Sourced from datasets's releases.

5.0.0

Datasets Features

Agent traces

• Parse Agent traces messages for SFT using teich by @​lhoestq in huggingface/datasets#8232

  • Agent traces from claude_code/pi/codex and others can now be loaded with load_dataset
  • Using the teich library (new optional dependency), traces are parsed to messages to enable training on traces using e.g. trl
  • Load the data:

  >>> from datasets import load_dataset
  >>> ds = load_dataset("lhoestq/agent-traces-example", split="train")
  >>> ds[0]["messages"]
  [{'role': 'user', 'content': 'Download a random dataset from Hugging Face, use DuckDB to inspect it, and come back with a short report about it. Be concise and include: dataset name, what files/format you found, row count or rough size if you can determine it,...'
   ...]

  • Train on agent traces:

  trl sft --dataset-name lhoestq/agent-traces-example ...

  • find all the Agent traces datasets on HF here: https://huggingface.co/datasets?format=format:agent-traces&sort=trending

Next-level shuffling in streaming mode

• Use multiple input shards for shuffle buffer by @​lhoestq in huggingface/datasets#8194

  ds = load_dataset(..., streaming=True)
  ds = ds.shuffle(seed=42)
  # or configure local buffer shuffling manually, default is:
  ds = ds.shuffle(seed=42, buffer_size=1000, max_buffer_input_shards=10)

  before👎:

  after✨:

  toy example comparison

  from datasets import IterableDataset
  ds = IterableDataset.from_dict({"i": range(123_456_789)}, num_shards=1024)
  ds = ds.shuffle(seed=42)
  print("Cold start ids:")

... (truncated)

• 68ac1a9 Release: 5.0.0 (#8239)
• cfe4492 Support composed splits in streaming datasets (#8220)
• fd67320 Keep None as a real null in Json() columns instead of the string "null" (#8231)
• 10cdc81 Fix iterable skip over full Arrow blocks (#8236)
• b7c064d Parse agent traces messages for SFT using teich (#8232)
• 31e92f1 fix: embed_external_files=True for mesh support (#8224)
• d168d5f feat: add TsFile (Apache IoTDB) packaged builder with per-device wide format ...
• 992f3cf fix(map): fix progress bar exceeding total when load_from_cache_file=False (#...
• 8474a91 Fix single lance file form pylance 7.0 (#8225)
• d4284e9 feat: add 3D mesh support and MeshFolder builder (#8055)
• Additional commits viewable in compare view

Sourced from transformers's releases.

Patch release v5.12.1

Updated the lower bound for PEFT and a fix for auto tokenizer to properly resolve the mistral tokenizer (when mistral-common is installed). This is similar to v.5.10.3 minus the fixes that were already included in the main release - vLLM will first target 5.10.3 🤗

• Fix peft lower bound #46605 by @​hmellor (#46605)
• mistral common backend fix #46667 by @​itazap (#46667)

Full Changelog: huggingface/transformers@v5.12.0...v5.12.1

• ddb849a v5.12.1
• d81db16 mistral common backend fix (#46667)
• 07c294c Fix peft lower bound (#46605)
• e0e7504 Revert "fix security issue with allow all kernels"
• 7904f30 fix security issue with allow all kernels
• d77d573 fix kernel path
• f7999c7 v5.12.0
• 4c5d4fd Add minimax m3vl (#46600)
• 5957e6f [CI] capture checkers output in OTEL (#46601)
• 2d68208 Lfm2: thread seq_idx through ShortConv for packed/varlen inputs (#46588)
• Additional commits viewable in compare view

Sourced from trl's releases.

v1.6.0

Features

AsyncGRPO rollout worker now runs in a separate process

AsyncRolloutWorker is no longer a thread — it's a spawned child process with its own GIL. The trainer's autograd engine no longer competes with recursive_parse / accuracy_reward for the GIL, which was causing 1-5s stalls in real Qwen3-30B-A3B @ 16k runs and ultimately NCCL watchdog timeouts on other ranks.

Architectural changes:

• AsyncRolloutWorker (parent) owns the child process + shared mp.Queue / mp.Value / mp.Event.
• _AsyncRolloutLoop (child-only) handles tokenization, dataset iteration, reward funcs, and asyncio loops.
• A new WeightTransferClient owns the NCCL group with vLLM (/pause, /resume, /init_weight_transfer_engine, /update_weights); the rollout child only talks to /v1/completions.

Two correctness fixes shipped alongside (they would have conflicted otherwise): broader aiohttp retry (now catches ClientPayloadError) with bounded exponential backoff, and all-NaN reward columns are now preserved — np.nansum was silently returning 0, giving unscorable completions a real advantage signal and pushing the policy away from correct answers (~30% of DeepMath / OpenR1-Math rows).

