fix: skip cooperative top-K on SM120 (#47164)
Signed-off-by: Zihua Wu <13583761+lucifer1004@users.noreply.github.com>
Gabriel Wu committed
Jul 1, 2026 at 12:32 UTC
9a08a5118e4c527e76dbcc2ba589dc0d58b2eac9
1 file changed
+1
vllm/model_executor/layers/sparse_attn_indexer.py
+1
@@ -578,6 +578,7 @@ def sparse_attn_indexer(
578
and num_rows <= 32
579
and logits.stride(0) % 4 == 0 # TMA 16-byte alignment
580
and current_platform.has_device_capability(90)
581
+ and not current_platform.is_device_capability_family(120)
582
)
583
use_persistent_topk = current_platform.is_cuda() and topk_tokens in (
584
512,