chore(deps): bump the python-runtime group across 1 directory with 11 updates

[dependabot]

Bumps the python-runtime group with 11 updates in the /requirements directory:

Package	From	To
cryptography	46.0.5	50.0.1
openai	3.3.1	3.6.0
piper-tts	1.4.2	1.7.0
playwright	1.55.0	1.62.0
pydantic	2.13.4	2.13.5
scipy	1.18.0	1.18.1
tokenizers	0.22.2	0.23.1
vllm	0.27.1	0.28.0
huggingface-hub	1.28.0	1.29.0
ninja	1.13.0	1.13.2
transformers	5.15.1	5.16.1

Updates cryptography from 46.0.5 to 50.0.1

Changelog

Sourced from cryptography's changelog.

50.0.1 - 2026-08-25

* Updated Windows, macOS, and Linux wheels to be compiled with OpenSSL 4.0.2.
.. _v50-0-0:
50.0.0 - 2026-07-31

• SECURITY ISSUE: :func:~cryptography.hazmat.primitives.serialization.pkcs7.pkcs7_decrypt_der and its PEM and S/MIME variants no longer expose distinguishable errors or timing when unwrapping a RecipientInfo's encryptedKey, which could act as a Bleichenbacher oracle for callers that decrypt untrusted messages. A random key is now substituted on failure, as described in :rfc:3218. Credit to @​X1AOxiang for reporting the issue. CVE-2026-69247
• Deprecated Diffie-Hellman key exchange over finite fields (FFDH). Everything FFDH is deprecated, including the types in cryptography.hazmat.primitives.asymmetric.dh and loading FFDH keys or parameters with the key loading APIs. Users should migrate to a more modern key exchange algorithm.
• Added xof() class methods to :class:~cryptography.hazmat.primitives.hashes.SHAKE128 and :class:~cryptography.hazmat.primitives.hashes.SHAKE256 for constructing algorithm instances configured for use with :class:~cryptography.hazmat.primitives.hashes.XOFHash.
• The :mod:X.509 verification <cryptography.x509.verification> APIs are now considered stable and are subject to our API stability policy.
• Added the :doc:/cobblestone recipe, an implementation of the Cobblestone-128 and Cobblestone-256 instantiations of the C2SP chunked-encryption specification <https://c2sp.org/chunked-encryption>_ for streaming authenticated encryption of large messages.
• Parsing a Signed Certificate Timestamp list now rejects encodings that carry trailing bytes after the list or after an individual SCT, instead of silently ignoring them.
• Added support for using :class:~cryptography.x509.Name as a field type in the :doc:/hazmat/asn1/index module.
• Loading a public key or an EC private key now rejects DER where the subjectPublicKey (or EC publicKey) BIT STRING declares a non-zero number of unused bits, instead of silently ignoring it.
• Parsing a CRL entry's InvalidityDate extension now rejects a GeneralizedTime that carries fractional seconds or another non-DER form, matching the strict encoding already required for every other X.509 time field.
• :func:~cryptography.x509.ocsp.load_der_ocsp_request and :func:~cryptography.x509.ocsp.load_der_ocsp_response now reject a request or response whose version field is not v1, the only version defined by RFC 6960, matching the version validation already performed when loading

... (truncated)

Commits

• ffde75a bump for 50.0.1 + changelog (#15520)
• dcb7050 Prepare for 50.0.0 release (#15372)
• 53fccd9 Don't leak how PKCS#7 encryptedKey decryption failed (#15369)
• d472f97 Add from __future__ import annotations to all src/ Python files (#15371)
• 908773d Bump downstream dependencies in CI (#15368)
• 2cc07cc Bump BoringSSL, OpenSSL, AWS-LC in CI (#15367)
• c94ede9 chore(deps): bump ruff from 0.16.0 to 0.16.1 (#15366)
• 67a8308 chore(deps): bump virtualenv from 21.7.0 to 21.7.1 (#15365)
• 95018ff Release the GIL in one-shot AEAD encrypt/decrypt (#15361)
• 6954733 Release the GIL during DH and DSA parameter generation (#15364)
• Additional commits viewable in compare view

Updates openai from 3.3.1 to 3.6.0

Release notes

Sourced from openai's releases.

v3.6.0

3.6.0 (2026-08-27)

