🚨 [Kernels] Fix kernel function registration#45420
Merged
vasqu merged 5 commits intohuggingface:mainfrom Apr 20, 2026
Merged
Conversation
|
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. |
Cyrilvallez
reviewed
Apr 17, 2026
Member
Cyrilvallez
left a comment
Cyrilvallez
left a comment
There was a problem hiding this comment.
Nice! Mostly nits to avoid relying on internals too much!
vasqu
commented
Apr 17, 2026
Comment on lines
+228
to
+230
| Mode.TRAINING: FuncRepository( | ||
| repo_id="kernels-community/rotary", func_name="apply_rotary_transformers" | ||
| ), |
Contributor
Author
There was a problem hiding this comment.
Not sure why this wasnt included for training before but it runs with deepspeed just fine + https://github.com/Dao-AILab/flash-attention/blob/b65ae6b175f2438de55601695b6a21971fc5e429/flash_attn/layers/rotary.py#L38-L90
vasqu
commented
Apr 17, 2026
Comment on lines
+4476
to
+4483
| def attach_hidden_kernels(module): | ||
| for name, fn in getattr(module, "_hidden_kernels", {}).items(): | ||
| if name not in dict(module.named_children()): | ||
| module.register_module(name, fn) | ||
|
|
||
| def detach_hidden_kernels(module): | ||
| for name in getattr(module, "_hidden_kernels", {}): | ||
| delattr(module, name) |
Contributor
Author
There was a problem hiding this comment.
Removed the internals structure and rely on native APIs instead as suggested
vasqu
changed the title
Kernels] Fix kernel function registrationKernels] Fix kernel function registration
Cyrilvallez
approved these changes
Apr 20, 2026
Comment on lines
+4485
to
+4486
| self.apply(attach_hidden_kernels) | ||
| try: |
Member
There was a problem hiding this comment.
Would put the apply inside the try as well, but not a big deal at all!
lvliang-intel
pushed a commit
to lvliang-intel/transformers
that referenced
this pull request
Apr 21, 2026
* fix attmpt * proper fix - also works with deepspeed * rely less on internals and add rotary to training * move under the try as well
artem-spector
pushed a commit
to artem-spector/transformers
that referenced
this pull request
Apr 21, 2026
* fix attmpt * proper fix - also works with deepspeed * rely less on internals and add rotary to training * move under the try as well
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Breaking change
🚨 Slightly breaking change: We no longer register the hidden
rotary_fn. Users shouldn't have relied on those but in any case marking it, e.g.self.rotary_fn(...)within the Attention module does not work anymore as the reference is deleted from now onDescription
As per title, we do not want to have proper
nn.Modules to be registered for kernels exchanged functions - they are not proper modules (and they are never called as such)! They act as exchange format for kernels but functionally they should stay as pure functions only.The exact reasons are numerous, but one recent example is deepspeed zero 3 which cannot handle this as the module is never properly called in the forward on the module directly (untracable) and it changes module structures after model construction (fixable by changing order of inits tbh).
This PR changes the core functionality to make the module registration temporarily under the parent module, discover the exchangable functions, and delete them from the visible interface. For BC purposes, we still keep a self reference that already exists (now as simple attribute, not module).