, and make them be device agnostic
• accelerate.load_checkpoint_and_dispatch does not load GPT-OSS Models correctly
• Gradient accumulation gives worse results when using DeepSpeed ZeRO 2
• Why TP can't be used with pure DP?

See more issues on GitHub