Features

• api: add compute_units to Responses and Chat Completions usage (#3749) (52421d1)

Bug Fixes

• auth: harden X.509 workload identity integration (#3740) (fc3ad6c)

Chores

• deps-dev: bump @​stdy/cli from 0.22.1 to 0.22.2 (#3719) (4f5598c)
• deps-dev: bump mypy from 1.17 to 2.3.1 (#3747) (0b52c9e)
• deps-dev: bump pandas-stubs from 2.2.2.240807 to 2.3.3.260113 (#3659) (95f0b43)
• deps-dev: bump pyright from 1.1.399 to 1.1.413 (#3744) (9917c6e)
• deps-dev: bump rich from 14.2.0 to 15.0.0 (#3717) (7a5484d)
• deps: bump actions/checkout from 6.0.2 to 7.0.1 (#3665) (d0a2550)
• deps: bump actions/download-artifact from 6.0.0 to 8.0.1 (#3669) (f627619)
• deps: bump github/codeql-action/init from 4.37.1 to 4.37.7 (#3745) (a36010d)

Build System

• deps: bump actions/setup-python from 5.6.0 to 7.0.0 (#3672) (d765db7)
• deps: bump pypa/gh-action-pypi-publish from 1.14.0 to 1.14.2 (#3666) (dacbb66)

v3.5.0

3.5.0 (2026-08-27)

Features

• api: make function call output call IDs optional (#3738) (c74501d)

v3.4.0

3.4.0 (2026-08-25)

Features

• api: Add obfuscation field to ChatCompletionChunk (#3690) (c7d8e1d)
• api: add project residency configuration and cost quantity units (#3726) (bc4f8ef)

Bug Fixes

... (truncated)

Changelog

Sourced from openai's changelog.

3.6.0 (2026-08-27)

Features

• api: add compute_units to Responses and Chat Completions usage (#3749) (52421d1)

Bug Fixes

• auth: harden X.509 workload identity integration (#3740) (fc3ad6c)

Chores

• deps-dev: bump @​stdy/cli from 0.22.1 to 0.22.2 (#3719) (4f5598c)
• deps-dev: bump mypy from 1.17 to 2.3.1 (#3747) (0b52c9e)
• deps-dev: bump pandas-stubs from 2.2.2.240807 to 2.3.3.260113 (#3659) (95f0b43)
• deps-dev: bump pyright from 1.1.399 to 1.1.413 (#3744) (9917c6e)
• deps-dev: bump rich from 14.2.0 to 15.0.0 (#3717) (7a5484d)
• deps: bump actions/checkout from 6.0.2 to 7.0.1 (#3665) (d0a2550)
• deps: bump actions/download-artifact from 6.0.0 to 8.0.1 (#3669) (f627619)
• deps: bump github/codeql-action/init from 4.37.1 to 4.37.7 (#3745) (a36010d)

Build System

• deps: bump actions/setup-python from 5.6.0 to 7.0.0 (#3672) (d765db7)
• deps: bump pypa/gh-action-pypi-publish from 1.14.0 to 1.14.2 (#3666) (dacbb66)

3.5.0 (2026-08-27)

Features

• api: make function call output call IDs optional (#3738) (c74501d)

3.4.0 (2026-08-25)

Features

• api: Add obfuscation field to ChatCompletionChunk (#3690) (c7d8e1d)
• api: add project residency configuration and cost quantity units (#3726) (bc4f8ef)

Bug Fixes

• api: encode Realtime call offers and session configuration (#3736) (555ac48)
• apply consistent origin checks to WebSocket redirects (#3693) (1b324d0)

... (truncated)

Commits

• 1cfa80a release: 3.6.0 (#3752)
• 9917c6e chore(deps-dev): bump pyright from 1.1.399 to 1.1.413 (#3744)
• 0b52c9e chore(deps-dev): bump mypy from 1.17 to 2.3.1 (#3747)
• 4f5598c chore(deps-dev): bump @​stdy/cli from 0.22.1 to 0.22.2 (#3719)
• 95f0b43 chore(deps-dev): bump pandas-stubs from 2.2.2.240807 to 2.3.3.260113 (#3659)
• a36010d chore(deps): bump github/codeql-action/init from 4.37.1 to 4.37.7 (#3745)
• 7a5484d chore(deps-dev): bump rich from 14.2.0 to 15.0.0 (#3717)
• d0a2550 chore(deps): bump actions/checkout from 6.0.2 to 7.0.1 (#3665)
• dacbb66 build(deps): bump pypa/gh-action-pypi-publish from 1.14.0 to 1.14.2 (#3666)
• d765db7 build(deps): bump actions/setup-python from 5.6.0 to 7.0.0 (#3672)
• Additional commits viewable in compare view