[!NOTE] reward_funcs / tools / environment_factory must now be picklable, and the child runs CPU-only (CUDA_VISIBLE_DEVICES="").

by @​AmineDiro in huggingface/trl#5749

New experimental A2PO trainer (Optimal Advantage Regression)

A new A2POTrainer implements A*-PO from "Accelerating RL for LLM Reasoning with Optimal Advantage Regression". Two stages: an offline V* estimation pass from reference policy samples (with optional filter_all_incorrect to drop prompts where every reference completion fails), then on-policy training with one generation per prompt and a plain least-squares loss on β₂·log(π/π_ref) vs r − V*. No group, no critic, no clipping, no reward normalization.

from trl.experimental.a2po import A2POConfig, A2POTrainer
trainer = A2POTrainer(
model="Qwen/Qwen3-4B",
args=A2POConfig(num_value_samples=8, filter_all_incorrect=True),
train_dataset=dataset,
reward_funcs=accuracy_reward,
)
trainer.train()

Designed for binary verifiable rewards (math/code), not open-ended problems.

by @​raghulchandramouli in huggingface/trl#5940

KTO now supports VLMs + big alignment push

The biggest KTO ↔ DPO alignment cycle yet — KTOTrainer now supports vision-language models, plus a deep restructuring of compute_loss, KL dataset generation, ref-logp precomputation, activation offloading, sampler strategy, metrics, and more. KTO graduation is very close.

from trl.experimental.kto import KTOConfig, KTOTrainer
trainer = KTOTrainer(
model="Qwen/Qwen2.5-VL-3B-Instruct",
args=KTOConfig(...),
train_dataset=vision_kto_dataset,
</tr></table>

... (truncated)

• 0dac440 Release: v1.6 (#6009)
• 6842058 docs: clarify PPO entropy metrics in PPO trainer docs (#5289)
• cb5ca23 fix(cli): drop duplicate "to" in trl skills install description (#6008)
• 8226159 Hide DeepSpeed/FSDP distributed backend boilerplate (#6000)
• fa286a8 Padding-free invariance test (#5842)
• eab8bc8 Announce upcoming SFT loss_type default change from 'nll' to `'chunked_nl...
• 4520e4b [CI] Check that training chat templates keep the stop token in the loss mask ...
• e28c6d9 Document bnb_4bit_quant_storage and normalize docstring param headers (#5993)
• 3f6f7d2 Align KTO with DPO: Inline kto_loss in _compute_loss (#5999)
• b84b487 Align KTO with DPO: Rename kto_loss_fn to liger_loss_fn (#5998)
• Additional commits viewable in compare view

Sourced from bitsandbytes's releases.

0.49.2

Highlights

• The default blocksize of 64 for 4bit quantization is now supported on ROCm. Previously the default was 128, which was a mismatch from the default for other devices.
• ROCm 7.2 build is now included.

What's Changed

• bug: fix 8bitoptim support with fsdp by @​ved1beta in bitsandbytes-foundation/bitsandbytes#1840
• Fix xpu 4bit kernel by @​jiqing-feng in bitsandbytes-foundation/bitsandbytes#1839
• ROCm 7.2 build and doc changes by @​sstamenk in bitsandbytes-foundation/bitsandbytes#1845
• Add CUDA kernel support for 4-bit quantization with blocksize=32 by @​Abdennacer-Badaoui in bitsandbytes-foundation/bitsandbytes#1854
• Add blocksize=64 4-bit quantization support for ROCm CDNA (warp64) GPUs by @​Abdennacer-Badaoui in bitsandbytes-foundation/bitsandbytes#1856
• [Docs Update] QLoRA 4-bit Support on ROCm by @​Abdennacer-Badaoui in bitsandbytes-foundation/bitsandbytes#1857
• [ROCm] Make blocksize=64 default for 4bit by @​matthewdouglas in bitsandbytes-foundation/bitsandbytes#1873
• Handle non-contiguous tensors in quantize/dequantize ops by @​TimDettmers in bitsandbytes-foundation/bitsandbytes#1859
• Fix AdEMAMix scheduler guard and add state_dict round-trip test by @​TimDettmers in bitsandbytes-foundation/bitsandbytes#1861

New Contributors

• @​Abdennacer-Badaoui made their first contribution in bitsandbytes-foundation/bitsandbytes#1854

Full Changelog: bitsandbytes-foundation/bitsandbytes@0.49.1...0.49.2

0.49.1

What's Changed

• Update AMD targets by @​sstamenk in bitsandbytes-foundation/bitsandbytes#1832
• add guard for self.weight.quant_state attribute by @​winglian in bitsandbytes-foundation/bitsandbytes#1837

Full Changelog: bitsandbytes-foundation/bitsandbytes@0.49.0...0.49.1

0.49.0

Highlights

x86-64 CPU Improvements

CPU performance for 4bit is significantly improved on x86-64, with optimized kernel paths for CPUs that have AVX512 or AVX512BF16 support.

