--- a/vllm/sparse_attn_indexer_kpool.py
+++ b/vllm/sparse_attn_indexer_kpool.py
@@ -807,7 +807,11 @@
         else:
             topk_dst = topk_indices_buffer[:num_padded_tokens, :topk_tokens]
 
-        if current_platform.is_cuda() and select_k in (512, 1024, 2048):
+        if (
+            current_platform.is_cuda()
+            and select_k in (512, 1024, 2048)
+            and not current_platform.is_device_capability_family(120)
+        ):
             workspace_manager = current_workspace_manager()
             (topk_workspace,) = workspace_manager.get_simultaneous(
                 ((RADIX_TOPK_WORKSPACE_SIZE,), torch.uint8),