Bump the minor-update group across 1 directory with 138 updates

[dependabot]

Bumps the minor-update group with 138 updates in the / directory:

Package	From	To
regex	2026.2.28	2026.9.3
tqdm	4.67.3	4.70.0
transformers	5.15.1	5.16.1
tokenizers	0.22.2	0.23.2
torchaudio	2.10.0	2.11.0+cpu
tpu-inference	0.27.0	0.28.0
absl-py	2.1.0	2.5.0
accelerate	1.13.0	1.14.0
aiohappyeyeballs	2.6.1	2.7.1
alembic	1.16.4	1.19.2
annotated-types	0.7.0	0.8.0
anyio	4.6.2.post1	4.15.1
arctic-inference	0.2.0	0.3.0
argcomplete	3.5.1	3.7.2
arrow	1.3.0	1.4.0
audioread	3.0.1	3.1.0
azure-core	1.38.2	1.41.0
azure-storage-blob	12.28.0	12.30.1
blobfile	3.0.0	3.3.0
bm25s	0.2.13	0.3.11
boto3	1.35.57	1.43.89
botocore	1.35.57	1.43.89
cffi	2.0.0	2.1.1
charset-normalizer	3.4.0	3.5.1
click	8.1.7	8.5.0
colorlog	6.10.1	6.12.0
coverage	7.10.6	7.16.0
cramjam	2.9.0	2.11.0
cuda-bindings	13.0.3	13.3.1
cuda-pathfinder	1.3.3	1.8.1
datamodel-code-generator	0.26.3	0.76.2
dataproperty	1.0.1	1.1.1
decorator	5.1.1	5.3.1
dill	0.3.8	0.4.1
distlib	0.3.9	0.4.3
dnspython	2.7.0	2.8.0
docker	7.1.0	7.2.0
einx	0.3.0	0.4.3
email-validator	2.2.0	2.3.0
fonttools	4.55.0	4.64.0
frozenlist	1.5.0	1.8.0
genson	1.3.0	1.4.0
google-api-core	2.34.0	2.36.0
google-auth	2.40.2	2.57.1
google-cloud-core	2.4.3	2.7.0
google-cloud-storage	3.4.0	3.13.1
google-crc32c	1.7.1	1.8.0
google-resumable-media	2.7.2	2.10.2
greenlet	3.2.3	3.5.5
h11	0.14.0	0.16.0
h2	4.3.0	4.4.1
harfile	0.3.0	0.5.0
hf-xet	1.4.3	1.6.0
hiredis	3.0.0	3.4.1
hpack	4.1.0	4.2.0
httpx	0.27.2	0.28.1
huggingface-hub	1.10.2	1.30.0
humanize	4.11.0	4.16.0
hypothesis	6.131.0	6.167.1
hypothesis-graphql	0.11.1	0.13.1
idna	3.10	3.19
iniconfig	2.0.0	2.3.0
jmespath	1.0.1	1.1.0
joblib	1.4.2	1.6.0
jsonpointer	3.0.0	3.1.1
jsonschema	4.23.0	4.26.0
kiwisolver	1.4.7	1.5.1
lazy-loader	0.4	0.5
mako	1.3.10	1.4.1
matplotlib	3.9.2	3.10.9
mpmath	1.3.0	1.4.1
msal	1.34.0	1.38.0
msgpack	1.1.0	1.2.2
mteb	2.8.3	2.20.10
multidict	6.1.0	6.7.1
mypy-extensions	1.0.0	1.1.0
networkx	3.2.1	3.4.2
nltk	3.9.1	3.10.3
nvidia-cublas	13.1.0.3	13.6.1.10
nvidia-cuda-cupti	13.0.85	13.3.75
nvidia-cuda-nvrtc	13.0.88	13.3.33
nvidia-cuda-runtime	13.0.96	13.3.29
nvidia-cudnn-cu13	9.15.1.9	9.25.1.1
nvidia-cufft	12.0.0.61	12.3.0.29
nvidia-cufile	1.15.1.6	1.18.1.6
nvidia-cusolver	12.0.4.66	12.2.6.9
nvidia-cusparse	12.6.3.3	12.8.2.51
nvidia-cusparselt-cu13	0.8.0	0.9.1
nvidia-nccl-cu13	2.28.9	2.31.2
nvidia-nvjitlink	13.0.88	13.3.33
nvidia-nvshmem-cu13	3.4.5	3.7.2
nvidia-nvtx	13.0.85	13.3.29
orjson	3.11.5	3.12.0
pathvalidate	3.2.1	3.3.1
perceptron	0.1.4	0.3.5
platformdirs	4.3.6	4.11.7
pluggy	1.5.0	1.6.0
polars	1.29.0	1.44.1
pooch	1.8.2	1.9.0
propcache	0.2.0	0.5.2
pycryptodomex	3.22.0	3.23.0
pydantic-core	2.46.4	2.48.0
pydantic-extra-types	2.10.5	2.11.1
pygments	2.18.0	2.21.0
pyjwt	2.11.0	2.13.0
pyparsing	3.2.0	3.3.2
pystemmer	3.0.0	3.1.0
pytest-forked	1.6.0	1.7.5
pytest-mock	3.14.0	3.15.1
pytest-subtests	0.14.1	0.15.0
pytest-timeout	2.3.1	2.4.0
python-rapidjson	1.20	1.23
rapidfuzz	3.12.1	3.14.5
referencing	0.35.1	0.37.0
responses	0.25.3	0.26.3
s3transfer	0.10.3	0.19.2
sacrebleu	2.4.3	2.6.0
scikit-learn	1.5.2	1.7.2
scipy	1.13.1	1.15.3
six	1.16.0	1.17.0
soundfile	0.12.1	0.14.0
statsmodels	0.14.4	0.15.0
sympy	1.13.3	1.14.0
tabulate	0.9.0	0.10.0
tblib	3.1.0	3.2.2
termcolor	3.1.0	3.3.0
threadpoolctl	3.5.0	3.6.0
tomli	2.2.1	2.4.1
triton	3.6.0	3.8.0
tritonclient	2.64.0	2.71.0
typer	0.15.2	0.27.2
urllib3	2.2.3	2.7.0
uvicorn	0.35.0	0.52.4
vector-quantize-pytorch	1.21.2	1.31.4
wcwidth	0.2.13	0.8.3
wrapt	2.3.0	2.4.0
yarl	1.17.1	1.24.5
griffelib	2.2.0	2.3.0

