---
name: Fine-tuning
slug: fine-tuning
organization_slug: hugging-face
category: ai
source_count: 2
canonical: https://releases.sh/hugging-face/fine-tuning
---

# Fine-tuning

Libraries for efficient model fine-tuning and alignment

## Sources (2)

- [PEFT](https://releases.sh/hugging-face/peft) — `github`
- [TRL](https://releases.sh/hugging-face/trl) — `github`

## Recent Releases

_Summaries below — fetch `url` for the original source, or `canonical` for the releases.sh copy of the full content._

<Release source="trl" version="v1.17.1" date="October 30, 2026" published="2026-10-30T14:15:28.000Z" url="https://github.com/huggingface/trl/releases/tag/v1.17.1" canonical="https://releases.sh/release/rel_GDI4J5YyX8kJnXdY8-y6k" truncated="true">
Fixed LFM2.5 tool-call parsing so a parenthesis inside an argument no longer closes the call prematurely.
</Release>

<Release source="trl" version="v1.17.0" date="October 29, 2026" published="2026-10-29T15:51:15.000Z" url="https://github.com/huggingface/trl/releases/tag/v1.17.0" canonical="https://releases.sh/release/rel_-8FUEc-qSLV5FkBKAOdLq" truncated="true">
Adds a LoRA SFT example for mistralai/Mistral-Large-4-1T-A52B that fine-tunes the 1T-parameter model on a single 8×B300 node on top of its FP8 weights, requiring transformers>=5.22.0, peft>=0.21.0 and accelerate>=1.15.0. GRPO and AsyncGRPO now renormalize log-probs over the top-k/top-p candidate set, which raises an error when top_p<1.0 or min_p>0.0 is set without top_k>0, and score centering is now the default AsyncGRPO objective.
</Release>

<Release source="trl" version="v1.16.0" date="October 15, 2026" published="2026-10-15T18:00:33.000Z" url="https://github.com/huggingface/trl/releases/tag/v1.16.0" canonical="https://releases.sh/release/rel_sYn3OMOJV8sislSKeJ_20" truncated="true">
TRL now trains on AWS Trainium with a new guide and pad_to_multiple_of defaulting to "auto" (128 on Neuron, None elsewhere), and AsyncGRPO packs micro-batch rows as a prefix tree, forwarding shared-prompt tokens once (up to 3.1x faster with peak memory dropping from 35.2 to 13.5 GiB). Adds Muse Spark Open support, AMD GPUs in async trainers, and JSON-schema-typed tool-call arguments.
</Release>

<Release source="trl" version="v1.15.0" date="October 8, 2026" published="2026-10-08T19:26:48.000Z" url="https://github.com/huggingface/trl/releases/tag/v1.15.0" canonical="https://releases.sh/release/rel_-QlXNUAbicPpBnhrHxbmg" truncated="true">
A new fused LM head, on by default in SFT, DPO, KTO, GRPO, RLOO and Distillation, projects hidden states through the LM head in Triton-kernel tiles to avoid building the full logits tensor, raising max trainable sequence length up to 6.9x and cutting peak memory 52% to 82%. The full-logits scoring path, use_liger_kernel chunked path and _forward_redirection are removed, so a PEFT adapter on lm_head now raises and use_liger_kernel=True is deprecated for v2.0.0 removal.
</Release>

<Release source="trl" version="v1.14.2" date="October 6, 2026" published="2026-10-06T20:30:35.000Z" url="https://github.com/huggingface/trl/releases/tag/v1.14.2" canonical="https://releases.sh/release/rel_H7vNfWdQk6TovPDaKnm17" truncated="true">
GRPO, RLOO, and Distillation on the default non-vLLM generation path now stop completions at every eos id the model declares, fixing silent training on tokens generated past the end of the turn that showed up as completions/clipped_ratio near 1. Also fixes DFT loss on batches with no trainable tokens, precompute_ref_log_probs under FSDP with no reference model, completion offsets where the tokenized prompt and prompt+completion diverge, and guards the requests and urllib3 imports in the vLLM client.
</Release>

<Release source="peft" version="v0.21.2" date="October 1, 2026" published="2026-10-01T10:29:50.000Z" url="https://github.com/huggingface/peft/releases/tag/v0.21.2" canonical="https://releases.sh/release/rel_BVxRC-qWaAqKKUl9adg9r" truncated="true">
Fixes an issue that prevented encoder-decoder models from working when using Transformers 5.18.0 or later.
</Release>

<Release source="trl" version="v1.14.1" date="September 29, 2026" published="2026-09-29T13:41:37.000Z" url="https://github.com/huggingface/trl/releases/tag/v1.14.1" canonical="https://releases.sh/release/rel_vPHZ9rpgU-8rpBiQcc-Kj" truncated="true">
Fixed a server-mode crash at the first weight sync on vLLM 0.20 through 0.25, and vLLM server mode now generates one completion per sample after tool calls. CLI scripts again honor resume_from_checkpoint and preserve explicit model_init_kwargs.
</Release>

<Release source="peft" version="v0.21.1" date="September 29, 2026" published="2026-09-29T09:28:26.000Z" url="https://github.com/huggingface/peft/releases/tag/v0.21.1" canonical="https://releases.sh/release/rel_LAKJasTgBfDpiQOVEd5tZ" truncated="true">
PEFT now works properly with Tensor Parallel (TP), and requires Transformers 5.17.0 or later.
</Release>

<Release source="trl" version="v1.14.0" date="September 25, 2026" published="2026-09-25T06:40:07.000Z" url="https://github.com/huggingface/trl/releases/tag/v1.14.0" canonical="https://releases.sh/release/rel_oGhz1Hr_wFgUSzwDM751v" truncated="true">
The trl.losses module introduced in v1.13 is removed, with DPO, KTO, and GRPO now streaming their own per-token log-probs through _ChunkedLogProbFunction, which also unlocks DPO with precomputed reference log-probs that did not exist before. A fused Triton logprob and entropy kernel now ships in-tree at trl.kernels and is on by default on CUDA, ROCm, and XPU, and six experimental trainers were removed from trl.experimental.
</Release>

<Release source="peft" version="v0.21.0" date="September 15, 2026" published="2026-09-15T15:31:52.000Z" url="https://github.com/huggingface/peft/releases/tag/v0.21.0" canonical="https://releases.sh/release/rel_RkH5ThUioDq7NBUVmU8of" truncated="true">
Four new fine-tuning methods ship: a Riemannian-preconditioned LoRA optimizer, the KaSA LoRA variant, Super-Tuning sparsity allocation, and ShadowPEFT. Also fixed a LoRA scaling applied twice in the SVD path of add_weighted_adapter, sped up MoE target parameter computation, and reduced Orthogonal Subspace Fine-tuning memory 22% and training time 46%.
</Release>

**Tags:** `fine-tuning`, `python`, `rlhf`
