Accelerate
Hugging FaceAccelerate is a Hugging Face PyTorch library that lets users run raw PyTorch training scripts across CPUs, multi-GPU, and TPU configurations with mixed precision, by adding only a few lines of code. It abstracts the boilerplate of distributed setup and device placement while leaving the training loop under the user's control, and provides a CLI for configuration. It is widely used as the distributed-training backbone for other Hugging Face libraries.
Verified live 2026-06-22 via primary sources. LICENSE file is the standard, unmodified Apache 2.0 text.
Openness
5 high confidence- license
- Apache-2.0(OSI)
- source
- public
- core-gated
- ungated
LICENSE file is the standard, unmodified Apache 2.0 text.
- https://github.com/huggingface/accelerate/blob/main/LICENSE recorded 2026-06-22
Standard Apache License 2.0 full text
Adoption
5 high confidenceRe-scored 4->5. 25.01M PyPI downloads in the last 30 days. The map's prevailing usage_volume level-5 floor is >10M/mo (cf. pydantic-ai ~31M, llama-index ~10.18M at level 5); the prior note bucketed it into a 5M-50M=L4 band used only in one ml_frameworks batch, below map peers of equal volume. Accelerate is the standard distributed-training launcher under the HF stack (TRL/PEFT depend on it) and clears the prevailing floor.
- https://pepy.tech/projects/accelerate recorded 2026-06-25
25,013,751 downloads in last 30 days
Capability
4 high confidenceCentral infrastructure for distributed training but scoped to orchestration, not modeling.
- https://github.com/huggingface/accelerate recorded 2026-06-22
Describes few-line distributed/mixed-precision training across devices
Unchanged since 2026-07-30 (last edited, not re-checked)