Updates regex from 2026.2.28 to 2026.9.3

Changelog

Sourced from regex's changelog.

Version: 2026.9.3

New version of cibuildwheel no longer supports building for free-threaded Python 3.13.

Version: 2026.9.2

New version of cibuildwheel no longer needs nor supports cpython-freethreading option.

Version: 2026.9.1

Updated cibuildwheel.
Support Python 3.15.

Version: 2026.8.31

Fixed version.

Version: 2026.8.30

Git issue 611: Heap out-of-bounds write at compile time
Git issue 612: count_one() size underflow through the stale required-string cache
Git issue 613: (*SKIP) inside an atomic group, plus an equality-only scan stop
Git issue 614: build_GROUP() does not propagate the match direction

Version: 2026.8.12

Support Python 3.15.

Version: 2026.7.19

Git issue 607: SIGSEGV: NULL pointer dereference in `basic_match()` when `(?R)`/`(?0)` recursion is used inside a fuzzy-matching quantifier
Git issue 608: SIGSEGV: out-of-bounds read in `bytes1_char_at()` with `DOTALL` + a fuzzy constraint + `(?r)` reverse-direction matching

Version: 2026.7.10

Clarified that the POSIX flag makes it look for only the longest overall match.
Git issue 606: regex's compiled-pattern cache eviction is not free-threading-safe

Version: 2026.6.28

Git issue 604: regex fails on free-threading python because setlocale is not thread-safe

Version: 2026.5.9

Reverse matching with full unicode casefolding could lead to out-of-range string indexes.