Updates piper-tts from 1.4.2 to 1.7.0

Release notes

Sourced from piper-tts's releases.

v1.7.0

• Add Japanese phonemizer using OpenJTalk (pyopenjtalk-plus) in the new ja extra
  • --data.phoneme_type japanese for training; "phoneme_type": "japanese" in a voice config for synthesis
  • espeak-ng has no kanji coverage (it reads out Unicode character names) and no pitch accent
  • Full-context labels are parsed for pitch accent and mapped to IPA, so Japanese voices stay compatible with the IPA-based (espeak) warmstart
• Add script/setup --ja, and install the ja extra in CI so the Japanese tests run
• libpiper: add piper_create_options and piper_create_with_options(), with piper_create() kept as a wrapper for ABI compatibility

v1.6.1

• Run the g2pW model through piper.g2pw_onnx instead of g2pw.api, dropping torch (~750 MB installed) and requests from the zh extra
  • g2pw.api imports torch only to build padded tensors and iterate batches; the model itself already ran under onnxruntime
  • Also 1.5-2x faster, since it no longer forks DataLoader worker processes on every call
  • g2pW is still required, for its pinyin/bopomofo lookup tables

v1.6.0

• Add Hebrew phonemizer using Nakdimon

v1.5.0

• Add libpiper C++ CLI executable ported from the legacy Piper repository, plus a C++ test suite
• Fix libpiper builds on Windows (MSVC, MSYS2-GCC) and Windows CI
• Bump embedded espeak-ng version
• Add default speaker id for multi-speaker voices
• Add vowel clustering support (--data.vowel_clusters)
• Add in-memory patching for alignments
• Training: add MRD (Multi-Resolution STFT) discriminator, loss/MOS tracking with UTMOS, silence-trim fixes, and dataloader performance improvements
• Pass custom phoneme id map when training

Changelog

Sourced from piper-tts's changelog.

1.7.0

• Add Japanese phonemizer using OpenJTalk (pyopenjtalk-plus) in the new ja extra
  • --data.phoneme_type japanese for training; "phoneme_type": "japanese" in a voice config for synthesis
  • espeak-ng has no kanji coverage (it reads out Unicode character names) and no pitch accent
  • Full-context labels are parsed for pitch accent and mapped to IPA, so Japanese voices stay compatible with the IPA-based (espeak) warmstart
• Add script/setup --ja, and install the ja extra in CI so the Japanese tests run
• libpiper: add piper_create_options and piper_create_with_options(), with piper_create() kept as a wrapper for ABI compatibility

1.6.1

• Run the g2pW model through piper.g2pw_onnx instead of g2pw.api, dropping torch (~750 MB installed) and requests from the zh extra
  • g2pw.api imports torch only to build padded tensors and iterate batches; the model itself already ran under onnxruntime
  • Also 1.5-2x faster, since it no longer forks DataLoader worker processes on every call
  • g2pW is still required, for its pinyin/bopomofo lookup tables

1.6.0

• Add Hebrew phonemizer using Nakdimon

1.5.0

• Add libpiper C++ CLI executable ported from the legacy Piper repository, plus a C++ test suite
• Fix libpiper builds on Windows (MSVC, MSYS2-GCC) and Windows CI
• Bump embedded espeak-ng version
• Add default speaker id for multi-speaker voices
• Add vowel clustering support (--data.vowel_clusters)
• Add in-memory patching for alignments
• Training: add MRD (Multi-Resolution STFT) discriminator, loss/MOS tracking with UTMOS, silence-trim fixes, and dataloader performance improvements
• Pass custom phoneme id map when training

