PEFT

license

Apache-2.0(OSI)

source

public(huggingface/peft)

governance

Hugging Face

managed-tier

none

core-gated

ungated

Apache-2.0, an OSI license, with the full source public and no proprietary tier.

• https://raw.githubusercontent.com/huggingface/peft/main/LICENSE recorded 2026-08-09

  Full unmodified Apache License, Version 2.0 text, opening "Apache License / Version 2.0, January 2004 / http://www.apache.org/licenses/" and "TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION", closing with the standard APPENDIX carrying the unfilled template placeholder "Copyright [yyyy] [name of copyright owner]". No added use restriction or field-of-use clause.

• https://github.com/huggingface/peft recorded 2026-08-09

  Repository metadata reports "license":{"spdxId":"Apache-2.0","name":"Apache License 2.0"} and "isArchived":false, with description "🤗 PEFT: State-of-the-art Parameter-Efficient Fine-Tuning." The only occurrences of "enterprise" or "premium" in the page are GitHub's own site-wide navigation chrome (Copilot/Enterprise nav links to github.com/enterprise and github.com/pricing) — none reference peft itself, and the page names no gated feature or paid edition of the library.

• https://raw.githubusercontent.com/huggingface/peft/main/setup.py recorded 2026-08-09

  setup() declares name="peft", license="Apache", license_files=["LICENSE"], and classifiers including "License :: OSI Approved :: Apache Software License". extras_require is limited to quality, docs_specific, dev and test — no commercial or enterprise extra — and package_dir/packages point at the same src/peft tree the repo ships, so the installable package builds from the published source with nothing withheld.

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

  Project page headed "peft 0.20.0". Sidebar License reads "Apache Software License (Apache)" and the classifier list carries "License :: OSI Approved :: Apache Software License". Provides Extra lists exactly dev, docs-specific, quality and test, matching setup.py — no paid or enterprise extra on the distributed wheel.

• https://huggingface.co/pricing recorded 2026-08-09

  Hugging Face's paid plans: PRO at "$9" per month, Team at "$20", Enterprise at "$50", plus per-TB Hub storage and hourly Spaces/inference hardware. Every listed benefit is Hub- or account-scoped (storage, inference credits, SSO, audit logs); a case-insensitive search of the body for "peft" returns zero matches, so no PEFT feature is gated behind a plan.

11.69M PyPI downloads in the trailing 30 days, per pypistats, comfortably clear of the top band's floor of more than 10M a month. PEFT is the canonical adapter and LoRA library, and a dependency of TRL and of most LoRA fine-tuning stacks.

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

  Full JSON body: {"data":{"last_day":471851,"last_month":11691121,"last_week":2916121},"package":"peft","type":"recent_downloads"} — 11,691,121 downloads in the trailing month, comfortably inside the >10M level-5 band the adoption score already rests on.

• https://pypistats.org/packages/peft recorded 2026-08-09

  HTML package-statistics page. Rendered text reads "Downloads last month: 11,691,121" and "Downloads last week: 2,916,121", with "Latest version: 0.20.0" — corroborates the API figure from an actively-released package rather than download residue.

Reference-grade depth on parameter-efficient methods: the LoRA, QLoRA, AdaLoRA, LoHa, LoKr, IA3 and prefix/prompt-tuning family, with 4- and 8-bit quantization and BF16, and parallelism inherited from Accelerate. Its scope stops at adapters, though. It does not orchestrate full RLHF the way TRL and OpenRLHF do, and it does not train at frontier scale the way Megatron-LM does, so it sits one band below the top.

• https://raw.githubusercontent.com/huggingface/peft/main/README.md recorded 2026-08-09

  Quickstart wraps a Transformers model with `get_peft_model` and a `LoraConfig`. The PEFT integrations section documents Transformers ("directly integrated"), Diffusers ("conveniently managing different adapters"), Accelerate ("distributed training and inference for really big models"), and TRL ("PEFT can also be applied to training LLMs with RLHF components such as the ranker and policy"). The memory table reports full fine-tuning of a 12B model running out of GPU memory versus fitting in 56GB with PEFT-LoRA, and a Stable Diffusion LoRA checkpoint at "only 8.8MB".

• https://raw.githubusercontent.com/huggingface/peft/main/src/peft/utils/peft_types.py recorded 2026-08-09

  The PeftType enum lists over 40 method identifiers today — LORA, ADALORA, IA3, PROMPT_TUNING, PREFIX_TUNING, P_TUNING, BOFT, LOHA, LOKR, OFT, VERA, FOURIERFT, HRA and BONE among the established ones, plus a long tail of recent additions (DELORA, GRALORA, CARTRIDGE, WAVEFT and more). All remain adapter/PEFT-family techniques; none add full-model training orchestration or frontier-scale parallelism.

• https://raw.githubusercontent.com/huggingface/peft/main/docs/source/developer_guides/quantization.md recorded 2026-08-09

  The Quantization guide combines PEFT LoRA with 4/8-bit bitsandbytes quantization, AWQ, GPTQ and AQLM, stating "QLoRA is a method that quantizes a model to 4-bits and then trains it with LoRA" and that this "allows you to finetune a 65B parameter model on a single 48GB GPU". The worked example sets `bnb_4bit_compute_dtype=torch.bfloat16`, matching the recorded value's BF16 precision claim.