torchtune

license

BSD-3-Clause(OSI)

source

public(pytorch/torchtune)

governance

PyTorch Foundation/Meta

core-gated

ungated

BSD-3-Clause, an OSI license, and the text is clean, unmodified three-clause BSD with no field-of-use or commercial carve-out -- worth reading directly, since this license is off the category's Apache-2.0 norm. The full source is public at the project's current home, meta-pytorch/torchtune, to which the old pytorch/torchtune address redirects. There is no enterprise or pricing tier, and openness is unaffected by the project being wound down.

• https://github.com/pytorch/torchtune recorded 2026-08-09

  Redirects (final_url https://github.com/meta-pytorch/torchtune) to the repo now hosted under the meta-pytorch GitHub org. Metadata shows "isArchived":false and "visibilityLabel":"Public", license spdxId "BSD-3-Clause", 5,794 stargazers and 744 forks. The README opens with "Torchtune is no longer actively maintained: torchtune development wound down in 2025" and lists "Hackable training recipes for SFT, knowledge distillation, DPO, PPO, GRPO, and quantization-aware training", "PyTorch FSDP2" for distributed training, single-device recipes, and a February 2025 note that multi-node training is "officially open for business in torchtune". The repository is an ordinary public source tree (recipes, docs, tests, configs) with no enterprise, pricing, or paid-tier directory; the page's own 'enterprise'/'pricing' strings are GitHub's own site navigation, not anything torchtune-specific.

• https://raw.githubusercontent.com/meta-pytorch/torchtune/main/LICENSE recorded 2026-08-09

  The full raw LICENSE file is a standard three-clause BSD license: "BSD 3-Clause License", copyright "Copyright 2024 Meta", the three ordinary conditions (retain the copyright notice in source; reproduce it in binary distributions; the copyright holder's or contributors' names may not be used to endorse or promote derived products without permission), and the standard AS-IS warranty disclaimer. No added field-of-use restriction, commercial carve-out, or non-standard clause beyond the three.

315,525 PyPI downloads in the trailing month, solidly inside the 100K-1M band. The project is officially deprecated and development has wound down, so usage is on a declining trajectory, and its ~5.8K stars are roughly flat. The level rests on current download volume, with that decline as the standing caveat.

• https://pypistats.org/api/packages/torchtune/recent recorded 2026-08-09

  Whole JSON body: {"data":{"last_day":10553,"last_month":315525,"last_week":65273},"package":"torchtune","type":"recent_downloads"} - 315,525 downloads in the trailing month, which stays inside the 100K-1M band the adoption score already rests on.

• https://pypi.org/project/torchtune/ recorded 2026-08-09

  Project page titled "torchtune · PyPI". Sidebar lists Author "PyTorch Team" (packages@pytorch.org), License "BSD 3-Clause License", and "Released:" with datetime 2025-04-07 - confirming the download figure belongs to the product's own package and that its most recent release predates the 2025 wind-down the GitHub README describes.

• https://github.com/pytorch/torchtune recorded 2026-08-09

  The repo (now at meta-pytorch/torchtune) records 5,794 stargazers and 744 forks, roughly flat against the ~5.8K stars previously recorded, corroborating a plateaued rather than growing user base consistent with the deprecated status.

The feature set -- SFT, knowledge distillation, DPO, PPO and GRPO, quantization-aware training, FSDP2, scaling from a single device through multi-device to multi-node, and integrations with the Hugging Face Hub and Weights & Biases -- has not changed since development wound down, and no newer method or precision mode such as FP8 or FP4 has been added. Memory efficiency remains its focus. That leaves it competent but behind actively developed peers and well behind Megatron-LM, which is what the middle band reflects.

• https://github.com/pytorch/torchtune recorded 2026-08-09

  The README's feature list still matches the recorded value: "Hackable training recipes for SFT, knowledge distillation, DPO, PPO, GRPO, and quantization-aware training", "PyTorch FSDP2" for distributed training, single-device recipes, and a February 2025 note that multi-node training is "officially open for business in torchtune"; Hugging Face Hub and Weights & Biases integrations are both referenced. No new method, precision mode, or capability has been added since development wound down in 2025.