Commits

• 7b8e8f7 Merge pull request #275 from OHF-Voice/claude/version-1-7-0-release-2b37ec
• 222159f Silence pylint redefined-outer-name in the phonemizer tests
• af4ff76 Make the mypy exclude for vendored VITS code actually work
• c0f3289 Release 1.7.0: add --ja setup flag and test Japanese in CI
• 4ab2d74 Merge pull request #274 from OHF-Voice/synesthesiam-20260811-japanese
• 711b7b4 Refactor: introduce piper_create_options with same-name overloading (#270)
• 5a9a99d Bump version
• ae655b9 Merge pull request #269 from OHF-Voice/torch-free-g2pw
• 03ee769 Add license
• 677992e Run g2pW under onnxruntime without torch
• Additional commits viewable in compare view

Updates playwright from 1.55.0 to 1.62.0

Release notes

Sourced from playwright's releases.

v1.62.0

🖼️ WebP screenshots

page.screenshot() and locator.screenshot() can now capture screenshots in the WebP format. Playwright infers the format from a .webp file extension, or you can set the type explicitly. Quality 100 (the default) is lossless, while lower values use lossy compression:

page.screenshot(path="homepage.webp", quality=50)

New APIs

• New scroll option ("auto" | "none") on actions to opt out of Playwright's automatic scrolling into view.
• New locator.wait_for_function() waits until a function called with the matching element returns a truthy value.
• New apiResponse.timing() returns resource timing information for an API response.

Announcements

• 📋 The clipboard is now isolated from the operating system in headless mode, so tests that use navigator.clipboard no longer read or overwrite the clipboard of the machine running them.
• ⚠️ Debian 11 is not supported anymore.

Browser Versions

• Chromium 151.0.7922.34
• Mozilla Firefox 153.0
• WebKit 26.5

This version was also tested against the following stable channels:

• Google Chrome 151
• Microsoft Edge 151

v1.61.0

🔑 WebAuthn passkeys

New Credentials virtual authenticator, available via browserContext.credentials, lets tests register passkeys and answer navigator.credentials.create() / navigator.credentials.get() ceremonies in the page — no real hardware key required, works in all browsers:

context = browser.new_context()
Seed a passkey your backend provisioned for a test user.
context.credentials.create("example.com",
id=credential_id,
user_handle=user_handle,
private_key=private_key,
public_key=public_key,
)
context.credentials.install()
page = context.new_page()
page.goto("https://example.com/login")
The page's navigator.credentials.get() is answered with the seeded passkey.

You can also let the app register a passkey once in a setup test, read it back with credentials.get(), and seed it into later tests — see Credentials for details.

... (truncated)

Commits

• 3b7c24c chore: roll Playwright to 1.62.0 (#3161)
• 1db079f build(deps): bump typing-extensions from 4.15.0 to 4.16.0 (#3162)
• 8335307 build(deps): bump greenlet from 3.5.3 to 3.5.4 (#3163)
• 5c71b2b build(deps): bump mypy from 2.1.0 to 2.3.0 (#3150)
• d0c8d03 build(deps-dev): bump setuptools-scm from 9.2.2 to 10.2.1 (#3151)
• ca0319c build(deps): bump build from 1.4.4 to 1.5.0 (#3152)
• 53802fe build(deps): bump requests from 2.33.0 to 2.34.2 (#3153)
• 9bc998f build(deps): bump actions/setup-python from 6 to 7 in the actions group (#3154)
• 9a10128 perf: speed up Python stack capture (#3147)
• c3eff76 fix: show not_ prefix in call log when negative assertions fail (#3033)
• Additional commits viewable in compare view

Updates pydantic from 2.13.4 to 2.13.5

Release notes

Sourced from pydantic's releases.

v2.13.5 (2026-08-28)

What's Changed

Fixes

• Allow reuse of validators when plugins are set by @​Viicos in #13535
• Fix missing GC traversal on some pydantic-core struct fields by @​Viicos in #13624
• Fix missing GC traversal in pydantic-core for GeneralFieldsSerializer by @​Viicos in #13629
• Count validated model fields once in smart unions by @​tamird in #13731

Changelog

Sourced from pydantic's changelog.

v2.13.5 (2026-08-28)

GitHub release

What's Changed

Fixes

• Allow reuse of validators when plugins are set by @​Viicos in #13535
• Fix missing GC traversal on some pydantic-core struct fields by @​Viicos in #13624
• Fix missing GC traversal in pydantic-core for GeneralFieldsSerializer by @​Viicos in #13629
• Count validated model fields once in smart unions by @​tamird in #13731

