Bump the pip group across 10 directories with 3 updates

[dependabot]

⚠️ Dependabot is rebasing this PR ⚠️

Rebasing might not happen immediately, so don't worry if this takes some time.

Note: if you make any changes to this PR yourself, they will take precedence over the rebase.

---

Bumps the pip group with 2 updates in the /examples/pytorch/diffusion_model/diffusers/flux directory: transformers and torch.
Bumps the pip group with 1 update in the /examples/pytorch/diffusion_model/diffusers/framepack directory: transformers.
Bumps the pip group with 1 update in the /examples/pytorch/diffusion_model/diffusers/wan directory: gdown.
Bumps the pip group with 1 update in the /examples/pytorch/nlp/huggingface_models/language-modeling/quantization/auto_round/deepseek directory: transformers.
Bumps the pip group with 1 update in the /examples/pytorch/nlp/huggingface_models/language-modeling/quantization/auto_round/llama3 directory: transformers.
Bumps the pip group with 1 update in the /examples/pytorch/nlp/huggingface_models/language-modeling/quantization/llm_qat directory: transformers.
Bumps the pip group with 1 update in the /examples/pytorch/nlp/huggingface_models/language-modeling/quantization/smooth_quant directory: torch.
Bumps the pip group with 1 update in the /examples/pytorch/nlp/huggingface_models/language-modeling/quantization/static_quant/pt2e directory: transformers.
Bumps the pip group with 1 update in the /examples/pytorch/nlp/huggingface_models/language-modeling/quantization/transformers/weight_only/text-generation directory: torch.
Bumps the pip group with 1 update in the /examples/tensorflow/nlp/large_language_models/quantization/ptq/smoothquant directory: transformers.

Updates transformers from 4.57.3 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 torch from 2.10.0 to 2.13.0

Release notes

Sourced from torch's releases.

PyTorch 2.13.0 Release Notes

• Highlights
• Backwards Incompatible Changes
• Deprecations
• New Features
• Improvements
• Bug fixes
• Performance
• Documentation
• Developers

Highlights

For more details about these highlighted features, you can look at the release blogpost. Below are the full release notes for this release.

Tracked Regressions

ROCm wheels break torch.compile on CPU in environments without a GPU

Running a torch==2.13.0+rocm7.2 wheel in an environment where no GPU is available (torch.cuda.is_available() is False) breaks torch.compile on the CPU path: the first compile raises RuntimeError: Can't detect vectorized ISA for CPU (#189194). This is a regression from torch==2.12.1+rocm7.2, which compiles CPU code fine (detecting e.g. VecAVX2) in the same setup. The 2.13 ROCm wheel appears to rely on something present in the ROCm builder image to detect the CPU vectorized ISA, so it works when run on a ROCm image but fails on a plain CPU-only image.

Workaround: run the +rocm wheel on a ROCm image, or install a standard CPU/CUDA build for GPU-less environments.

Backwards Incompatible Changes

• Stop building CPython 3.13t (free-threaded) binaries (#182951)

  Upstream pypa/manylinux removed CPython 3.13t (free-threaded) on 2026-05-07, because 3.13t was experimental and has been superseded by the now-non-experimental CPython 3.14t. As a result, PyTorch 2.13 no longer ships cp313t wheels (Linux, Triton, and related artifacts). Users on the free-threaded interpreter should move to Python 3.14t.