Version: 2026.4.4

A fix for older Python versions before free-threading was  supported.

... (truncated)

Commits

• 57f366f New version of cibuildwheel no longer supports building for free-threaded Pyt...
• a1d72ba New version of cibuildwheel no longer needs nor supports cpython-freethreadin...
• c8d0aa4 Updated cibuildwheel. Support Python 3.15.
• b283bb2 Updated version.
• 5268ae7 Python 3.15 not released yet
• 9398a6d Git issue 614: build_GROUP() does not propagate the match direction
• b77694a Git issue 613: (*SKIP) inside an atomic group, plus an equality-only scan stop
• 8244055 Git issue 612: count_one() size underflow through the stale required-string c...
• 1c90270 Git issue 611: Heap out-of-bounds write at compile time
• 1760a20 Support Python 3.15.
• Additional commits viewable in compare view

Updates tqdm from 4.67.3 to 4.70.0

Release notes

Sourced from tqdm's releases.

tqdm v4.70.0 stable

• contrib.concurrent: major improvements
  • support process_map(mp_context, max_tasks_per_child), thread_map(thread_name_prefix) (#1265)
  • fix total based on shortest iterable length (#1473)
  • use default max_workers (#1543 <- #1530, #1518)
  • support timeout, buffersize (#1576)
  • improve ETA (#1708 <- #1161)
  • update as_completed (#1709 <- #1565)
  • add tqdm.concurrent.intepreter_map (#1777)
• asyncio: support iterables with only __aiter__ (#1714 <- #1686)
• support reset(float("inf")) (#1783 <- #1781, #651)
• framework: test & reduce wheel size (#1782)

tqdm v4.69.1 stable

• contrib.logging: preserve filters (#1775 <- #1581)
• rich: misc fixes (#1773)
  • support generators (without len) (#1391)
  • fix potential AttributeError on exit (#1674)
• fix total=float('inf') (#1781 <- #651)
• fix potential AttributeError on exit (#1778 <- #1668, #1537, #261)
• fix minor docs typos (#1780, #1746)

tqdm v4.69.0 stable

• add tqdm.asyncio.gather(..., return_exceptions=False) (#1776, #1671 <- #1286)
• misc minor framework updates
  • bump workflow actions & pre-commit hooks

tqdm v4.68.4 stable

• trim to ncols even when '{bar}' not in bar_format (#1717, #1767 <- #1658)
• fix tqdm.write when stdout=None (#1768 <- #1654)

tqdm v4.68.3 stable

• utils: delay os.get_terminal_size (#1763 <- #1760)
• autonotebook: support QtConsole, Spyder, JupyterLite (#1763, #1628, #1559 <- #1283, #1098, #512)
• minor docs updates
  • fix typo (#1762)
  • use git-fame
• misc minor framework updates
  • fix & update CI build
  • pre-commit: add docs & metadata generation
  • move tox.ini -> pyproject.toml, move tox-gh-actions -> tox-gh
  • add Python 3.14, drop 3.7 support

tqdm v4.68.2 stable

• revert accidental change to ascii default (fixes #1760)
  • UnicodeEncodeError: 'charmap' codec can't encode characters in position 6-7: character maps to <undefined> can be fixed by installing tqdm!=4.68.0,!=4.68.1
• misc docs updates
  • fix links
  • replace stray rst -> md syntax
  • consistent "progress bar" terminology (#1737)

... (truncated)

Commits

• 96f2e60 Merge pull request #1777 from shermansiu/feat/interpreter-pool
• c27393e misc tidy
• 061c623 Disable tqdm.monitor_interval for subinterpreters because they do not have ...
• 9fc160b Update how we check for interpreter_map support in the tests
• b42463a Ensure that subinterpreters can import tqdm while unpickling the initializer
• 16d5486 Add support for nested progress bars
• 9f5890f Add initial implementation for interpreter_map
• 321f920 Merge pull request #1783 from LuShadowX/reset-inf-total
• 4664b57 minor tidy
• 426a098 Treat inf total as unknown in reset() too
• Additional commits viewable in compare view

Updates transformers from 5.15.1 to 5.16.1

Release notes

Sourced from transformers's releases.