Commits

• 001dea0 Bump pypa/gh-action-pypi-publish action to v1.14.2
• 558379f Bump twine to v7.0.0
• 2cfd5d3 Do not check for docs build
• a735bee Fix more Clippy lints
• 7eed4a1 Fix Clippy 0.1.95 warnings
• b353bbb Prepare release v2.13.5
• 63d2ccc Count validated model fields once in smart unions
• a53ec2e Speed up PyPy CI tests
• d65e0f9 Workaround circular import error in Mypy
• 47a6dbf Fix missing GC traversal in pydantic-core for GeneralFieldsSerializer
• Additional commits viewable in compare view

Updates scipy from 1.18.0 to 1.18.1

Release notes

Sourced from scipy's releases.

SciPy 1.18.1 Release Notes

SciPy 1.18.1 is a bug-fix release with no new features compared to 1.18.0. This release includes binaries on PyPI for Python 3.15, and the minimum required version of the GCC toolchain has been increased to 10.3.0.

Authors

• Name (commits)
• Wasim Akram (2) +
• Jake Bowhay (1)
• Michael Bratsch (1)
• Dietrich Brunn (1)
• Evgeni Burovski (1)
• Aadya Chinubhai (1)
• Lucas Colley (3)
• Tekin Ertekin (1) +
• Fuyugithub (2) +
• Ralf Gommers (11)
• Joren Hammudoglu (1)
• Ijtihed Kilani (1) +
• Andrew Nelson (2)
• Ilhan Polat (1)
• Tyler Reddy (39)
• romao05 (1) +
• Michael Simacek (1) +
• Jacob Vanderplas (2)

A total of 18 people contributed to this release. People with a "+" by their names contributed a patch for the first time. This list of names is automatically generated, and may not be fully complete.

Note that the source and binary assets associated with this release were published to PyPI using trusted publishing, and so the trusted assets and their hashes are made available more securely at https://pypi.org/project/scipy/1.18.1/ rather than providing them here in a less secure manner.

A complete list of issues and pull requests associated with this release is available in the associated README.txt.

Commits

• e4e854e REL: SciPy 1.18.1 rel commit
• 3c8e5f8 BLD, MAINT: pythran bounds for 1.18.1 (#25978)
• 54c0f95 BLD, MAINT: pythran bounds for 1.18.1
• 45a355d Merge pull request #25884 from tylerjereddy/treddy_backports_1.18.1
• a717e28 DOC: PR 25884 revisions
• bb489c0 DOC: PR 25884 revisions
• 09bca24 BLD: package: pin meson on win-64 to avoid clang-cl bug (#25913)
• 095fe56 DOC, BLD: PR 25884 revisions
• e66ea64 DOC: update minimum GCC version in toolchain roadmap
• c021653 TST: tolerance bumps and skips on 32-bit platforms
• Additional commits viewable in compare view

Updates tokenizers from 0.22.2 to 0.23.1

Release notes

Sourced from tokenizers's releases.

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.

---

⚡ Performance — measured locally on this Mac, not lifted from PRs

Run with cargo bench --bench <name> -- --save-baseline v0_22_2 on v0.22.2, then --baseline v0_22_2 on v0.23.1. Numbers are point-in-time wall clock on a single laptop; relative deltas are what matters, absolute numbers will differ on CI hardware.

Added-vocabulary deserialize — the headline win (#1995, #1999)

bench: improve added_vocab_deserialize to reflect real-world workloads (#2000) is now representative of how transformers actually loads tokenizer.json files. The combined effect of daachorse for the matching automaton plus the normalize-on-insert refactor is enormous on this workload:

benchmark	v0.22.2	v0.23.1	change
100k tokens, special, no norm	~410 ms	248 ms	−40%
100k tokens, non-special, no norm	~7.1 s	273 ms	−96%
100k tokens, special, NFKC	~395 ms	235 ms	−40%
100k tokens, non-special, NFKC	~7.4 s	290 ms	−96%
400k tokens, special, no norm	~15 s	980 ms	−94%

Real-world impact: loading a Llama-3-style tokenizer with a large set of added tokens dropped from "noticeable pause" to "instant".

