Conversation
DeepGEMM-side validation that intra-node compute compensation needs zero kernel change: a helper GPU absorbs peer experts' chunks purely by extending the weight table (b=[local experts ++ migrated peer experts]) and pointing the migrated chunks' task_queue expert_idx at the appended slots — the existing bf16_chunk_gemm_nn / chunk_weighted_swiglu / chunk_zip kernels index the weight table by task.expert_idx and cover all groups. - test_compensation_chunk.py: helper absorbs peer experts; chunk-GEMM output vs m_grouped reference is bitwise diff=0 (all-valid and migrated-only), plus a build_helper_schedule constructor. - test_fp8_compensation_chunk.py: FP8 variant. Co-Authored-By: Claude Code <noreply@anthropic.com>
…ramoe-compensation
Co-Authored-By: Claude Code <noreply@anthropic.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
DeepGEMM-side validation that TeraMoE intra-node compute compensation needs zero kernel change.
A helper GPU can absorb peer experts' chunks purely by:
b = [local experts ++ migrated peer experts], andtask_queueexpert_idxat the appended slots.The existing
bf16_chunk_gemm_nn/chunk_weighted_swiglu/chunk_zipkernels already index the weight table bytask.expert_idxand cover all groups, so no kernel modification is required.Added
tests_overlap/test_compensation_chunk.py: a helper absorbs peer experts (extended-table slots); chunk-GEMM output vs them_groupedreference is bitwisediff=0(both all-valid and migrated-only), plus abuild_helper_scheduleconstructor.tests_overlap/test_fp8_compensation_chunk.py: FP8 variant.Testing
o1/o3) vsm_groupedreference: bitwisediff=0for all-valid and migrated-only schedules.🤖 Generated with Claude Code