Release v5.16.1

This is a special release as we include GLM! (and a few small fixes)

GLM-5.3-Flash

GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series. With 320B total parameters and just 18B active parameters, it outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks.

GLM-5.3-Flash starts from a newly trained base model, with its architecture and training recipe redesigned around capability and efficiency. For the first time in the GLM series, we introduce a hybrid architecture combining sparse and linear attention, sharply reducing long-context serving costs while preserving precise long-context capabilities. The model also adopts Manifold-Constrained Hyper-Connections (mHC) to further improve scaling efficiency. Together with our latest 30T-token multimodal pre-training corpus, these changes enable GLM-5.3-Flash to deliver more intelligence with less compute.

Links: Documentation

• [Glm 5.3 Flash] GLM 5.3 Flash Support (#48342) by @​Dovis01 in #48342

Small patch fixes

Mainly BC behavior for TP and pinning a hf kernel for security reasons 🤗

• Restore BC for the tensor-parallel API (#48300) by @​ArthurZucker
• Fix kernel commit and repo paths for ESMFold2 (#48186) by @​Rocketknight1

Full Changelog: huggingface/transformers@v5.16.0...v5.16.1

Release: v5.16.0

Release v5.16.0

New Model additions

Qwen4-Exp

Qwen4-Exp builds on Qwen3.5's hybrid text and multimodal architecture with three key components: GatedResidual (GR), Qwen Sparse Attention (QSA), and Per-Layer Embedding (PLE).

GR is a Qwen-developed residual architecture that combines Hyper-Connection with GatedNorm. It mixes multiple residual streams with fine-grained elementwise gating before each attention and Mixture-of-Experts (MoE) block, then controls how much of the block output is injected back into each stream.

QSA uses multiple query heads to score compressed key blocks, selects the most relevant contiguous token blocks, and keeps the incomplete trailing block uncompressed. This block-level selection reduces indexing overhead and improves memory locality for long sequences. Combined with Gated DeltaNet, QSA makes Qwen4-Exp the first hybrid architecture to integrate linear and sparse attention, substantially improving inference efficiency for long-context workloads.

PLE enriches selected decoder layers with layer-specific lexical features derived from hashed token n-grams and a dilated depthwise convolution.

Links: Documentation

• Add Qwen4Exp model (#48337) by @​Cyrilvallez in #48337

GraniteSpeech5

... (truncated)

Commits

• 93c8b7b v5.16.1
• cdaca83 [Glm 5.3 Flash] GLM 5.3 Flash Support (#48342)
• b366ad3 Fix kernel commit and repo paths for ESMFold2 (#48186)
• f2f4327 Restore BC for the tensor-parallel API (#48300)
• 93d1bcf release: v5.16.0
• 276c5c4 Add Qwen4Exp model (#48337)
• d3cdaec Fix video-llama modular conversion (#48336)
• 949e301 CI: gate the hunyuan-moe slow test (#48330)
• 5fcf605 Add a regression test for force_accelerate_hooks signature preservation (#48...
• 27e7f6c Add an opt-in per-frame pixel cap (cap_pixels_per_frame) to the Qwen3-VL vide...
• Additional commits viewable in compare view

Updates tokenizers from 0.22.2 to 0.23.2

Release notes

Sourced from tokenizers's releases.

v0.23.2

This is the last v0 release, we are moving to v1!!

