[Bugfix][ROCm] Add record_logical_topk_ready to ROCMAiterMLASparseImpl (GLM-5.3-Flash boot crash) (#57252)
Signed-off-by: Mustafa YILDIRIM <mustafa@character.ai> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com>
Mustafa YILDIRIM committed
Sep 17, 2026 at 15:34 UTC
e0050f287aae8b3ddbe2947a5af1c5dd23db2781
1 file changed
+5
vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py
+5
@@ -808,6 +808,11 @@ class ROCMAiterMLASparseImpl(
808
(q_concat_shape, vllm_config.model_config.dtype),
809
)
810
811
+ def record_logical_topk_ready(self) -> None:
812
+ # This impl shares the top-k indices buffer via SharedTopkIndicesBuffer
813
+ # but does not participate in sparse-MLA index groups.
814
+ pass
815
+
816
def _forward_mla(
817
self,
818
layer: AttentionLayer,