Reland spec v2 tree drafting (eagle topk>1) with page_size==1 (#26866)#26997
Merged
Conversation
hnyls2002
requested review from
Qiaolin-Yu,
Ying1123 and
merrymercy
as code owners
June 2, 2026 02:46
Contributor
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
Collaborator
Author
|
/tag-and-rerun-ci extra |
hnyls2002
requested review from
hanming-lu,
hzh0425,
ispobock,
xiezhq-hermann and
yizhang2077
as code owners
June 2, 2026 03:03
Collaborator
Author
|
/rerun-test test_spec_eagle_page.py test_eagle_dp_attention.py test_eagle_infer_beta_dp_attention_large.py |
hnyls2002
requested review from
Fridge003,
HaiShaw,
alphabetc1 and
hebiao064
as code owners
June 2, 2026 05:12
Collaborator
Author
|
/rerun-test test_constrained_decoding_spec_reasoning.py |
Contributor
|
Results for 🚀 |
Collaborator
Author
|
/rerun-test test_qwen3_next_models_mtp.py |
Contributor
|
Results for 🚀 |
# Conflicts: # python/sglang/srt/speculative/eagle_info_v2.py
hnyls2002
force-pushed
the
lsyin/reland-26866
branch
from
June 3, 2026 06:04
8fb54b3 to
2b402cc
Compare
Collaborator
Author
|
Latest PR Test (Base): ✅ Run #26898846215 |
Oasis-Git
added a commit
to Oasis-Git/sglang
that referenced
this pull request
Jun 3, 2026
…acy writes Conflicts: - python/sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py: PR sgl-project#26997 (Reland spec v2 tree drafting for eagle topk>1 with page_size==1) resizes the cuda-graph buffer width from speculative_num_steps + 1 to speculative_num_draft_tokens. Take upstream — this is a correctness fix for topk > 1 draft-extend buffer overflow. Silent legacy writes flagged by post-merge sweep (NOT introduced by this merge — pre-existing breakage in cg-refactor's earlier disable_ cuda_graph migration; surfaced now while we have the audit running): - python/sglang/compile_deep_gemm.py refine_server_args() and python/sglang/srt/arg_groups/pd_disaggregation_hook.py both wrote server_args.disable_cuda_graph = True after ServerArgs.__post_init__ had already consumed that field. The post_init read-once-then-mirror pattern means those late writes were silent no-ops. Switch both to set cuda_graph_config[Phase.DECODE].backend / [Phase.PREFILL].backend = Backend.DISABLED directly — same semantic, propagates correctly. Validation: attention unittests 181 passed / 0 failures / 548 subtests.
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.
Relands #26866 (reverted in #26981).
CI States
Latest PR Test (Base): 🚫 Run #26907731073
Latest PR Test (Extra): ✅ Run #26907730270