More details coming soon 👀

What's Changed

• Update daachorse from 1.0.1 to 3.0.0 by @​musicinmybrain in huggingface/tokenizers#2050
• Ci test fixes by @​sebpop in huggingface/tokenizers#2079
• Avoid full vocab clone in get_vocab_size() by @​eunseo9311 in huggingface/tokenizers#2074
• ci: pin benchmark jobs to a larger dedicated runner by @​ArthurZucker in huggingface/tokenizers#2084
• ci: target benchmark runners by group, not nonexistent label by @​ArthurZucker in huggingface/tokenizers#2090
• Bindings: acquire model read-lock once per call instead of per pre-token by @​sebpop in huggingface/tokenizers#2072
• ci: split audit in a separate job by @​SBrandeis in huggingface/tokenizers#2103
• ci: install libcairo on benchmark runners by @​SBrandeis in huggingface/tokenizers#2100
• feat(ci): enable npm's trusted publisher for releases by @​McPatate in huggingface/tokenizers#2111
• chore: bump pyo3 to version 0.29 by @​SBrandeis in huggingface/tokenizers#2115
• perf: add benchmarks for the decode API by @​SBrandeis in huggingface/tokenizers#2099
• ci: cache HF test data and authenticate Hub downloads to avoid rate limits by @​SBrandeis in huggingface/tokenizers#2102
• Implementing Parity-aware BPE by @​cimeister in huggingface/tokenizers#1974
• Make doc builds faster by @​mishig25 in huggingface/tokenizers#2158
• chore(deps-dev): bump websocket-driver from 0.7.4 to 0.7.5 in /tokenizers/examples/unstable_wasm/www by @​dependabot[bot] in huggingface/tokenizers#2208
• chore(deps-dev): bump shell-quote from 1.8.3 to 1.10.0 in /tokenizers/examples/unstable_wasm/www by @​dependabot[bot] in huggingface/tokenizers#2244
• chore: remove stale examples (kills 50% of dependabot traffic) by @​ArthurZucker in huggingface/tokenizers#2285
• chore(node): drop 8 unused devDependencies by @​ArthurZucker in huggingface/tokenizers#2286
• ci: pin every action to one SHA, drop stale audit suppressions by @​ArthurZucker in huggingface/tokenizers#2291
• chore(node): take node security alerts from 33 to 0 by @​ArthurZucker in huggingface/tokenizers#2287
• fix(python): pin the ruff rule set so a ruff release can't redden main by @​ArthurZucker in huggingface/tokenizers#2292
• Add a ParityBpeTrainer example and list it in the trainers API docs by @​cimeister in huggingface/tokenizers#2217

New Contributors

• @​eunseo9311 made their first contribution in huggingface/tokenizers#2074
• @​cimeister made their first contribution in huggingface/tokenizers#1974

Full Changelog: huggingface/tokenizers@v0.23.1...v0.23.2

Release v0.23.1

TL;DR

tokenizers 0.23.1 is the first proper stable release in the 0.23 line — 0.23.0 only ever shipped as rc0 because the release pipeline itself was broken (Node side hadn't shipped multi-platform binaries since 2023, Python side was on pyo3 0.27 without free-threaded support). 0.23.1 is the version where everything actually goes out the door together: full Node multi-platform wheels for the first time in years, Python 3.14 (regular and free-threaded 3.14t), full type hints for every Python class, and a stack of measurable perf wins on the BPE / added-vocab hot paths.

There is no functional 0.23.0 published — we tag 0.23.1 directly so users don't accidentally pull a never-shipped version.

---

🚨 Breaking changes

• Drop Python 3.9 (#1952) — requires-python = ">=3.10"; 3.9 users stay on 0.22.x.
• add_tokens normalizes content at insertion (#1995) — re-saved tokenizer.json may differ in the added_tokens block. Existing files load unchanged.
• Type stubs are precise (#1928, #1997) — methods that returned Any now return real types; mypy --strict may surface previously-hidden errors. Stub layout also moved from tokenizers/<sub>/__init__.pyi to tokenizers/<sub>.pyi. This breaks the surface of some of the processors like RobertaProcessign's __init__ .
• 3.14t-only: setters/getters return PyResult<T> because of Arc<RwLock<Tokenizer>>; a poisoned lock surfaces as PyException instead of a panic.

