Remove stale and redundant _no_split_modules entries by guarin · Pull Request #47645 · huggingface/transformers

guarin · 2026-07-30T12:12:58Z

This PR removes or updates _no_split_modules entires with the addition of a new mlinter rule: huggingface/transformers-mlinter#12

The new rule verifies that _no_split_modules contains only entries from modules defined in modeling_*.py, modular_*.py or a file in the same model directory.

This helps avoiding wrong _no_split_modules entries. It also simplifies reviews.

Updates can be grouped into these cases:

remove unknown class (e.g. Idefics2DecoderLayer)
fix typo (e.g. Data2VecTextForTextEmbeddings -> Data2VecTextEmbeddings)
remove classes already in _no_split_modules of submodule (e.g. T5Block in blip_2)
empty _no_split_modules to avoid inheriting entry for parent class (e.g. in DeepseekVLPreTrainedModel to avoid inheriting JanusVisionEncoderLayer)

There are some special cases that are automatically handled by the linter:

Parametrized* classes created through torch.nn.utils.parametrize are ignored as they are only created at runtime
TimmWrapperForImageClassification is ignored

Note

There are still some blind spots for the linter. As the linter only checks non-autogenerated files, it can miss cases where a modular file doesn't define _no_split_modules and the generated modeling.py file inherits a wrong name from the parent class. For example AriaPreTrainedModel doesn't have _no_split_modules:

transformers/src/transformers/models/aria/modular_aria.py

Lines 791 to 795 in 71c6f69

    
           class AriaPreTrainedModel(LlamaPreTrainedModel): 
        
               config: AriaConfig 
        
               base_model_prefix = "model" 
        
               _can_compile_fullgraph = False  # MoE models don't work with torch.compile (dynamic slicing) 
        
               _supports_attention_backend = True

which gets automatically extended to this in modeling_aria.py:

transformers/src/transformers/models/aria/modeling_aria.py

Lines 597 to 601 in 71c6f69

    
           class AriaPreTrainedModel(PreTrainedModel): 
        
               config: AriaConfig 
        
               base_model_prefix = "model" 
        
               supports_gradient_checkpointing = True 
        
               _no_split_modules = ["AriaDecoderLayer"]

where AriaDecoderLayer is automatically renamed from LlamaDecoderLayer but there is no AriaDecoderLayer.

Code Agent Policy

The Transformers repo is currently being overwhelmed by a large number of PRs and issue comments written by
code agents. These often are low-quality, or fix extremely minor issues that occur rarely or never in practice.
As a result, we're instituting a rule that first-time contributors should not use code agents to submit PRs or issues.
We'd also ask autonomous "OpenClaw"-like agents not to open any PRs or issues.

Issues/PRs from first-time contributors that violate this rule will probably just be closed without review, and we
might block you, especially if you open more than one or appear to be deliberately ignoring this. We especially do not
want new contributors to jump in on random issues to contribute an agent-written fix. This creates lots of noise
for reviewers and other users and will almost certainly get you blocked.

For more information, please read CONTRIBUTING.md.

(First-time contributors only): I confirm that this PR description and code is not written by an LLM or code agent

Before submitting

This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
Did you read the contributor guideline and the
Pull Request checks?
Was this discussed/approved via a Github issue or the forum? Please add a link
to it if that's the case.
Did you make sure to update the documentation with your changes according to the guidelines?
Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

Some models list `_no_split_modules` entries naming classes that do not exist anywhere in transformers, so `device_map` never matches them and the modules they were meant to keep together get split across devices: - idefics2: Idefics2DecoderLayer - idefics3: Idefics3DecoderLayer - video_llava: VideoLlavaVisionAttention Others hardcode classes owned by a submodel that is loaded through the Auto* classes. `post_init` already collects `_no_split_modules` from child submodels, so these entries are redundant, and in blip_2's case actively wrong: it declared both T5Block and OPTDecoderLayer regardless of which language model the checkpoint actually uses. - blip_2: T5Block, OPTDecoderLayer - depth_anything: DPTViTEmbeddings (the backbone is an AutoBackbone, so this never matched; it now correctly reports the backbone's Dinov2Layer) - llava_next: LlamaDecoderLayer - moshi: MimiTransformerLayer The resulting runtime sets are more accurate than the hardcoded lists were. Generated modeling files updated with the modular converter, which propagates the removals to smolvlm, llava_next_video, llava_onevision, kyutai_speech_to_text and prompt_depth_anything. deepseek_ocr2 and granite4_vision only shift the position of their own override.