AMD ROCm Experimental Wheels

• Experimental support for AMD devices is now included in our PyPI wheels on Linux x86-64.
• We've added additional GPU target devices as outlined in our docs.
• Support for using the default blocksize of 64 for 4bit was added for RDNA GPUs in #1748.

macOS 14+ Wheels

• We're now publishing wheels for macOS 14+!
• The 4bit and 8bit quantization features are supported on MPS by slow implementations. We plan to enable Metal kernels with improved performance in the future.

🚨 Breaking Changes

• Dropped support for Python 3.9.
• Dropped compilation support for Maxwell GPUs in the CUDA backend.

What's Changed

• [ROCm] Update build targets by @​matthewdouglas in bitsandbytes-foundation/bitsandbytes#1788

... (truncated)

• f0e6ca3 Release 0.49.2
• cace65c Fix AdEMAMix scheduler guard and add state_dict round-trip test (#1861)
• 505a00a Handle non-contiguous tensors in quantize/dequantize ops (#1859)
• c2ae381 [ROCm] Make blocksize=64 default for 4bit (#1873)
• 943e42d Skip typo check on agent markdown files
• b075afc update docs (#1857)
• de89ff7 docs: Add PR review posting guide, agent reference docs, and lint fixes
• a2c92f7 docs: Add human+agent issue triage workflow guide
• 252e6ff docs: Update issue patterns from 43 recently closed issues
• 88c6c71 style: Fix ruff format violation in test_linear4bit.py
• Additional commits viewable in compare view

Sourced from setuptools's changelog.

v82.0.1

Bugfixes

• Fix the loading of launcher manifest.xml file. (#5047)
• Replaced deprecated json.__version__ with fixture in tests. (#5186)

Improved Documentation

• Add advice about how to improve predictability when installing sdists. (#5168)

Misc

• #4941, #5157, #5169, #5175

v82.0.0

Deprecations and Removals

• pkg_resources has been removed from Setuptools. Most common uses of pkg_resources have been superseded by the importlib.resources <https://docs.python.org/3/library/importlib.resources.html>_ and importlib.metadata <https://docs.python.org/3/library/importlib.metadata.html>_ projects. Projects and environments relying on pkg_resources for namespace packages or other behavior should depend on older versions of setuptools. (#3085)

v81.0.0

Deprecations and Removals

• Removed support for the --dry-run parameter to setup.py. This one feature by its nature threads through lots of core and ancillary functionality, adding complexity and friction. Removal of this parameter will help decouple the compiler functionality from distutils and thus the eventual full integration of distutils. These changes do affect some class and function signatures, so any derivative functionality may require some compatibility shims to support their expected interface. Please report any issues to the Setuptools project for investigation. (#4872)

v80.10.2

Bugfixes

• Update vendored dependencies. (#5159)

Misc

... (truncated)

• 5a13876 Bump version: 82.0.0 → 82.0.1
• 51ab8f1 Avoid using (deprecated) 'json.version' in tests (#5194)
• f9c37b2 Docs/CI: Fix intersphinx references (#5195)
• 8173db2 Docs: Fix intersphinx references
• 09bafbc Fix past tense on newsfragment
• 461ea56 Add news fragment
• c4ffe53 Avoid using (deprecated) 'json.version' in tests
• 749258b Cleanup pkg_resources dependencies and configuration (#5175)
• 2019c16 Parse ext-module.define-macros from pyproject.toml as list of tuples (#5169)
• b809c86 Sync setuptools schema with validate-pyproject (#5157)
• Additional commits viewable in compare view

Sourced from setuptools-scm's releases.

setuptools-scm v10.1.1

Fixed

• Update CI to use PyPy 3.11 as cryptography has no PyPy 3.10 build available (#1421)

setuptools-scm v10.1.0

Added

• Add backward-compatible shims in setuptools_scm.git, setuptools_scm.hg, setuptools_scm.hg_git, and setuptools_scm.scm_workdir so that external code calling get_scm_version(config) or run_describe(config) with an explicit Configuration continues to work. The shim automatically wires _config and VcsEnvironment onto the workdir. (#compat-shims)
• Write scm_version.json and scm_file_list.json into egg-info directories during egg_info, enabling sdist fallback version inference when no VCS is present. Add ScmEggInfoMixin for workdir-based file finding in find_sources(). (#egg-info-metadata)
• Add write_to_source pyproject.toml option to control whether version files are written to the source tree. When unset, a deprecation warning advises setting it explicitly before the default changes in a future major release. The SETUPTOOLS_SCM_WRITE_TO_SOURCE environment variable overrides this setting. (#1301)
• Adopt the workdir-centric pipeline from vcs-versioning: version discovery now follows an explicit env → config → workdir → version chain instead of relying on ambient globals and parse entry points. The egg_info command writes scm_version.json and scm_file_list.json metadata so sdists can infer versions without a VCS checkout. Requires vcs-versioning >= 2.0.0.dev0. (#1378)

