torchao

TorchAO is an architecture optimization library for PyTorch, it provides high performance dtypes, optimization techniques and kernels for inference and training, featuring composability with native PyTorch features like torch.compile, FSDP etc.. Some benchmark numbers can be found here.

Before you begin, make sure you have Pytorch version 2.5, or above, and TorchAO installed:

pip install -U torch torchao

Usage

Now you can quantize a model by passing a TorchAoConfig to from_pretrained(). This works for any model in any modality, as long as it supports loading with Accelerate and contains torch.nn.Linear layers.

Usage

Resources

< > Update on GitHub