... (truncated)

Commits

• 88a4498 add lock
• e4ea65f real release
• 7c5964b add lock
• bc405ad push rc0
• d582781 Add a ParityBpeTrainer example and list it in the trainers API docs (#2217)
• 447890f fix(python): pin the ruff rule set so a ruff release can't redden main (#2292)
• 68b3ab7 chore(node): take node security alerts from 33 to 0 (#2287)
• 74ef81f ci: pin every action to one SHA, drop stale audit suppressions (#2291)
• b26727f chore(node): drop 8 unused devDependencies (#2286)
• 7c7cfd8 chore: remove stale examples (kills 50% of dependabot traffic) (#2285)
• Additional commits viewable in compare view

Updates torchaudio from 2.10.0 to 2.11.0+cpu

Updates tpu-inference from 0.27.0 to 0.28.0

Release notes

Sourced from tpu-inference's releases.

v0.28.0

• Pypi
• Docker

tpu-inference v0.28.0

Cut: 2026-08-22 (initial), 2026-08-23 (post cherry-pick).
Tag commit: 9eca2a5a7.
vLLM LKG: d626108b1 (was f5bb701fa at v0.27.0 cut).
Compared to: releases/v0.27.0 — ~43 substantive PRs.

Highlights

• Fused MoE kernel integration and optimization (#3388), plus additional fused EP MoE kernels (#3040) and reduced shared-expert output over in-group TP under attention DP (#3435).
• DeepSeek V4 kernel wave continues (#3392, #3359, #3409, #3366, #3391, #3426, #3428) — mHC kernels ported, a fused reverse-rope-quant-wo-a kernel, per-KV-cache-group jit metadata, forward-pass copy elimination, and correctness hardening.
• PCP + ring attention (#3277). Ring attention support added for prefill context parallelism and enabled by default.
• Kernel-tuning infra: batched-RPA decode tuning and a gmm_v2 kernel tuner, with error recovery/resume for tuning runs (#3290);

New features

Model support

• DeepSeek V4 (continued, experimental): mHC kernels ported from g3 (#3392), fused reverse-rope-quant-wo-a kernel (#3359), attention metadata passed to jit once per KV-cache group (#3409), forward-pass parameter-copy elimination (#3366).
• gemma-4 on the torchax path: E2B/E4B accuracy fix — inline PLE + skip KV-cache writes for KV-shared layers (#3406); 26B/31B benchmarks pass (HBM headroom, startup timeout, variant bars) (#3408).

Capabilities

• RL rollouts: VllmSampler rollout class (#3257); engine pause/resume made functional on TPU (#3362).
• PCP + ring attention (#3277).
• transformers 5.15 per-layer Gemma4 configs supported alongside the LKG advance (#3372).
• Agentic benchmark: replay recorded rollout traces (#3365), replay tool-call idle time (#3383), honour --group-size and read traces from GCS (#3389).

Performance & kernel improvements

MoE / DP

• Fused MoE kernel integration and optimization (#3388).
• Additional fused EP MoE kernels (#3040).
• Reduce shared-expert output over in-group TP under attention DP (#3435).

Attention / MLA

• Stacked-RPA made mode-specific, with optimizations (#3351); decode stops over-fetching KV (#3361).
• Separate q,k,v projections in sliding-attention layers (#3376).
• ViT flash-attention scoped-vmem limit raised to 64MiB for multimodal (#3407).

Tuning

• Batched-RPA decode tuning + gmm_v2 kernel tuner; tuning infra gains error recovery/resume (#3290).

... (truncated)

Commits

• 9eca2a5 update docs for v0.28.0 release
• 9101adb [releases/v0.28.0] Cherry-pick #3446: mark Qwen2.5-VL-7B benchmarks unverifie...
• dacd27d Update verified commit hashes
• b8692ff [skip ci] Update nightly support matrices for vllm (v6e/v7x)
• 3582484 [tpu-inference] Add block-diffusion strategy configuration (#3242)
• 7482cf9 Update verified commit hashes
• 0fbf23a [skip ci] Update nightly support matrices (v6e/v7x)
• d253b79 Use tokamax release version to unblock TPU-inference release. (#3441)
• 57987c2 [Fix] Reduce shared-expert output over in-group TP under attention DP (#3435)
• ee1e17a [gemma4] Fix torchax-path E2B/E4B accuracy: inline PLE + skip KV-cache writes...
• Additional commits viewable in compare view

Updates absl-py from 2.1.0 to 2.5.0

Release notes

Sourced from absl-py's releases.

v2.5.0

Added

• (app) Add support for enabling Python profiling via the ABSL_PYTHON_PROFILE_FILE environment variable.
• (flags) Introduce --only_check_flags built-in flag to validate flag definitions without executing main().
• (flagsaver) Add support for async test functions in @flagsaver.flagsaver.
• (testing) Add record_property and get_recorded_properties methods to absltest.TestCase.

