huggingface/accelerate

Health v2 · maintenance only

Adoption not part of the score

9838 stars · 1444 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

Hugging Face Accelerate is a Python library that lets you run the same PyTorch training and inference code on any device or distributed configuration (multi-GPU, TPU, fp16/fp8 mixed precision, FSDP, DeepSpeed) by adding just a few lines of code. It abstracts the boilerplate of distributed training while leaving your training loop unchanged.

Use cases

• train pytorch models on multiple gpus
• run the same training script on gpu, tpu, or cpu
• add mixed precision fp16 or fp8 to my training loop
• scale pytorch training with fsdp or deepspeed
• run inference with very large models that don't fit in memory
• launch distributed training without writing torchrun boilerplate
• fine-tune large language models across distributed hardware

When to choose

• you write custom PyTorch training loops and want minimal changes to scale them
• you need portability across GPUs, TPUs, and distributed setups with one codebase
• you want easy mixed precision, FSDP, or DeepSpeed integration without learning each framework's API
• you're in the Hugging Face ecosystem (Transformers, Diffusers, TRL)

When to avoid

• you use a high-level training framework like PyTorch Lightning or Keras and don't need fine-grained control
• you need fully managed distributed training rather than a library you integrate into your own loop
• your project isn't PyTorch-based (e.g., JAX or TensorFlow)

Facets

library · maturity stable

machine-learning llm-training gpu-computing cli deep-learning machine-learning large-language-models developer-tools python cross-platform cli pytorch distributed-training mixed-precision fsdp deepspeed multi-gpu tpu huggingface gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="huggingface/accelerate")