guarin · 2026-07-30T12:22:56Z

run-slow: blip_2, deepseek_ocr2, depth_anything, granite4_vision, idefics2, idefics3, kyutai_speech_to_text, llava_next, llava_next_video, llava_onevision, moshi, prompt_depth_anything, smolvlm, video_llava

github-actions · 2026-07-30T12:24:22Z

Workflow Run ⚙️

This comment contains run-slow, running the specified jobs:

models: ["models/blip_2", "models/deepseek_ocr2", "models/depth_anything", "models/granite4_vision", "models/idefics2", "models/idefics3", "models/kyutai_speech_to_text", "models/llava_next", "models/llava_next_video", "models/llava_onevision", "models/moshi", "models/prompt_depth_anything", "models/smolvlm", "models/video_llava"]
quantizations: []

HuggingFaceDocBuilderDev · 2026-07-30T12:30:02Z

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

github-actions · 2026-07-30T12:56:17Z

CI Results

Workflow Run ⚙️

Commit Info

Context	Commit	Description
RUN	3e618b40	workflow commit (merge commit)
PR	35a211fc	branch commit (from PR)
main	560f36cc	base commit (on `main`)

✅ No failing test specific to this PR 🎉 👏 !

…n/transformers into fix-stale-no-split-modules

molbap

🧹 🧹
Nice and clean! Happy for the linter too 🎉

molbap · 2026-08-03T11:27:29Z

I think above this one, cohere_asr also has a CohereEncoderLayer that does not exist, but maybe more are like this, your linter rule will help!

I think this is the loophole that still exists, see Note in the PR description

Hmm, I can't really follow tbh. Shouldn't the linter follow the modeling files? Or is there an conflict of order of applications re fix/check repo?

ah indeed, same as you described, I think it's fixable in modular_model_converter (and should be): we could either include it as a regex str match in convert_modular_file , just would need re.findall(r"_no_split_modules = \ somewhere and would not cost much

Hmm, I can't really follow tbh. Shouldn't the linter follow the modeling files? Or is there an conflict of order of applications re fix/check repo?

The issue is that the linter doesn't check modeling files that are generated. Agree with @molbap that this should be fixed in modular converter.

The issue is that the linter doesn't check modeling files that are generated

That kind of seems like a self contained issue in itself? I would expect modeling to be checked as well. But yea anyways, agree that it could be done at modular level as well (I just find it a bit less intuitive)

Yes I was surprised to learn that the linter doesn't check the auto-generated files. The motivation seems to be that errors should be reported in the modular files instead. But in this case here it would definitely also make sense to check the modeling file. Will ask Tarek once he is back.

vasqu

Some small questions but nothing major, thanks 🫡 maybe also run slow ci for some whom that should fix?

vasqu · 2026-08-03T12:08:31Z

Hmm, I can't really follow tbh. Shouldn't the linter follow the modeling files? Or is there an conflict of order of applications re fix/check repo?

guarin · 2026-08-03T13:49:04Z

run-slow: blip_2, data2vec, deepseek_ocr2, deepseek_vl, deepseek_vl_hybrid, depth_anything, diffusion_gemma, exaone4_5, gemma3, granite4_vision, higgs_audio_v2_tokenizer, idefics2, idefics3, janus, kyutai_speech_to_text, llava_next

github-actions · 2026-08-03T13:49:37Z

[For maintainers] Suggested jobs to run (before merge)

run-slow: blip_2, dac, data2vec, deepseek_ocr2, deepseek_vl, deepseek_vl_hybrid, depth_anything, diffusion_gemma, exaone4_5, gemma3, granite4_vision, higgs_audio_v2_tokenizer, idefics2, idefics3, janus, kyutai_speech_to_text

github-actions · 2026-08-03T13:50:41Z

Workflow Run ⚙️

This comment contains run-slow, running the specified jobs:

models: ["models/blip_2", "models/data2vec", "models/deepseek_ocr2", "models/deepseek_vl", "models/deepseek_vl_hybrid", "models/depth_anything", "models/diffusion_gemma", "models/exaone4_5", "models/gemma3", "models/granite4_vision", "models/higgs_audio_v2_tokenizer", "models/idefics2", "models/idefics3", "models/janus", "models/kyutai_speech_to_text", "models/llava_next"]
quantizations: []

github-actions · 2026-08-03T14:04:09Z

CI recap

Dashboard: View test results in Grafana
Latest run: 30819654874:1
Result: failure | Jobs: 16 | Tests: 148,999 | Failures: 1 | Duration: 15h 39m

github-actions · 2026-08-03T14:24:27Z

CI Results

Workflow Run ⚙️

Commit Info

Context	Commit	Description
RUN	febffdb3	workflow commit (merge commit)
PR	59114f4e	branch commit (from PR)
main	583f1942	base commit (on `main`)

✅ No failing test specific to this PR 🎉 👏 !

TRF005 only validates the shape of `_no_split_modules`. TRF022 additionally checks that each entry names a class that actually exists, resolving it against the classes defined in the modeling file, the names it imports, and the classes defined by sibling modules of the same model directory. `device_map` matches these strings against `module.__class__.__name__` at runtime, so a stale or misspelled entry is silently ignored. Entries naming another model's classes are flagged too: `post_init` already collects `_no_split_modules` from child submodels, so hardcoding them is redundant. Corresponding transformers PR to make checks pass: huggingface/transformers#47645 --------- Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>

guarin mentioned this pull request Jul 30, 2026

Add TRF022: verify _no_split_modules entries name existing classes huggingface/transformers-mlinter#12

Merged

Merge branch 'main' into fix-stale-no-split-modules

35a211f

guarin marked this pull request as draft July 30, 2026 13:22

guarin added 2 commits July 30, 2026 16:09

update modular _no_split_modules

773c087

Merge branch 'fix-stale-no-split-modules' of https://github.com/guari…

ac957e8

…n/transformers into fix-stale-no-split-modules

guarin commented Jul 30, 2026

View reviewed changes

Comment thread src/transformers/models/higgs_audio_v2_tokenizer/modeling_higgs_audio_v2_tokenizer.py Outdated

guarin marked this pull request as ready for review July 30, 2026 15:56

molbap approved these changes Aug 3, 2026

View reviewed changes

vasqu approved these changes Aug 3, 2026

View reviewed changes

guarin added 3 commits August 3, 2026 13:27

switch to attributeerror

0eb2adc

fix dac

8bb79d7

Merge branch 'main' into fix-stale-no-split-modules

59114f4

guarin added this pull request to the merge queue Aug 3, 2026

github-merge-queue Bot removed this pull request from the merge queue due to failed status checks Aug 3, 2026

guarin added this pull request to the merge queue Aug 3, 2026

Merged via the queue into huggingface:main with commit d09f53a Aug 3, 2026
212 of 214 checks passed

guarin deleted the fix-stale-no-split-modules branch August 3, 2026 16:31

guarin mentioned this pull request Aug 4, 2026

Fix Dac offload tests #47775

Merged

	class AriaPreTrainedModel(LlamaPreTrainedModel):
	config: AriaConfig
	base_model_prefix = "model"
	_can_compile_fullgraph = False # MoE models don't work with torch.compile (dynamic slicing)
	_supports_attention_backend = True

	class AriaPreTrainedModel(PreTrainedModel):
	config: AriaConfig
	base_model_prefix = "model"
	supports_gradient_checkpointing = True
	_no_split_modules = ["AriaDecoderLayer"]

Uh oh!

Conversation

guarin commented Jul 30, 2026 • edited by github-actions Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Note

Code Agent Policy

Before submitting

Who can review?

Uh oh!

guarin commented Jul 30, 2026

Uh oh!

github-actions Bot commented Jul 30, 2026

Uh oh!

HuggingFaceDocBuilderDev commented Jul 30, 2026

Uh oh!

github-actions Bot commented Jul 30, 2026

CI Results

Commit Info

Uh oh!

Uh oh!

molbap left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

molbap Aug 3, 2026

Choose a reason for hiding this comment

Uh oh!

guarin Aug 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

vasqu Aug 3, 2026

Choose a reason for hiding this comment

Uh oh!

molbap Aug 3, 2026

Choose a reason for hiding this comment

Uh oh!

guarin Aug 3, 2026

Choose a reason for hiding this comment

Uh oh!

vasqu Aug 3, 2026

Choose a reason for hiding this comment

Uh oh!

guarin Aug 3, 2026

Choose a reason for hiding this comment

Uh oh!

vasqu left a comment

Choose a reason for hiding this comment

Uh oh!

vasqu Aug 3, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

guarin commented Aug 3, 2026

Uh oh!

github-actions Bot commented Aug 3, 2026

Uh oh!

github-actions Bot commented Aug 3, 2026

Uh oh!

github-actions Bot commented Aug 3, 2026

CI recap

Uh oh!

github-actions Bot commented Aug 3, 2026

CI Results

Commit Info

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

guarin commented Jul 30, 2026 •

edited by github-actions Bot

Loading

guarin Aug 3, 2026 •

edited

Loading