  PyTorch 2.12:

  # cp313t (free-threaded 3.13) wheels were available
  python3.13t -m pip install torch

  PyTorch 2.13:

... (truncated)

Commits

• cf30153 [release/2.13] Strip +PTX from CUDA arch list on release/RC builds (#188914) ...
• 3e3e24b [release/2.13] Restrict cuda-bindings to Python < 3.15 for CUDA 12.9 builds (...
• 7986b06 [release/2.13] Bump binary build timeout 280 -> 400 minutes (#188551)
• 0bdbc26 [release/2.13] Add CUDA 12.9 to TORCH_CUDA_ARCH_LIST tables (#188443)
• 9cabb45 [release/2.13] Update manywheel docker image pin to 78e737ad (#188409)
• 78e737a [release/2.13] Revert "Tighten generalized scatter graph target (#184075)" (#...
• 0bb9b5b [release/2.13] Revert "dynamo: round-trip torch.cuda.stream ctx mgr across gr...
• aaac2bf [release/2.13] Revert "[Reland] Port D104346887/PR 182675 for index_add fast ...
• 9330813 Fix build_with_debinfo.py broken by CONFIGURE_DEPENDS globbing (#188192)
• 4e077a7 Remove setuptools upper bound (#188190)
• Additional commits viewable in compare view

Updates transformers from 4.57.3 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 torch from 2.10.0 to 2.13.0

Release notes

Sourced from torch's releases.

PyTorch 2.13.0 Release Notes

• Highlights
• Backwards Incompatible Changes
• Deprecations
• New Features
• Improvements
• Bug fixes
• Performance
• Documentation
• Developers

Highlights

For more details about these highlighted features, you can look at the release blogpost. Below are the full release notes for this release.

Tracked Regressions

ROCm wheels break torch.compile on CPU in environments without a GPU

Running a torch==2.13.0+rocm7.2 wheel in an environment where no GPU is available (torch.cuda.is_available() is False) breaks torch.compile on the CPU path: the first compile raises RuntimeError: Can't detect vectorized ISA for CPU (#189194). This is a regression from torch==2.12.1+rocm7.2, which compiles CPU code fine (detecting e.g. VecAVX2) in the same setup. The 2.13 ROCm wheel appears to rely on something present in the ROCm builder image to detect the CPU vectorized ISA, so it works when run on a ROCm image but fails on a plain CPU-only image.

Workaround: run the +rocm wheel on a ROCm image, or install a standard CPU/CUDA build for GPU-less environments.

Backwards Incompatible Changes

• Stop building CPython 3.13t (free-threaded) binaries (#182951)

  Upstream pypa/manylinux removed CPython 3.13t (free-threaded) on 2026-05-07, because 3.13t was experimental and has been superseded by the now-non-experimental CPython 3.14t. As a result, PyTorch 2.13 no longer ships cp313t wheels (Linux, Triton, and related artifacts). Users on the free-threaded interpreter should move to Python 3.14t.

  PyTorch 2.12:

  # cp313t (free-threaded 3.13) wheels were available
  python3.13t -m pip install torch

  PyTorch 2.13:

... (truncated)

Commits

• cf30153 [release/2.13] Strip +PTX from CUDA arch list on release/RC builds (#188914) ...
• 3e3e24b [release/2.13] Restrict cuda-bindings to Python < 3.15 for CUDA 12.9 builds (...
• 7986b06 [release/2.13] Bump binary build timeout 280 -> 400 minutes (#188551)
• 0bdbc26 [release/2.13] Add CUDA 12.9 to TORCH_CUDA_ARCH_LIST tables (#188443)
• 9cabb45 [release/2.13] Update manywheel docker image pin to 78e737ad (#188409)
• 78e737a [release/2.13] Revert "Tighten generalized scatter graph target (#184075)" (#...
• 0bb9b5b [release/2.13] Revert "dynamo: round-trip torch.cuda.stream ctx mgr across gr...
• aaac2bf [release/2.13] Revert "[Reland] Port D104346887/PR 182675 for index_add fast ...
• 9330813 Fix build_with_debinfo.py broken by CONFIGURE_DEPENDS globbing (#188192)
• 4e077a7 Remove setuptools upper bound (#188190)
• Additional commits viewable in compare view

Updates transformers from 4.57.3 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 4.57.3 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 4.57.3 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 4.57.3 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 4.57.3 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) b...

  Description has been truncated