Changed

• (flags) Apply defensive copying in save_flag_values to prevent dictionary mutation flakiness.
• (logging) Update type signatures for absl.logging functions to support standard logging arguments (exc_info, stack_info, stacklevel, extra).
• (cleanup) Expand usage of FlagHolder objects in app and absltest.

Fixed

• (flags) Improve formatting and clarity of DuplicateFlagError message.
• (testing) Fix assertion logic in absltest.TestCase.create_file.
• (typechecking) Fix type annotations and compatibility with MyPy and Pyrefly.

v2.4.0

Added

• Added support for Python 3.14.
• Allow $PYTHONBREAKPOINT to affect runcall and post_mortem debugging.
• (logging) Propagate **kwargs in conditional/rate-limited logging functions.

Changed

• Dropped support for Python 3.8 and 3.9.
• (flags) Change internals of absl.flags.get_help_width() implementation.
• (cleanup) Modernize Bazel setup with MODULE.bazel.
• (cleanup) Modernize type annotations using Python 3.10+ features.

Fixed

• (flags) Fix duplicate flag definition when reloading a module.
• (typechecking) Correct type signature of absltest.skipThisClass.

v2.3.1

Changed

• (cleanup) Removed leftover code supporting Python < 3.8, as well as other references to older Python versions.

Fixed

• (typechecking) Fixed typechecking errors that appeared under mypy release 1.16

v2.3.0

... (truncated)

Changelog

Sourced from absl-py's changelog.

2.5.0 (2026-07-03)

Added

• (app) Add support for enabling Python profiling via the ABSL_PYTHON_PROFILE_FILE environment variable.
• (flags) Introduce --only_check_flags built-in flag to validate flag definitions without executing main().
• (flagsaver) Add support for async test functions in @flagsaver.flagsaver.
• (testing) Add record_property and get_recorded_properties methods to absltest.TestCase.

Changed

• (flags) Apply defensive copying in save_flag_values to prevent dictionary mutation flakiness.
• (logging) Update type signatures for absl.logging functions to support standard logging arguments (exc_info, stack_info, stacklevel, extra).
• (cleanup) Expand usage of FlagHolder objects in app and absltest.

Fixed

• (flags) Improve formatting and clarity of DuplicateFlagError message.
• (testing) Fix assertion logic in absltest.TestCase.create_file.
• (typechecking) Fix type annotations and compatibility with MyPy and Pyrefly.

2.4.0 (2026-01-28)

Added

• Added support for Python 3.14.
• Allow $PYTHONBREAKPOINT to affect runcall and post_mortem debugging.
• (logging) Propagate **kwargs in conditional/rate-limited logging functions.

Changed

• Dropped support for Python 3.8 and 3.9.
• (flags) Change internals of absl.flags.get_help_width() implementation.
• (cleanup) Modernize Bazel setup with MODULE.bazel.
• (cleanup) Modernize type annotations using Python 3.10+ features.

Fixed

• (flags) Fix duplicate flag definition when reloading a module.
• (typechecking) Correct type signature of absltest.skipThisClass.

2.3.1 (2025-07-03)

Changed

• (cleanup) Removed leftover code supporting Python < 3.8, as well as other references to older Python versions.