Fixed

• Fix worktree file listing test to expect relative paths from the file finder. The test now passes on Linux; Windows remains xfail due to a subprocess limitation with worktree directories. (#620)
• Remove the _warn_on_old_setuptools() check that incorrectly warned when a custom build-backend caused setuptools.__version__ to return the project version instead of setuptools' version. The minimum setuptools version is now enforced via build-system requirements. (#1192)
• Wrap version in setuptools.sic() when normalize = false to prevent setuptools from re-normalizing the version after our hook returns. This preserves CalVer zero-padding (e.g. 2024.01.05) and other non-canonical version strings in dist.metadata.version. (#1354)
• Skip writing non-package version files to build_lib, fixing incorrect inclusion of root-level version files in wheels. (#1364)

Documentation

• Rewrite the GitHub Actions CI/CD example to use a dedicated build job (via build-and-inspect-python-package) and OIDC Trusted Publishers instead of building in publishing jobs with long-lived API tokens. (#1215)

setuptools-scm v10.0.5

Fixed

• Allow dump_version() deprecation warning to be silenced by passing scm_version=None. (#1286)
• Remove [tool.uv.sources] from setuptools-scm/pyproject.toml to fix sdist builds outside the workspace — the workspace root already declares the source mapping for development. (#1330)

setuptools-scm v10.0.4

Fixed

• Anchor get_version in setup.py with relative_to and fallback_root so SCM fallbacks (e.g. PKG-INFO) do not resolve against the wrong directory when the build cwd is the workspace or repo root. (#1302)
• Enter GlobalOverrides for SETUPTOOLS_SCM when using setuptools_scm.get_version / _get_version, avoiding implicit context warnings for direct API callers. (#1314)

Miscellaneous

• Upgrade pre-commit hooks (Ruff, mypy, codespell), align locked Ruff with hooks, and add Ruff per-file configuration for setuptools_scm re-export modules. (#1311)

setuptools-scm v10.0.3

Fixed

• Remove monorepo-only ../vcs-versioning/src from build-system.backend-path so sdists install under PEP 517 (paths must stay inside the source tree). (#1306)

... (truncated)

• b59248f Merge pull request #1342 from pypa/release/main
• ddd14aa Prepare release: setuptools-scm v10.1.1, vcs-versioning v2.0.1
• 72a4eec Merge pull request #1421 from RonnyPfannschmidt/fix/pypy-3.11
• 4bbee8a fix(ci): update PyPy 3.10 to PyPy 3.11 in test matrix
• 4b11690 Merge pull request #1345 from pypa/release/develop
• 1dbe0af Prepare release: setuptools-scm v10.1.0, vcs-versioning v2.0.0
• d4cbbf3 Merge pull request #1419 from RonnyPfannschmidt/fix/setup-py-dist-name
• 1084ff3 fix: add dist_name to setup.py get_version() calls
• 16dd9ad Merge pull request #1418 from RonnyPfannschmidt/fix/release-branch-versioning
• 35242d1 fix(ci): set pretend versions on release PRs to fix wheel builds
• Additional commits viewable in compare view

You can trigger Dependabot actions by commenting on this PR:

• @dependabot rebase will rebase this PR
• @dependabot recreate will recreate this PR, overwriting any edits that have been made to it
• @dependabot show <dependency name> ignore conditions will show all of the ignore conditions of the specified dependency
• @dependabot ignore <dependency name> major version will close this group update PR and stop Dependabot creating any more for the specific dependency's major version (unless you unignore this specific dependency's major version or upgrade to it yourself)
• @dependabot ignore <dependency name> minor version will close this group update PR and stop Dependabot creating any more for the specific dependency's minor version (unless you unignore this specific dependency's minor version or upgrade to it yourself)
• @dependabot ignore <dependency name> will close this group update PR and stop Dependabot creating any more for the specific dependency (unless you unignore this specific dependency or upgrade to it yourself)
• @dependabot unignore <dependency name> will remove all of the ignore conditions of the specified dependency
• @dependabot unignore <dependency name> <ignore condition> will remove the ignore condition of the specified dependency and ignore conditions