TorchAO

license

BSD-3-Clause(OSI)

source

public

core features withheld

no

The LICENSE file contains the BSD-3-Clause text. The repository is public and the README describes no paid tier, enterprise edition or license-gated build, so the core is ungated. The default install and the float8, quantization-aware training, int4 and int8 paths need no non-OSI package; the proprietary nvidia-cutlass-dsl is an optional dependency for the MoE mxfp8 training kernels only.

• https://docs.nvidia.com/cutlass/latest/media/docs/pythonDSL/license.html recorded 2026-09-30

  The NVIDIA CUTLASS DSL page is a "Software License Agreement" that "governs the use of the NVIDIA CUTLASS DSLs software and materials that NVIDIA delivers to you under this Agreement".

• https://raw.githubusercontent.com/meta-pytorch/mslk/main/LICENSE recorded 2026-09-30

  MSLK LICENSE opens "BSD License" and "For MSLK software" and carries "Copyright (c) Meta Platforms, Inc. and affiliates. All rights reserved."

• https://raw.githubusercontent.com/pytorch/ao/main/LICENSE recorded 2026-09-30

  LICENSE body is the verbatim three-clause BSD text under Meta and Arm copyright lines, with "Redistribution and use in source and binary forms, with or without modification, are permitted" and "Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products".

• https://raw.githubusercontent.com/pytorch/ao/main/README.md recorded 2026-09-30

  README carries the badge "license-BSD_3--Clause" pointing at ./LICENSE and lists MSLK as "is an optional runtime dependency that provides accelerated kernels for some of the workflows in torchao" while the default install is plain "pip install torchao". Searched the body for enterprise, license key, paid, commercial and proprietary: none.

• https://raw.githubusercontent.com/pytorch/ao/main/setup.py recorded 2026-09-30

  setup.py builds the package from the repository, with the comment "Auto-enable experimental builds on ARM64 macOS when USE_CPP=1" and env flags such as USE_CPP. Searched the body for enterprise, license key, paid and proprietary: none.

PyPI downloads of `torchao` are the adoption signal here.

• https://pypistats.org/api/packages/torchao/recent recorded 2026-08-18

  last_month downloads = 4,080,208 for torchao

TorchAO applies float8, MX-format and quantization-aware training on the training side and int4 and int8 quantization on the serving side, all as native PyTorch dtypes and tensor subclasses. That is depth across techniques within one framework, short of the multi-framework, multi-vendor compilation Apache TVM covers.

• https://docs.pytorch.org/ao/main/workflows/index.html recorded 2026-09-30

  Workflows docs page (torchao main documentation) states "our main training workflow is float8 quantized training" and lists inference workflows.

• https://raw.githubusercontent.com/pytorch/ao/main/README.md recorded 2026-09-30

  README describes float8 training with "from torchao.float8 import convert_to_float8_training", quantization-aware training with "from torchao.quantization.qat import QATConfig", and "any custom dtype or memory layout should work with our compiler".