Skip to content

[Bugfix] Restore overridden HF config fields and support index_skip_topk_offset for DSA topk sharing#27114

Merged
ispobock merged 7 commits into
sgl-project:mainfrom
zRzRzRzRzRzRzR:glm-bug-fix
Jun 6, 2026
Merged

[Bugfix] Restore overridden HF config fields and support index_skip_topk_offset for DSA topk sharing#27114
ispobock merged 7 commits into
sgl-project:mainfrom
zRzRzRzRzRzRzR:glm-bug-fix

Conversation

@zRzRzRzRzRzRzR

@zRzRzRzRzRzRzR zRzRzRzRzRzRzR commented Jun 3, 2026

Copy link
Copy Markdown
Contributor

Restore HF config fields clobbered by GlmMoeDsaConfig and add index_skip_topk_offset handling for DSA topk sharing

And in the PR description, reference the upstream transformers fix:

The GlmMoeDsaConfig drops/clobbers raw checkpoint fields the DSA path needs (qk_rope_head_dim, index_topk_freq), so we re-read them from config.json and restore. Fixed upstream by this pr — this workaround can be removed once SGLang requires transformers >= 5.10.

Also adds handling for the index_skip_topk_offset config. when set, skip_topk / next_skip_topk are computed relative to the offset instead of layer 1.


CI States

Latest PR Test (Base): ✅ Run #27008071717
Latest PR Test (Extra): ❌ Run #27008071538

@gemini-code-assist

Copy link
Copy Markdown
Contributor

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/tag-and-rerun-ci

@zRzRzRzRzRzRzR

Copy link
Copy Markdown
Contributor Author

/tag-and-rerun-ci

1 similar comment
@zRzRzRzRzRzRzR

Copy link
Copy Markdown
Contributor Author

/tag-and-rerun-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

1 similar comment
@zRzRzRzRzRzRzR

Copy link
Copy Markdown
Contributor Author

/rerun-failed-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

6 similar comments
@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

@zRzRzRzRzRzRzR

Copy link
Copy Markdown
Contributor Author

/rerun-failed-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

12 similar comments
@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

@JustinTong0323

Copy link
Copy Markdown
Collaborator

/rerun-failed-ci

@JustinTong0323 JustinTong0323 left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

reuse_mtp_topk_indices defaults to False (not None), which breaks two-batch-overlap on every DeepSeek-family run — the deepep CI jobs (TestMTPWithTBO / TestDeepseekMTP) die at scheduler init, not just the MTP path.

Comment thread python/sglang/srt/model_executor/forward_batch_info.py
@ispobock
ispobock merged commit 393d0e1 into sgl-project:main Jun 6, 2026
257 of 278 checks passed
@zRzRzRzRzRzRzR
zRzRzRzRzRzRzR deleted the glm-bug-fix branch June 6, 2026 07:25
JustinTong0323 added a commit to JustinTong0323/sglang that referenced this pull request Jun 14, 2026
The Spec V1 deprecation (sgl-project#25464) deleted eagle_worker.py, which carried
the only setter for forward_batch.reuse_mtp_topk_indices. The receiver
plumbing (deepseek_nextn, ForwardBatch fields, TBO filter) from sgl-project#27114
survived but went dead. Re-add the setup/reset block to the V2
EagleDraftWorker.draft_forward so DSA/NSA topk indices are shared across
MTP draft steps again (topk==1 only).
ovowei pushed a commit to kvcache-ai/sglang that referenced this pull request Jun 16, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants