[MoE] MoE Calibration with calibrate_all_experts
#1760
Draft
Conversation
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
Change Purpose:
- Add calibrate_all_experts option to improve MoE calibration
Change Details:
- Add `calibrate_all_experts` flag to MoE layers
- Update `replace_modules_for_calibration` and `moe_calibration_context`
to propagate the flag into modules
- Modify expert forward passes:
* Normal mode (default): compute output only for tokens routed to
top-k experts, and combine their weighted results in the final
output
* Calibration mode (`calibrate_all_experts=True`): compute output for
all tokens on every expert, but still apply the top-k gating to
decide which token outputs contribute to the final result.
Testing:
- Add unit test to verify all experts are triggered during MoE calibration
Signed-off-by: Kyle Sayers <[email protected]>
|
👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review. Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed. |
kylesayrs
changed the title
dsikka
reviewed
Aug 19, 2025
Signed-off-by: Kyle Sayers <[email protected]>
kylesayrs
changed the title
calibrate_all_experts
Signed-off-by: Kyle Sayers <[email protected]>
Signed-off-by: Kyle Sayers <[email protected]>
Signed-off-by: Kyle Sayers <[email protected]>
Signed-off-by: Kyle Sayers <[email protected]>
fynnsu
reviewed
Aug 29, 2025
There was a problem hiding this comment.
Left a comment below. I also agree with @brian-dellabetta's point that this could maybe be simplified by patching self.top_k temporarily.
Signed-off-by: Kyle Sayers <[email protected]>
dsikka
reviewed
Sep 2, 2025
|
Running those examples now |
brian-dellabetta
approved these changes
Sep 9, 2025
| @@ -974,7 +974,8 @@ def getattr_chain(obj: Any, chain_str: str, *args, **kwargs) -> Any: | |||
| return res | |||
|
|
|||
|
|
|||
| class DisableKVCache: | |||
| @contextlib.contextmanager | |||
| def disable_cache(module: torch.nn.Module): | |||
There was a problem hiding this comment.
definitely agree with these changes, but might be better in a separate PR or added in the PR summary. Seems orthogonal to calibrate_all_experts
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Coauthored with @dichn!
Purpose
calibrate_all_expertsoption, which sends all tokens to all experts, but still produces the same outputs as if tokens had been gatedChanges
calibrate_all_experts=Truetoken gating occurs after passing tokens to experts, rather than beforeTesting