Repository navigation
[Fix] [Intel] moe: Route dropped slots at weight 0 - #47664
Closed
dayanandav wants to merge 2 commits into
Closed
dayanandav wants to merge 2 commits into
dayanandav wants to merge 2 commits into
Conversation
dayanandav
requested review from
Alisehen,
AniZpZ,
BBuf,
Edwardf0t1,
FlamingoPg,
HaiShaw,
OrangeRedeng,
b8zhong and
ch-wan
as code owners
October 31, 2026 14:48
dayanandav
requested review from
Fridge003,
Ying1123,
ispobock and
merrymercy
as code owners
October 31, 2026 14:48
This comment has been minimized.
This comment has been minimized.
… weight 0 Motivation: EP and padded rows put -1 in topk_ids; sgl-kernel-xpu fused_experts has no drop sentinel, so the device is lost on the first prefill with --ep-size > 1. Modifications: - moe/utils: route_dropped_experts_to_local maps -1 to expert 0 at weight 0. - Use it at the four XPU fused_experts call sites (fused_moe, unquant, fp8, mxfp4). Other backends are unchanged. - Tests: unit/layers/moe/test_moe_utils.py (CPU), xpu/test_moe_ep_drop_sentinel.py. Accuracy Tests: gpt-oss-20b on Arc Pro B60, TP2/4 + EP2/4: greedy first tokens match TP 16/16. Speed Tests and Profiling: EP prefill 0.70-0.77x TP on 2 cards, 0.56-0.61x on 4 (previously crashed). Checklist: - [x] pre-commit - [x] unit tests - [x] accuracy and speed results Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
dayanandav
force-pushed
the
xpu-ep-drop-sentinel
branch
from
October 31, 2026 15:06
6421695 to
d90a81d
Compare
This comment has been minimized.
This comment has been minimized.
airMeng
pushed a commit
to airMeng/sglang
that referenced
this pull request
Nov 4, 2026
Contributor
Author
|
@siju-samuel @ANSHUMAN87 help to review. |
airMeng
pushed a commit
to airMeng/sglang
that referenced
this pull request
Nov 5, 2026
airMeng
pushed a commit
to airMeng/sglang
that referenced
this pull request
Nov 6, 2026
airMeng
pushed a commit
to airMeng/sglang
that referenced
this pull request
Nov 6, 2026
airMeng
pushed a commit
to airMeng/sglang
that referenced
this pull request
Nov 7, 2026
Contributor
Author
|
Merged as part of #47956 consolidation. |
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.
Motivation
moe: Route EP-dropped expert slots to a local expert at weight 0, EP and padded rows put -1 in topk_ids; sgl-kernel-xpu fused_experts has no drop sentinel, so the device is lost on the first prefill with --ep-size > 1.
Modifications
Accuracy Tests
gpt-oss-20b on Arc Pro B60, TP2/4 + EP2/4: greedy first tokens match TP 16/16.
Speed Tests and Profiling
EP prefill 0.70-0.77x TP on 2 cards, 0.56-0.61x on 4 (previously crashed).
Checklist
CI States
Latest PR Test (Base): ❌ Run #41569384253
Latest PR Test (Extra): ❌ Run #41569384196
Latest PR Test (AMD ROCm 10): ❌ Run #41569384214
Tip
Base CI only covers the core cases, and its workload may not exercise your change. Use
/rerun-testto run the tests your change touches, or copy this prompt for an agent: