Use eigenvalue EMA over per-step outer products for per-factor correction#264
Open
runame wants to merge 7 commits into
Open
Use eigenvalue EMA over per-step outer products for per-factor correction#264runame wants to merge 7 commits into
runame wants to merge 7 commits into
Conversation
…rsUnwrapped base class Consolidate duplicated eigendecomposition logic from EigendecomposedShampooKroneckerFactorsUnwrapped and EigenvalueCorrectedShampooKroneckerFactorsUnwrapped into a shared base class. The base class provides _perform_eigendecomposition and _amortized_computation, with subclass behavior controlled via hasattr checks on field presence. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…or_outer_product hook Inline the outer product loop into BaseShampooPreconditionerList._update_factor_matrices and introduce _transform_grad_for_outer_product as the single extension point. The base returns grad unchanged; KL-Shampoo subclasses override it to precondition the gradient. This eliminates _compute_outer_product_list from all three classes that defined it. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Introduce PerFactorEigenvalueCorrectedShampoo, which stores m+n eigenvalues per block (one per factor dimension) computed directly as diag(Q^T M Q), where Q are cached eigenvectors and M is the already-accumulated factor matrix. This is more memory-efficient than EShampoo/SOAP's m*n eigenvalues while still providing eigenvalue correction. New classes: - PerFactorEigenvalueCorrectedShampooKroneckerFactorsUnwrapped - PerFactorEigenvalueCorrectedShampooPreconditionerList - PerFactorEigenvalueCorrectedKLShampooPreconditionerList (KL variant) - PerFactorEigenvalueCorrectedShampooPreconditionerConfig - PerFactorEigenvalueCorrectedKLShampooPreconditionerConfig Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Test the combined PerFactor+KL variant which recomputes eigenvalues every step and preconditions gradients before outer products. Uses beta2=0 and epsilon=1.0 to get clean expected values, leveraging the perturb_before_computation happy path where KL is effectively a no-op when eigenvalues are equal. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…sses Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…tion
Instead of recomputing eigenvalues as diag(Q^T M Q) every step (which causes
discontinuities when eigenvectors Q change at amortized computation steps),
maintain a direct EMA: E_t = beta2 * E_{t-1} + w * diag(Q^T O_t Q).
Key changes:
- Add _get_eigenvalue_allocator hook for zero-init eigenvalues (EMA from zero)
- Add _post_outer_product_hook for eigenvalue EMA update using outer products
- Replace hasattr dispatch with _include_eigenvalues_in_amortized_computation
ClassVar (addresses existing TODO), set False on PerFactor instances to
prevent amortized computation from overwriting EMA'd eigenvalues
- Add _prepare_eigenvalues_for_preconditioning hook for bias correction
- Add zero-eigenvalue guard in KL variant to skip preconditioning at init
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
meta-cla
Bot
added
the
CLA Signed
|
@hjmshi has imported this pull request. If you are a Meta employee, you can view this in D104875559. |
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
1 participant
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.
Summary
diag(Q^T M Q)eigenvalue computation (whereMis the accumulated factor matrix) with an EMA over per-step projections:E_t = beta2 * E_{t-1} + w * diag(Q^T O_t Q), whereO_tis the per-step outer product._post_outer_product_hookand_get_eigenvalue_allocatoroverrides, with the amortized computation's eigenvalue overwrite disabled per-instance.Stack
This PR is part of a stack adding per-factor eigenvalue correction to Distributed Shampoo:
_transform_grad_for_outer_producthook (KL refactor)Test plan
distributed_shampoo/tests/,distributed_shampoo/preconditioner/tests/)make type-check)Generated with Claude Code