BPE encode

benchmark	v0.22.2	v0.23.1	change
BPE GPT2 encode batch, no cache	530 ms	446 ms	−16%
BPE GPT2 encode batch (cached)	690 ms	685 ms	noise
BPE GPT2 encode (single)	1.95 s	1.94 s	noise
BPE Train (small)	32.6 ms	31.5 ms	−3%
BPE Train (big)	1.01 s	988 ms	−2%

The BPE per-thread cache PR (#2028) shows much larger wins on highly-parallel workloads (+47–62% at 88+ threads on a server box, per the PR's own measurements on Vera). Single-thread batch numbers above are flat or slightly improved because cache-hit overhead was already low without contention.

Llama-3 encode

... (truncated)

Commits

• 7f1623b Bump version to 0.23.1
• bbe43ad ci: release workflow fixes (node + python) (#2043)
• ab0c5d8 Fix node release (#2034)
• decd8e0 bindings/python: free-threaded Python (3.14t) support (#2041)
• 3992692 update for release (#2033)
• bcdd25b BPE cache: per-thread read-through cache to avoid RwLock atomics on hits (#2028)
• 618eb38 Bump follow-redirects in /tokenizers/examples/unstable_wasm/www (#2024)
• b6b1688 chore: bump doc-builder SHA for PR upload workflow (#2025)
• 19015d6 fix: use uvx --with cairosvg instead of uv pip install --system (#2021)
• efbcc68 Ci benchmarks (#2019)
• Additional commits viewable in compare view

Updates vllm from 0.27.1 to 0.28.0

Release notes

Sourced from vllm's releases.

v0.28.0

Highlights

This release features 584 commits from 270 contributors (76 new)!

• Kimi-K3 performance push: a major optimization effort for Kimi-K3 across the stack — Decode Context Parallel (DCP) support (#50484), fused FlashKDA decode and prefill kernels (#50654, #51311, #52458), SiTU activation support for MegaMoE (#50510), GEMM-RS for sequence parallelism (#52079), combined all-gathers with 1.5~3x kernel-level speedup (#51070), an adaptive speculative token budget delivering ~60% better DSpark TTFT (#51725), and optional shared-expert sharding saving ~17 GiB of memory per GPU (#50912). Kimi-K3 also now runs on ROCm with the V2 model runner (#51653).
• DeepSeek V4: sparse MLA now works end-to-end for plain decode, MTP, and DSpark speculative decoding (#51538), joined by AMD Quark NVFP4 support (#47972), reasoning-effort prompts and mappings (#50580), sparse top-k metadata kernel optimizations (#52084, #51967), narrowed eager CUDA graph regions (#51430, #52401), and ROCm enablement on gfx11 and gfx950 (#47017, #52212).
• Speculative decoding advances: DFlash2 with local convolution and a candidate selector (#52816), DSpark confidence-scheduled verification (#47808), and async scheduling auto-enabled for draft models (#48341).
• Model Runner V2 maturation: E/P/D disaggregation (#38390), weight offloading (#51413), multi-layer MTP KV cache support (#50062), encoder CUDA graphs (#49852), decoder token-wise pooling (#50931) plus Transformers pooling models (#52425), attention-free models (#52374), and thinking_token_budget support (#46727).
• Tiered KV cache offloading: disk offloading support (#49644), out-of-tree secondary tier managers via module_path (#51007), partial secondary-tier load results (#50321), tiering metrics (#48798), and a canonical CPU layout for parallelism-agnostic offload (#48414).
• Rust frontend & gRPC: a standalone renderer (#50289), multimodal image inference over gRPC (#50368), explicit data-parallel rank routing (#51178), and RL lifecycle control (#51316), with protobuf schemas now published to Buf (#51276).
• New defaults: max_num_batched_tokens raised from 8192 to 16384 (#51726), prefix caching enabled by default for Mamba models (#50991), and the Blackwell CUDA graph capture default raised to 1024 (#49390).
• Breaking changes: bitsandbytes support migrated to an out-of-tree plugin (#43529); Transformers bumped to 5.15.0 (#51668); the deprecated calculate_kv_scales runtime KV scale calculation was removed (#49389); override_attention_dtype was removed (#48684).

