{{ message }}
Repository navigation
[Fix] dsv4: Write idle SWA stores to an idle request-window layout - #45586
Merged
Merged
Conversation
hnyls2002
requested review from
Fridge003,
HaiShaw,
Qiaolin-Yu,
Ying1123,
alphabetc1,
hanming-lu,
hebiao064,
huangtingwei9988,
hzh0425,
ispobock,
merrymercy,
xiezhq-hermann and
yizhang2077
as code owners
October 21, 2026 10:28
Collaborator
Author
Collaborator
Author
|
/tag-and-rerun-ci |
Contributor
kkHuang-amd
pushed a commit
to HaiShaw/sglang
that referenced
this pull request
Oct 22, 2026
Resolve conflicts with sgl-project#45579 (history_src precomputed in the layout: drop the HIP history-row cache, pass zero_row, add history_src to the direct layout) and sgl-project#45586 (idle request-window layout). Gate the direct window branches on _is_hip, move the direct test to AMD CI, and give the HIP floor test mock a zero_row.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
DeepseekV4AttnBackend.get_swa_out_cache_locno longer reads attention metadata on an idle forward under--enable-encoder-swa-bounded-replay. It returns the write locations ofRequestWindow.activate_idle, a cached layout whose rows land in the workspace past an empty history.init_forward_metadata. The store then read the previous forward's request-window layout, orNoneafter an unpadded idle. The result was either aTensorMatcherfailure infused_k_norm_rope_flashmla(the stalewrite_loclength does not match the padded rows) orAttributeError: 'NoneType' object has no attribute 'core_attn_metadata'.--moe-a2a-backend megamoe --speculative-algorithm DSPARK --enable-encoder-swa-bounded-replay, GSM8K through sgl-eval (200 examples, 64 threads, max tokens 4096, temperature 0, no thinking). The server crashed at 197/200 when one rank ran an idle verify. With this change the run completes 200/200 at 97.50%, the same score as without encoder replay.engram serves extend, target-verify and decode, not 4on the same idle forward; [Fix] dsv4: Skip DP-attention pad rows in the engram hasher and low-ratio sources #40424 has since merged.No behavior change outside idle forwards with
--enable-encoder-swa-bounded-replay; idle rows were already discarded and never committed.CI States
Latest PR Test (Base): 🚫 Run #39683121972
Latest PR Test (Extra): ❌ Run #39683122015
Latest PR Test (AMD ROCm 10): 🚫 Run #39683121706