From d72e59508b7554045cb51827f9b8d0f08c7a3abc Mon Sep 17 00:00:00 2001 From: Qiaolin Yu Date: Mon, 14 Sep 2026 14:05:41 -0700 Subject: [PATCH] Reland fix(qsa): clamp the compress gather to the rows (#38346) (#39446) Co-authored-by: ehuaa --- python/sglang/srt/layers/attention/qsa/qsa_indexer.py | 1 + 1 file changed, 1 insertion(+) diff --git a/python/sglang/srt/layers/attention/qsa/qsa_indexer.py b/python/sglang/srt/layers/attention/qsa/qsa_indexer.py index 4de4c6a8d..a15585835 100644 --- a/python/sglang/srt/layers/attention/qsa/qsa_indexer.py +++ b/python/sglang/srt/layers/attention/qsa/qsa_indexer.py @@ -321,6 +321,7 @@ class QSAIndexer(MultiPlatformOp): self.compress_ratio, device=member_rows.device, dtype=torch.long ) source_keys = token_k + group_locs = group_locs.clamp_max(source_keys.shape[0] - 1) source_rope = metadata.extend_rope_matrix if source_rope is None: source_rope = build_rope_position_matrix(