Bump the pip group across 4 directories with 4 updates

[dependabot]

Bumps the pip group with 3 updates in the /.aitk/requirements directory: transformers, aiohttp and tornado.
Bumps the pip group with 1 update in the /Qwen-Qwen3.5-2B/cuda directory: transformers.
Bumps the pip group with 1 update in the /Qwen-Qwen3.5-2B/webgpu directory: transformers.
Bumps the pip group with 4 updates in the /meta-llama-Llama-3.1-8B-Instruct/QAIRT directory: transformers, aiohttp, tornado and bleach.

Updates transformers from 4.51.3 to 5.5.0

Release notes

Sourced from transformers's releases.

Release v5.5.0

New Model additions

Gemma4

Gemma 4 is a multimodal model with pretrained and instruction-tuned variants, available in 1B, 13B, and 27B parameters. The architecture is mostly the same as the previous Gemma versions. The key differences are a vision processor that can output images of fixed token budget and a spatial 2D RoPE to encode vision-specific information across height and width axis.

You can find all the original Gemma 4 checkpoints under the Gemma 4 release.

The key difference from previous Gemma releases is the new design to process images of different sizes using a fixed-budget number of tokens. Unlike many models that squash every image into a fixed square (like 224×224), Gemma 4 keeps the image's natural aspect ratio while making it the right size. There a a couple constraints to follow:

• The total number of pixels must fit within a patch budget
• Both height and width must be divisible by 48 (= patch size 16 × pooling kernel 3)

[!IMPORTANT] Gemma 4 does not apply the standard ImageNet mean/std normalization that many other vision models use. The model's own patch embedding layer handles the final scaling internally (shifting values to the [-1, 1] range).

The number of "soft tokens" (aka vision tokens) an image processor can produce is configurable. The supported options are outlined below and the default is 280 soft tokens per image.

Soft Tokens	Patches (before pooling)	Approx. Image Area
70	630	~161K pixels
140	1,260	~323K pixels
280	2,520	~645K pixels
560	5,040	~1.3M pixels
1,120	10,080	~2.6M pixels

To encode positional information for each patch in the image, Gemma 4 uses a learned 2D position embedding table. The position table stores up to 10,240 positions per axis, which allows the model to handle very large images. Each position is a learned vector of the same dimensions as the patch embedding. The 2D RoPE which Gemma 4 uses independently rotate half the attention head dimensions for the x-axis and the other half for the y-axis. This allows the model to understand spatial relationships like "above," "below," "left of," and "right of."

NomicBERT

NomicBERT is a BERT-inspired encoder model that applies Rotary Position Embeddings (RoPE) to create reproducible long context text embeddings. It is the first fully reproducible, open-source text embedding model with 8192 context length that outperforms both OpenAI Ada-002 and OpenAI text-embedding-3-small on short-context MTEB and long context LoCo benchmarks. The model generates dense vector embeddings for various tasks including search, clustering, and classification using specific instruction prefixes.

Links: Documentation | Paper

• Internalise the NomicBERT model (#43067) by @​ed22699 in #43067

MusicFlamingo

Music Flamingo is a fully open large audio–language model designed for robust understanding and reasoning over music. It builds upon the Audio Flamingo 3 architecture by including Rotary Time Embeddings (RoTE), which injects temporal position information to enable the model to handle audio sequences up to 20 minutes. The model features a unified audio encoder across speech, sound, and music with special sound boundary tokens for improved audio sequence modeling.

Links: Documentation | Paper

• Add Music Flamingo (#43538) by @​lashahub in #43538

... (truncated)

Commits

• c1c3424 update
• 20bff68 update release workflow
• 8956441 v5.5.0
• 5135e5e casually dropping the most capable open weights on the planet (#45192)
• a594e09 Internalise the NomicBERT model (#43067)
• 4932e97 Fix resized LM head weights being overwritten by post_init (#45079)
• 57e8413 [Qwen3.5 MoE] Add _tp_plan to ForConditionalGeneration (#45124)
• b10552e Fix TypeError: 'NoneType' object is not iterable in GenerationMixin.generate ...
• 423f2a3 fix(models): Fix dtype mismatch in SwitchTransformers and TimmWrapperModel (#...
• ade7a05 Generalize gemma vision mask to videos (#45185)
• Additional commits viewable in compare view

Updates aiohttp from 3.14.0 to 3.14.1

Changelog

Sourced from aiohttp's changelog.

