From 9a08a5118e4c527e76dbcc2ba589dc0d58b2eac9 Mon Sep 17 00:00:00 2001
From: Gabriel Wu <13583761+lucifer1004@users.noreply.github.com>
Date: Wed, 1 Jul 2026 12:32:54 +0800
Subject: [PATCH] fix: skip cooperative top-K on SM120 (#47164)

Signed-off-by: Zihua Wu <13583761+lucifer1004@users.noreply.github.com>
---
 vllm/model_executor/layers/sparse_attn_indexer.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/vllm/model_executor/layers/sparse_attn_indexer.py b/vllm/model_executor/layers/sparse_attn_indexer.py
index 678b66eab8..ceb52e5d32 100644
--- a/vllm/model_executor/layers/sparse_attn_indexer.py
+++ b/vllm/model_executor/layers/sparse_attn_indexer.py
@@ -578,6 +578,7 @@ def sparse_attn_indexer(
             and num_rows <= 32
             and logits.stride(0) % 4 == 0  # TMA 16-byte alignment
             and current_platform.has_device_capability(90)
+            and not current_platform.is_device_capability_family(120)
         )
         use_persistent_topk = current_platform.is_cuda() and topk_tokens in (
             512,