Fixed

... (truncated)

Commits

• a742ec3 Release Abseil-py 2.5.0
• f05ee36 Apply defensive programming in save_flag_values by copying the input.
• 1275a98 Fix a new Mypy issue in absl-py
• b7decb2 Suppressed a few Pyrefly errors
• 97f1f72 Fix typing errors.
• 44da4bd Replace "random" flag with a different one
• 5cd5771 Remove pyrefly suppressions that are not needed anymore.
• 3e729eb Add type annotations for absl.logging functions and fix tests to pass type ch...
• 354ee20 Another batch of Pyrefly-related fixes
• 5889704 Add record_property and get_recorded_properties in absltest.TestCase.
• Additional commits viewable in compare view

Updates accelerate from 1.13.0 to 1.14.0

Release notes

Sourced from accelerate's releases.

v1.14.0: AMD ROCm support, FSDP2 hardening

FSDP2 Improvements

This release brings a large batch of FSDP2 fixes and quality-of-life improvements: correct dtype handling on load, sharding of embeddings/norms, QLoRA crash prevention, and a more robust auto-wrap policy.

• Fsdp2 fully_shard embedding and norm by @​SunMarc in #4015
• Fix fsdp2 load full state dict dtype mismatch by @​SunMarc in #4021
• Fix region compilation fsdpv2 by @​SunMarc in #4022
• [FSDP2] Cast model to uniform dtype before fully_shard to fix mixed-dtype AssertionError by @​roycho96 in #3985
• [FSDP2] Auto-exclude non-floating frozen Params4bit from fully_shard to prevent QLoRA crash by @​roycho96 in #3987
• fix(FSDP2): auto-wrap policy ignoring _no_split_modules fallback by @​JohnGiorgi in #3999
• fix: use key-based matching in fsdp2_load_full_state_dict by @​roycho96 in #3982
• fix: add missing model_has_params4bit guard to fsdp2_load_full_state_dict call by @​roycho96 in #3981
• Fix to-fsdp2: drop REMOVED / NOT_YET_IMPLEMENTED FSDP1 keys instead of leaking them by @​lollinng in #4065
• Prevent double-wrapping models in prepare_model() by @​joshuaswanson in #3977

AMD ROCm support

Accelerate now works end-to-end on AMD ROCm devices. Thanks @​Abdennacer-Badaoui!

• Make accelerate work end-to-end on AMD ROCm by @​Abdennacer-Badaoui in #4025

Neuron

Further Neuron improvements to reduce recompilation and cover missing device cases.

• Add padded allgather and broadcast for Neuron devices to reduce recompilation by @​czkkkkkk in #4000
• fix: add missing neuron device case by @​michaelbenayoun in #4042

Quantization & Offloading

We improved offloading support for quantized models, including Torchao, int8, and tied-weight handling.

• Torchao offload by @​SunMarc in #3973
• Fix int8 offload hook detachment statistics restoration by @​jiqing-feng in #4044
• Fix keep_in_fp32_modules not working for tied weights in load_and_quantize_model by @​jiqing-feng in #4043
• Fix dtype_byte_size for FP8 fnuz / e8m0fnu dtypes by @​lollinng in #4063

Data Loading

• Feat: Support dynamic batch size in BatchSamplerShard with even_batches by @​yuxinyuan in #3969
• Fix iterable dataset sharding condition when n_shards == num_processes by @​SunMarc in #3958
• Fix implicit padding in split_between_processes when apply_padding=False and num_samples < num_processes by @​3manifold in #4052

Minor fixes

• [DeepSpeed] allow kernels flash-attn in SP by @​kashif in #3959
• Fix: Conditionally import torch.distributed.algorithms.join in accelerator.py by @​0xDELUXA in #3962
• Fix is_hf_initialized attribute by @​SunMarc in #3976
• feat(utils): add max reduction type by @​imstevenpmwork in #4027

... (truncated)

Commits

• beb0672 Release: v1.14.0
• a5cb37c Fix all tests (#4072)
• 14d7cdc Support dynamic batch size in BatchSamplerShard with even_batches (#3969)
• 77eb5cf fix(tracker): guard init_trackers and log against None kwargs (#4026)
• 379c8e0 fix(security): remediate workflow vulnerability in .github/workflows/build-do...
• 61e6420 Fix to-fsdp...

  Description has been truncated