From e0050f287aae8b3ddbe2947a5af1c5dd23db2781 Mon Sep 17 00:00:00 2001 From: Mustafa YILDIRIM <2341474+mustafayildirim@users.noreply.github.com> Date: Thu, 17 Sep 2026 15:34:56 -0700 Subject: [PATCH] [Bugfix][ROCm] Add record_logical_topk_ready to ROCMAiterMLASparseImpl (GLM-5.3-Flash boot crash) (#57252) Signed-off-by: Mustafa YILDIRIM Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Co-authored-by: Andreas Karatzas --- vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py b/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py index e1365f4d30..69b6e342cd 100644 --- a/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py +++ b/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py @@ -808,6 +808,11 @@ class ROCMAiterMLASparseImpl( (q_concat_shape, vllm_config.model_config.dtype), ) + def record_logical_topk_ready(self) -> None: + # This impl shares the top-k indices buffer via SharedTopkIndicesBuffer + # but does not participate in sparse-MLA index groups. + pass + def _forward_mla( self, layer: AttentionLayer,