Release Artifacts

Python Wheels

Platform	Install
PyPI (CUDA 13.0)	pip install vllm
PyPI (CUDA 13.0, uv)	uv pip install vllm --torch-backend=auto
ROCm	pip install vllm --extra-index-url https://wheels.vllm.ai/rocm/0.28.0/rocm722

Docker Images

Platform	Docker Image
CUDA 13.0 (Default)	docker pull vllm/vllm-openai:v0.28.0 (v0.28.0-cu130 also works)
CUDA 12.9	docker pull vllm/vllm-openai:v0.28.0-cu129
CUDA 13.0 + Ubuntu 24.04	docker pull vllm/vllm-openai:v0.28.0-ubuntu2404
CUDA 12.9 + Ubuntu 24.04	docker pull vllm/vllm-openai:v0.28.0-cu129-ubuntu2404
ROCm	docker pull vllm/vllm-openai-rocm:v0.28.0
CPU	docker pull vllm/vllm-openai-cpu:v0.28.0
XPU	docker pull vllm/vllm-openai-xpu:v0.28.0

Other Artifacts

Pre-built release artifacts are available in the Assets section at the bottom of this page, including:

• Source distribution tarball
• CUDA 12.9 Python wheels for x86_64 and arm64
• CUDA 13.0 Python wheels for x86_64 and arm64
• CPU Python wheels for x86_64, arm64, and macOS

Model Support

• New models: Muse Glimmer (#51655), Ling 3.0 Flash with BF16, MTP, and parser support (#51045) plus an FP8 variant (#51265) and hybrid MXFP4 routed experts (#52114), Dots3 NOTE native multimodal support (#51255), and Interns2mobius (#51149).
• Qwen: Qwen3.8 enabled on AMD ROCm (#50068), fused CUDA post-conv MTP decode kernel for Qwen3.5 GDN (#51674), GDN gates aligned with speculative tokens (#51812), and Qwen3.5 fixes for text-only checkpoints (#50734, #50355).
• Transformers modeling backend: MLA support (#48250), hardware-agnostic model definition (#49458), fully generalized input embedding handling (#51247), logit softcapping (#52173), and a hardened multimodal path (#51408, #51657).
• LoRA: vision tower LoRA for Gemma4 (#42662), tower/connector LoRA for Keye (#51780) and Ultravox (#48215).

... (truncated)

Commits

• 2cf0a69 [CI/Build] Pin Cython below 3.3 for arm64 tilelang sdist (#53358)
• 9991487 Revert "[Kernel] Gemma-4 FA4 FP8 Kernel" (#52987)
• 1af5a38 [CI][Release] Extend DSv4 engine readiness timeout (#53252)
• 74a6576 [Spec Decode] DFlash2: local convolution + candidate selector (#52816)
• cf902bd [CI/Build] Fix accident pre-commit breakage due to concurrent merge (#52578)
• cd6ae1e [Bugfix][Security] Guard _load_ov2_processor with resolve_trust_remote_code (...
• b7d8e0f [kv_offload] fix(metrics): rename kv_offload_tiering_block_{queries,hits} → c...
• c0eac6f [ROCm]: Bump triton 3.7 commit (#52819)
• 855e09e [Kernel] SM120: stop routing misaligned-M blockwise FP8 GEMMs to the small-M ...
• 1b6ed9b [EPD] Allow KV consumers to omit MM embeddings (#52697)
• Additional commits viewable in compare view

Updates huggingface-hub from 1.28.0 to 1.29.0

Release notes

Sourced from huggingface-hub's releases.

[v1.29.0] Fix Xet downloads rate limits, bucket visibility updates, and security fixes

⚡ Xet downloads no longer make one API call per file

Since v1.19.0, downloading a repository with hf_xet eagerly requested a xet read token for every single file, because each download group was built without cached connection info. On repos with many files this quickly added up — a 77k-file repo made ~1,500 Hub API calls per minute — and eventually hit the rate limiter, leaving snapshot_download appearing stalled for minutes before failing with a 429 Too Many Requests. This release restores the Python-side connection info cache so the endpoint and token are reused across download groups, skipping the eager per-file token request entirely. Large downloads are both faster and far less likely to be rate-limited.

• [Xet] Cache connection info to avoid one token request per file...

  Description has been truncated