3.14.1 (2026-06-07)

Bug fixes

• Fixed a race condition in :py:class:~aiohttp.TCPConnector where closing the connector while a DNS resolution was in-flight could raise :py:exc:AttributeError instead of :py:exc:~aiohttp.ClientConnectionError -- by :user:goingforstudying-ctrl.

  Related issues and pull requests on GitHub: :issue:12497.

• Fixed CancelledError not closing a connection -- by :user:aiolibsbot.

  Related issues and pull requests on GitHub: :issue:12795.

• Tightened up some websocket parser checks -- by :user:Dreamsorcerer.

  Related issues and pull requests on GitHub: :issue:12817.

• Fixed :class:~aiohttp.CookieJar dropping the host-only flag of cookies when persisted with :meth:~aiohttp.CookieJar.save and reloaded with :meth:~aiohttp.CookieJar.load, so a cookie set without a Domain attribute is again scoped to the exact host that set it after a reload; the absolute expiration deadline is now persisted as well, so a reloaded cookie keeps its original lifetime instead of being rescheduled from the load time. :meth:~aiohttp.CookieJar.load now replaces the jar contents rather than merging onto prior state, and loaded cookies pass through the same acceptance rules as :meth:~aiohttp.CookieJar.update_cookies, so a cookie for an IP-address host is dropped when loaded into a jar created without unsafe=True -- by :user:bdraco.

  Related issues and pull requests on GitHub: :issue:12824.

• Scoped :class:~aiohttp.DigestAuthMiddleware credentials to the origin of the first request it handles, so a redirect to a different origin no longer triggers a digest response computed from the configured credentials; a challenge from another origin is only answered when that origin falls within a protection space advertised by the anchor origin through the RFC 7616 domain directive -- by :user:bdraco.

  Related issues and pull requests on GitHub: :issue:12825.

• Fixed the C HTTP parser not enforcing max_line_size on a request target or response reason phrase that is split across multiple reads; each fragment was checked on its own, so an accumulated line could exceed the limit without raising LineTooLong. The accumulated length is now checked, matching the pure-Python parser -- by :user:bdraco.

  Related issues and pull requests on GitHub:

... (truncated)

Commits

• 9c35d03 Release v3.14.1 (#12864)
• 38b956c [PR #12861/59684b5c backport][3.14] Revert "Drop list compression (#12857)" (...
• 8f31009 [PR #12857/69dff14d backport][3.14] Drop list compression (#12858)
• dfdfa9d [PR #12830/93a2b1c3 backport][3.14] Bound pipelined request queue per connect...
• 0e9cedd [PR #12827/ccf218ab backport][3.14] Numeric ipv4 resolver bypass (#12849)
• a762eda [PR #12831/1ac92dae backport][3.14] Payload close on disconnect (#12843)
• a329a7a [PR #12824/60b85e98 backport][3.14] Preserve host-only cookie scope across Co...
• 4f7480e [PR #12828/13b635d7 backport][3.14] Bounded unread compressed drain (#12845)
• 5ab61bb [PR #12826/36df6c13 backport][3.14] Enforce max_line_size on fragmented reque...
• 3912667 [3.14] Add test that env proxy auth is scoped to the redirect-selected proxy ...
• Additional commits viewable in compare view

Updates tornado from 6.5.5 to 6.5.7

Changelog

Sourced from tornado's changelog.

