Repository navigation
[ROCm][CI] Gate MoRI EPv2 modular-kernel cases on the production EPv2 checks - #65184
Merged
Merged
Conversation
|
✅ @djramic, CI is now available for this PR.
|
Collaborator
Author
|
/ci run |
|
✅ Triggered Buildkite CI #99242 for commit ℹ️ This PR is 1 commit behind upstream |
… checks Signed-off-by: Djordje Ramic <djoramic@amd.com>
djramic
force-pushed
the
mk-gate-mori-epv2
branch
from
November 2, 2026 16:00
683d39e to
c1bb3b9
Compare
AndreasKaratzas
approved these changes
Nov 2, 2026
Member
|
/ci run |
|
✅ Triggered Buildkite CI #99247 for commit ℹ️ This PR is 15 commits behind upstream |
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Overview
Since #61361,
:amd: (MI300) MoE Kernels Multi-GPUfails on main (amd-ci#15761): all 3MoriEPv2PrepareAndFinalizecases hit a MoRI assertion on MI300X. This test-only PR registers MoRI EPv2 in the modular-kernel harness only where production would select it.Claims
RocmPlatform.check_and_update_configuses.Validation
MI300X,
rocm/vllm-dev:ci_basewith MoRI 1.2.3.post1 (as in CI), 2 GPUs:cd tests VLLM_ROCM_USE_AITER=1 VLLM_ROCM_USE_AITER_MOE=1 pytest -v -s kernels/moe/test_modular_kernel_combinations.py -k multigpuOn main, the EPv2 cases fail locally with the CI error:
AssertionError: combine block_num 304 exceeds xdb_flag_slots 256.Details
The harness sets
mori_ep_v2=Truefor the"mori"backend. That skips the EPv1 fallback production uses unless MoRI >= 1.2.4 is installed and the GPU is gfx950/gfx1250. On MI300X, MoRI's untuned EPv2 path uses one combine block per CU (304), above its fixed limit of 256. ROCm/mori#876 fixes this, but no MoRI release includes it yet.AI-assisted tooling was used to analyze the CI logs, write the change, and run the validation. The change and the results were manually reviewed.