Conversation
Extends find_kv_cache_attention to match the GPT-OSS attention-sink path (slice(softmax(concat(scores, sink_col)))) so decode attention fuses (else it runs unfused static-shape ops that truncate to T=1 at decode). Adds 'concat' to valid_attn_ops, plus the GQA sliding-window decode-mask off-by-one + sink-slice (ends=-1 dynamic) fixes in parse_group_query_attention, and a rewrite_reduce tuple guard. Co-Authored-By: Claude <noreply@anthropic.com>
rlegithub
marked this pull request as ready for review
September 14, 2026 21:21
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Enables the GPT-OSS attention-sink path in decode-attention fusion. GPT-OSS inserts
concat(scores, sink_col) -> softmax -> slicebefore the V GEMM; the previousfind_kv_cache_attentionmatcher only handledsoftmax_input(mask), so the sink variant didn't fuse and ran as unfused static-shape ops that truncate to T=1 at decode.fuse_attention.cpp: matchslice(softmax(concat(any(), sink_col))); add"concat"tovalid_attn_opsparse_group_query_attention.cpp: sliding-window decode-mask off-by-one + sink-sliceends=-1(dynamic) fixesrewrite_reduce.cpp: tuple guard3 of 3 PRs enabling GPT-OSS-20B (with the fused MoE op and SLN-FP32 PRs).
Test plan