Release notes

.. toctree:: :maxdepth: 2

releases/v6.5.7 releases/v6.5.6 releases/v6.5.5 releases/v6.5.4 releases/v6.5.3 releases/v6.5.2 releases/v6.5.1 releases/v6.5.0 releases/v6.4.2 releases/v6.4.1 releases/v6.4.0 releases/v6.3.3 releases/v6.3.2 releases/v6.3.1 releases/v6.3.0 releases/v6.2.0 releases/v6.1.0 releases/v6.0.4 releases/v6.0.3 releases/v6.0.2 releases/v6.0.1 releases/v6.0.0 releases/v5.1.1 releases/v5.1.0 releases/v5.0.2 releases/v5.0.1 releases/v5.0.0 releases/v4.5.3 releases/v4.5.2 releases/v4.5.1 releases/v4.5.0 releases/v4.4.3 releases/v4.4.2 releases/v4.4.1 releases/v4.4.0 releases/v4.3.0 releases/v4.2.1 releases/v4.2.0 releases/v4.1.0 releases/v4.0.2 releases/v4.0.1 releases/v4.0.0 releases/v3.2.2 releases/v3.2.1

... (truncated)

Commits

• 48fc2d4 Merge pull request #3633 from bdarnell/curl-reset-65
• 4ae1ddd Release notes and version bump for 6.5.7
• 3154caa curl_httpclient: Reset the curl object before putting it on the freelist
• 7d869c0 Merge pull request #3631 from bdarnell/cve-links
• 288241f docs: Use the correct link syntax
• 8da981c docs: Add CVE links to 6.5.6 release notes
• aba2569 Merge pull request #3626 from bdarnell/fixes-656
• a24b260 httpclient_test: Accept an additional error message variant
• a74240a Release notes and version bump for 6.5.6.
• e8fc7ed simple_httpclient: Strip auth headers on cross-origin redirects
• Additional commits viewable in compare view

Updates transformers from 4.51.3 to 5.5.0

Release notes

Sourced from transformers's releases.

Release v5.5.0

New Model additions

Gemma4

Gemma 4 is a multimodal model with pretrained and instruction-tuned variants, available in 1B, 13B, and 27B parameters. The architecture is mostly the same as the previous Gemma versions. The key differences are a vision processor that can output images of fixed token budget and a spatial 2D RoPE to encode vision-specific information across height and width axis.

You can find all the original Gemma 4 checkpoints under the Gemma 4 release.

The key difference from previous Gemma releases is the new design to process images of different sizes using a fixed-budget number of tokens. Unlike many models that squash every image into a fixed square (like 224×224), Gemma 4 keeps the image's natural aspect ratio while making it the right size. There a a couple constraints to follow:

• The total number of pixels must fit within a patch budget
• Both height and width must be divisible by 48 (= patch size 16 × pooling kernel 3)

[!IMPORTANT] Gemma 4 does not apply the standard ImageNet mean/std normalization that many other vision models use. The model's own patch embedding layer handles the final scaling internally (shifting values to the [-1, 1] range).

The number of "soft tokens" (aka vision tokens) an image processor can produce is configurable. The supported options are outlined below and the default is 280 soft tokens per image.

Soft Tokens	Patches (before pooling)	Approx. Image Area
70	630	~161K pixels
140	1,260	~323K pixels
280	2,520	~645K pixels
560	5,040	~1.3M pixels
1,120	10,080	~2.6M pixels

To encode positional information for each patch in the image, Gemma 4 uses a learned 2D position embedding table. The position table stores up to 10,240 positions per axis, which allows the model to handle very large images. Each position is a learned vector of the same dimensions as the patch embedding. The 2D RoPE which Gemma 4 uses independently rotate half the attention head dimensions for the x-axis and the other half for the y-axis. This allows the model to understand spatial relationships like "above," "below," "left of," and "right of."

NomicBERT

NomicBERT is a BERT-inspired encoder model that applies Rotary Position Embeddings (RoPE) to create reproducible long context text embeddings. It is the first fully reproducible, open-source text embedding model with 8192 context length that outperforms both OpenAI Ada-002 and OpenAI text-embedding-3-small on short-context MTEB and long context LoCo benchmarks. The model generates dense vector embeddings for various tasks including search, clustering, and classification using specific instruction prefixes.

Links: Documentation | Paper

• Internalise the NomicBERT model (#43067) by @​ed22699 in #43067

MusicFlamingo

Music Flamingo is a fully open large audio–language model designed for robust understanding and reasoning over music. It builds upon the Audio Flamingo 3 architecture by including Rotary Time Embeddings (RoTE), which injects temporal position information to enable the model to handle audio sequences up to 20 minutes. The model features a unified audio encoder across speech, sound, and music with special sound boundary tokens for improved audio sequence modeling.

