# Hugging Face — How 🤗 Accelerate runs very large models thanks to PyTorch

- Company: Hugging Face (huggingface.co)
- Announced: 2022-09-27
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://huggingface.co/blog/accelerate-large-models
- Record: https://forck.live/items/2150-how-accelerate-runs-very-large-models-thanks-to-pytorch
- Subject: Platform
- Models affected: OPT-6.7B, BLOOM, OPT-176B, OPT-13b

Hugging Face's Accelerate library leverages PyTorch's meta device to load and run very large models that do not fit in memory, using techniques like empty model creation and device mapping.

## Evidence

Verbatim from https://huggingface.co/blog/accelerate-large-models:

> We'll explain how Accelerate leverages PyTorch features to load and run inference with very large models, even if they don't fit in RAM or one GPU.

---

Record: https://forck.live/items/2150-how-accelerate-runs-very-large-models-thanks-to-pytorch
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md