Fix GPT-OSS MXFP4 hidden size reshape on SM10X#27528
Merged
hnyls2002 merged 1 commit intoJun 8, 2026
Merged
Conversation
Signed-off-by: Mohammad Miadh Angkad <[email protected]>
Contributor
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
5 tasks
Collaborator
|
/rerun-test test/registered/models_e2e/test_gpt_oss_4gpu_mxfp4.py test/registered/lora/test_lora_gpt_oss_20b_logprob_diff.py |
Contributor
|
Results for 🚀 🚀 |
Oasis-Git
added a commit
to Oasis-Git/sglang
that referenced
this pull request
Jun 8, 2026
7 commits from main. Notable: - sgl-project#27528 Fix GPT-OSS MXFP4 hidden size reshape on SM10X — upstream landed the same fix I had locally for the prepad bug in forward_normal; dropped my local commit f3c169b85 before merging so the upstream version comes in cleanly. - sgl-project#25464 Deprecate Spec V1 — removed python/sglang/srt/speculative/{eagle_worker,multi_layer_eagle_worker, standalone_worker}.py. We had local edits on eagle_worker.py for the cuda_graph_bs -> cuda_graph_bs_decode rename; accepted the deletion (V2 was already the runtime path; no remaining call sites for V1 in srt/). - Other content (dsv4 ROCm fix, EAGLE SWA pool, TRTLLM target verify, NPU CI, LMSYS blog) auto-merged without conflicts. Stale-import surface re-checked post-merge: no leftover imports of piecewise_context_manager, the old piecewise_cuda_graph_runner, the old cuda_graph_runner module, or the deleted V1 spec workers. No stale server_args.cuda_graph_bs / disable_piecewise_cuda_graph accesses in python/sglang/srt/. Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
jeynmann
pushed a commit
to jeynmann/sglang
that referenced
this pull request
Jun 9, 2026
Signed-off-by: Mohammad Miadh Angkad <[email protected]>
5 tasks
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
2 participants
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
Fix a regression from #27063 where GPT-OSS used the backend-internal expert hidden size for the final MoE reshape.
On SM10X with FlashInfer MXFP4,
self.experts.hidden_sizeis padded from2880to3072, whileconfig.hidden_sizeremains2880, causing warmup to fail:I also checked H200, where serving succeeds because it uses
moe_runner_backend='triton_kernel'and does not pad the expert hidden size to3072.Test Plan
Before
On one GPU from a single GB300 node, GPT-OSS 20B failed during server warmup before this fix:
Failure:
After
With this fix, the same server command succeeds, and GPQA eval completes:
Result:
CI States
Latest PR Test (Base): ❌ Run #27116872906
Latest PR Test (Extra): ❌ Run #27116872875