Links: Documentation | Paper

• Add Music Flamingo (#43538) by @​lashahub in #43538

... (truncated)

Commits

• c1c3424 update
• 20bff68 update release workflow
• 8956441 v5.5.0
• 5135e5e casually dropping the most capable open weights on the planet (#45192)
• a594e09 Internalise the NomicBERT model (#43067)
• 4932e97 Fix resized LM head weights being overwritten by post_init (#45079)
• 57e8413 [Qwen3.5 MoE] Add _tp_plan to ForConditionalGeneration (#45124)
• b10552e Fix TypeError: 'NoneType' object is not iterable in GenerationMixin.generate ...
• 423f2a3 fix(models): Fix dtype mismatch in SwitchTransformers and TimmWrapperModel (#...
• ade7a05 Generalize gemma vision mask to videos (#45185)
• Additional commits viewable in compare view

Updates transformers from 4.51.3 to 5.5.0

Release notes

Sourced from transformers's releases.

Release v5.5.0

New Model additions

Gemma4

Gemma 4 is a multimodal model with pretrained and instruction-tuned variants, available in 1B, 13B, and 27B parameters. The architecture is mostly the same as the previous Gemma versions. The key differences are a vision processor that can output images of fixed token budget and a spatial 2D RoPE to encode vision-specific information across height and width axis.

You can find all the original Gemma 4 checkpoints under the Gemma 4 release.

The key difference from previous Gemma releases is the new design to process images of different sizes using a fixed-budget number of tokens. Unlike many models that squash every image into a fixed square (like 224×224), Gemma 4 keeps the image's natural aspect ratio while making it the right size. There a a couple constraints to follow:

• The total number of pixels must fit within a patch budget
• Both height and width must be divisible by 48 (= patch size 16 × pooling kernel 3)

[!IMPORTANT] Gemma 4 does not apply the standard ImageNet mean/std normalization that many other vision models use. The model's own patch embedding layer handles the final scaling internally (shifting values to the [-1, 1] range).

The number of "soft tokens" (aka vision tokens) an image processor can produce is configurable. The supported options are outlined below and the default is 280 soft tokens per image.

Soft Tokens	Patches (before pooling)	Approx. Image Area
70	630	~161K pixels
140	1,260	~323K pixels
280	2,520	~645K pixels
560	5,040	~1.3M pixels
1,120	10,080	~2.6M pixels

To encode positional information for each patch in the image, Gemma 4 uses a learned 2D position embedding table. The position table stores up to 10,240 positions per axis, which allows the model to handle very large images. Each position is a learned vector of the same dimensions as the patch embedding. The 2D RoPE which Gemma 4 uses independently rotate half the attention head dimensions for the x-axis and the other half for the y-axis. This allows the model to understand spatial relationships like "above," "below," "left of," and "right of."

NomicBERT

NomicBERT is a BERT-inspired encoder model that applies Rotary Position Embeddings (RoPE) to create reproducible long context text embeddings. It is the first fully reproducible, open-source text embedding model with 8192 context length that outperforms both OpenAI Ada-002 and OpenAI text-embedding-3-small on short-context MTEB and long context LoCo benchmarks. The model generates dense vector embeddings for various tasks including search, clustering, and classification using specific instruction prefixes.

Links: Documentation | Paper

• Internalise the NomicBERT model (#43067) by @​ed22699 in #43067

MusicFlamingo

Music Flamingo is a fully open large audio–language model designed for robust understanding and reasoning over music. It builds upon the Audio Flamingo 3 architecture by including Rotary Time Embeddings (RoTE), which injects temporal position information to enable the model to handle audio sequences up to 20 minutes. The model features a unified audio encoder across speech, sound, and music with special sound boundary tokens for improved audio sequence modeling.

Links: Documentation | Paper

• Add Music Flamingo (#43538) by @​lashahub in #43538

... (truncated)

Commits

• c1c3424 update
• 20bff68 update release workflow
• 8956441 v5.5.0
• 5135e5e casually dropping the most capable open weights on the planet (#45192)
• a594e09 Internalise the NomicBERT model (#43067)
• 4932e97 Fix resized LM head weights being overwritten by post_init (#45079)
• 57e8413 [Qwen3.5 MoE] Add _tp_plan to ForConditionalGeneration (#45124)
• b10552e Fix TypeError: 'NoneType' object is not iterable in GenerationMixin.generate ...
• 423f2a3 fix(models): Fix dtype mismatch in SwitchTransformers and TimmWrapperModel (#...
• ade7a05 Generalize gemma vision mask to videos (#45185)
• Additional commits viewable in compare view

Updates aiohttp from 3.14.0 to 3.14.1

Changelog

Sourced from aiohttp's changelog.

3.14.1 (2026-06-07)

Bug fixes

• Fixed a race condition in :py:class:~aiohttp.TCPConnector where closing the connector while a DNS resolution was in-flight could raise :py:exc:AttributeError instead of :py:exc:~aiohttp.ClientConnectionError -- by :user:goingforstudying-ctrl.

  Related issues and pull requests on GitHub: :issue:12497.

• Fixed CancelledError not closing a connection -- by :user:aiolibsbot.

  Related issues and pull requests on GitHub: :issue:12795.

• Tightened up some websocket parser checks -- by :user:Dreamsorcerer.

  Related issues and pull requests on GitHub: :issue:12817.

• Fixed :class:~aiohttp.CookieJar dropping the host-only flag of cookies when persisted with :meth:~aiohttp.CookieJar.save and reloaded with :meth:~aiohttp.CookieJar.load, so a cookie set without a Domain attribute is again scoped to the exact host that set it after a reload; the absolute expiration deadline is now persisted as well, so a reloaded cookie keeps its original lifetime instead of being rescheduled from the load time. :meth:~aiohttp.CookieJar.load now replaces the jar contents rather than merging onto prior state, and loaded cookies pass through the same acceptance rules as :meth:~aiohttp.CookieJar.update_cookies, so a cookie for an IP-address host is dropped when loaded into a jar created without unsafe=True -- by :user:bdraco.

  Related issues and pull requests on GitHub: :issue:12824.

• Scoped :class:~aiohttp.DigestAuthMiddleware credentials to the origin of the first request it handles, so a redirect to a different origin no longer triggers a digest response computed from the configured credentials; a challenge from another origin is only answered when that origin falls within a protection space advertised by the anchor origin through the RFC 7616 domain directive -- by :user:bdraco.

  Related issues and pull requests on GitHub: :issue:12825.

• Fixed the C HTTP parser not enforcing max_line_size on a request target or response reason phrase that is split across multiple reads; each fragment was checked on its own, so an accumulated line could exceed the limit without raising LineTooLong. The accumulated length is now checked, matching the pure-Python parser -- by :user:bdraco.

  Related issues and pull requests on GitHub:

... (truncated)

Commits

• 9c35d03 Release v3.14.1 (#12864)
• 38b956c [PR #12861/59684b5c backport][3.14] Revert "Drop list compression (#12857)" (...
• 8f31009 [PR #12857/69dff14d backport][3.14] Drop list compression (#12858)
• dfdfa9d [PR #12830/93a2b1c3 backport][3.14] Bound pipelined request queue per connect...
• 0e9cedd [PR #12827/ccf218ab backport][3.14] Numeric ipv4 resolver bypass (#12849)
• a762eda [PR #12831/1ac92dae backport][3.14] Payload close on disconnect (#12843)
• a329a7a [PR #12824/60b85e98 backport][3.14] Preserve host-only cookie scope across Co...
• 4f7480e [PR #12828/13b635d7 backport][3.14] Bounded unread compressed drain (#12845)
• 5ab61bb [PR #12826/36df6c13 backport][3.14] Enforce max_line_size on fragmented reque...
• 3912667 [3.14] Add test that env proxy auth is scoped to the redirect-selected proxy ...
• Additional commits viewable in compare view

Updates tornado from 6.5.5 to 6.5.7

Changelog

Sourced from tornado's changelog.

Release notes

.. toctree:: :maxdepth: 2

releases/v6.5.7 releases/v6.5.6 releases/v6.5.5 releases/v6.5.4 releases/v6.5.3 releases/v6.5.2 releases/v6.5.1 releases/v6.5.0 releases/v6.4.2 releases/v6.4.1 releases/v6.4.0 releases/v6.3.3 releases/v6.3.2 releases/v6.3.1 releases/v6.3.0 releases/v6.2.0 releases/v6.1.0 releases/v6.0.4 releases/v6.0.3 releases/v6.0.2 releases/v6.0.1 releases/v6.0.0 releases/v5.1.1 releases/v5.1.0 releases/v5.0.2 releases/v5.0.1 releases/v5.0.0 releases/v4.5.3 releases/v4.5.2 releases/v4.5.1 releases/v4.5.0 releases/v4.4.3 releases/v4.4.2 releases/v4.4.1 releases/v4.4.0 releases/v4.3.0 releases/v4.2.1 releases/v4.2.0 releases/v4.1.0 releases/v4.0.2 releases/v4.0.1 releases/v4.0.0 releases/v3.2.2 releases/v3.2.1

... (truncated)

Commits

• 48fc2d4 Merge pull request #3633 from bdarnell/curl-reset-65
• 4ae1ddd Release notes and version bump for 6.5.7
• 3154caa curl_httpclient: Reset the curl object before putting it on the freelist
• 7d869c0 Merge pull request #3631 from bdarnell/cve-links
• 288241f docs: Use the correct link syntax
• 8da981c docs: Add CVE links to 6.5.6 release notes
• aba2569 Merge pull request #3626 from bdarnell/fixes-656
• a24b260 httpclient_test: Accept an additional error message variant
• a74240a Release notes and version bump for 6.5.6.
• e8fc7ed simple_httpclient: Strip auth headers on cross-origin redirects
• Additional commits viewable in compare view

Updates transformers from 4.51.3 to 5.5.0

Release notes

Sourced from transformers's releases.

Release v5.5.0

New Model additions

Gemma4

Gemma 4 is a multimodal model with pretrained and instruction-tuned variants, available in 1B, 13B, and 27B parameters. The architecture is mostly the same as the previous Gemma versions. The key differences are a vision processor that can output images of fixed token budget and a spatial 2D RoPE to encode vision-specific information across height and width axis.

You can find all the original Gemma 4 checkpoints under the Gemma 4 release.

The key difference from previous Gemma releases is the new design to process images of different sizes using a fixed-budget number of tokens. Unlike many models that squash every image into a fixed square (like 224×224), Gemma 4 keeps the image's natural aspect ratio while making it the right size. There a a couple constraints to follow:

• The total number of pixels must fit within a patch budget
• Both height and width must be divisible by 48 (= patch size 16 × pooling kernel 3)

[!IMPORTANT] Gemma 4 does not apply the standard ImageNet mean/std normalization that many other vision models use. The model's own patch embedding layer handles the final scaling internally (shifting values to the [-1, 1] range).

The number of "soft tokens" (aka vision tokens) an image processor can produce is configurable. The supported options are outlined below and the default is 280 soft tokens per image.

Soft Tokens	Patches (before pooling)	Approx. Image Area
70	630	~161K pixels
140	1,260	~323K pixels
280	2,520	~645K pixels
560	5,040	~1.3M pixels
1,120	10,080	~2.6M pixels

To encode positional information for each patch in the image, Gemma 4 uses a learned 2D position embedding table. The position table stores up to 10,240 positions per axis, which allows the model to handle very large images. Each position is a learned vector of the same dimensions as the patch embedding. The 2D RoPE which Gemma 4 uses independently rotate half the attention head dimensions for the x-axis and the other half for the y-axis. This allows the model to understand spatial relationships like "above," "below," "left of," and "right of."

NomicBERT

NomicBERT is a BERT-inspired encoder model that applies Rotary Position Embeddings (RoPE) to create reproducible long context text embeddings. It is the first fully reproducible, open-source text embedding model with 8192 context length that outperforms both OpenAI Ada-002 and OpenAI text-embedding-3-small on short-context MTEB and long context LoCo benchmarks. The model generates dense vector embeddings for various tasks including search, clustering, and classification using specific instruction prefixes.

Links: Documentation | Paper

• Internalise the NomicBERT model (#43067) by @​ed22699 in #43067

MusicFlamingo

Music Flamingo is a fully open large audio–language model designed for robust understanding and reasoning over music. It builds upon the Audio Flamingo 3 architecture by including Rotary Time Embeddings (RoTE), which injects temporal position information to enable the model to handle audio sequences up to 20 minutes. The model features a unified audio encoder across speech, sound, and music with special sound boundary tokens for improved audio sequence modeling.

Links: Documentation | Paper

• Add Music Flamingo (#43538) by @​lashahub in #43538

... (truncated)

Commits

• c1c3424 update
• 20bff68 update release workflow
• 8956441 v5.5.0
• 5135e5e casually dropping the most capable open weights on the planet (#45192)
• a594e09 Internalise the NomicBERT model (#43067)
• 4932e97 Fix resized LM head weights being overwritten by post_init (#45079)
• 57e8413 [Qwen3.5 MoE] Add _tp_plan to ForConditionalGeneration (#45124)
• b10552e Fix TypeError: 'NoneType' object is not iterable in GenerationMixin.generate ...
• 423f2a3 fix(models): Fix dtype mismatch in SwitchTransformers and TimmWrapperModel (#...
• ade7a05 Generalize gemma vision mask to videos (#45185)
• Additional commits viewable in compare view

Updates transformers from 4.52.4 to 5.5.0

Release notes

Sourced from transformers's releases.

Release v5.5.0

New Model additions

Gemma4

Gemma 4 is a multimodal model with pretrained and instruction-tuned variants, available in 1B, 13B, and 27B parameters. The architecture is mostly the same as the previous Gemma versions. The key differences are a vision processor that can output images of fixed token budget and a spatial 2D RoPE to encode vision-specific information across height and width axis.

You can find all the original Gemma 4 checkpoints under the Gemma 4 release.

The key difference from previous Gemma releases is the new design to process images of different sizes using a fixed-budget number of tokens. Unlike many models that squash every image into a fixed square (like 224×224), Gemma 4 keeps the image's natural aspect ratio while making it the right size. There a a couple constraints to follow:

• The total number of pixels must fit within a patch budget
• Both height and width must be divisible by 48 (= patch size 16 × pooling kernel 3)

[!IMPORTANT] Gemma 4 does not apply the standard ImageNet mean/std normalization that many other vision models use. The model's own patch embedding layer handles the final scaling internally (shifting values to the [-1, 1] range).

The number of "soft tokens" (aka vision tokens) an image processor can produce is configurable. The supported options are outlined below and the default is 280 soft tokens per image.

Soft Tokens	Patches (before pooling)	Approx. Image Area
70	630	~161K pixels
140	1,260	~323K pixels
280	2,520	~645K pixels
560	5,040	~1.3M pixels
1,120	10,080	~2.6M pixels

To encode positional information for each patch in the image, Gemma 4 uses a learned 2D position embedding table. The position table stores up to 10,240 positions per axis, which allows the model to handle very large images. Each position is a learned vector of the same dimensions as the patch embedding. The 2D RoPE which Gemma 4 uses independently rotate half the attention head dimensions for the x-axis and the other half for the y-axis. This allows the model to understand spatial relationships like "above," "below," "left of," and "right of."

NomicBERT

NomicBERT is a BERT-inspired encoder model that applies Rotary Position Embeddings (RoPE) to create reproducible long context text embeddings. It is the first fully reproducible, open-source text embedding model with 8192 context length that outperforms both OpenAI Ada-002 and OpenAI text-embedding-3-small on short-context MTEB and long context LoCo benchmarks. The model generates dense vector embeddings for various tasks including search, clustering, and classification using specific instruction prefixes.

Links: Documentation | Paper

• Internalise the NomicBERT model (#43067) by @​ed22699 in #43067

MusicFlamingo

Music Flamingo is a fully open large audio–language model designed for robust understanding and reasoning over music. It builds upon the Audio Flamingo 3 architecture by including Rotary Time Embeddings (RoTE), which injects temporal position information to enable the model to handle audio sequences up to 20 minutes. The model features a unified audio encoder across speech, sound, and music with special sound boundary tokens for improved audio sequence modeling.

Links: Documentation | Paper

• Add Music Flamingo (#43538) by @​lashahub in #43538

... (truncated)

Commits

• c1c3424 update
• 20bff68 update release workflow
• 8956441 v5.5.0
• 5135e5e casually dropping the most capable open weights on the planet (#45192)
• a594e09 Internalise the NomicBERT model (#43067)
• 4932e97 Fix resized LM head weights being overwritten by post_init (#45079)
• 57e8413 [Qwen3.5 MoE] Add _tp_plan to ForConditionalGeneration (#45124)
• b10552e Fix TypeError: 'NoneType' object is not iterable in GenerationMixin.generate ...
• 423f2a3 fix(models): Fix dtype mismatch in SwitchTransformers and TimmWrapperModel (#...
• ade7a05 Generalize gemma vision mask to videos (#45185)
• Additional commits viewable in compare view

Updates transformers from 4.52.4 to 5.5.0

Release notes

Sourced from transformers's releases.

Release v5.5.0

New Model additions

Gemma4

Gemma 4 is a multimodal model with pretrained and instruction-tuned variants, available in 1B, 13B, and 27B parameters. The architecture is mostly the same as the previous Gemma versions. The key differences are a vision processor that can output images of fixed token budget and a spatial 2D RoPE to encode vision-specific information across height and width axis.

You can find all the original Gemma 4 checkpoints under the Gemma 4 release.

The key difference from previous Gemma releases is the new design to process images of different sizes using a fixed-budget number of tokens. Unlike many models that squash every image into a fixed square (like 224×224), Gemma 4 keeps the image's natural aspect ratio while making it the right size. There a a couple constraints to follow:

• The total number of pixels must fit within a patch budget
• Both height and width must be divisible by 48 (= patch size 16 × pooling kernel 3)

[!IMPORTANT] Gemma 4 does not apply the standard ImageNet mean/std normalization that many other vision models use. The model's own patch embedding layer handles the final scaling internally (shifting values to the [-1, 1] range).

The number of "soft tokens" (aka vision tokens) an image processor can produce is configurable. The supported options are outlined below and the default is 280 soft tokens per image.

Soft Tokens	Patches (before pooling)	Approx. Image Area
70	630	~161K pixels
140	1,260	~323K pixels
280	2,520	~645K pixels
560	5,040	~1.3M pixels
1,120	10,080	~2.6M pixels

To encode positional information for each patch in the image, Gemma 4 uses a learned 2D position embedding table. The position table stores up to 10,240 positions per axis, which allows the model to handle very large images. Each position is a learned vector of the same dimensions as the patch embedding. The 2D RoPE which Gemma 4 uses independently rotate half the attention head dimensions for the x-axis and the other half for the y-axis. This allows the model to understand spatial relationships like "above," "below," "left of," and "right of."

NomicBERT

NomicBERT is a BERT-inspired encoder model that applies Rotary Position Embeddings (RoPE) to create reproducible long context text embeddings. It is the first fully reproducible, open-source text embedding model with 8192 context length that outperforms both OpenAI Ada-002 and OpenAI text-embedding-3-small on short-context MTEB and long context LoCo benchmarks. The model generates dense vector embeddings for various tasks including search, clustering, and classification using specific instruction prefixes.

Links: Documentation | Paper

• Internalise the NomicBERT model (#43067) by @​ed22699 in #43067

MusicFlamingo

Music Flamingo is a fully open large audio–language model designed for robust understanding and reasoning over music. It builds upon the Audio Flamingo 3 architecture by including Rotary Time Embeddings (RoTE), which injects temporal position information to enable the model to handle audio sequences up to 20 minutes. The model features a unified audio encoder across speech, sound, and music with special sound boundary tokens for improved audio sequence modeling.

Links: Documentation | Paper

• Add Music Flamingo (#43538) by @​lashahub in #43538

... (truncated)

Commits

• c1c3424 update
• Description has been truncated

[Copilot]

Copilot can't review bot-authored pull requests automatically. A user with Copilot access can request a review manually.