diff --git a/vllm/platforms/cuda.py b/vllm/platforms/cuda.py
index 39101c43142f7..915392a4125f9 100644
--- a/vllm/platforms/cuda.py
+++ b/vllm/platforms/cuda.py
@@ -182,8 +182,8 @@ class CudaPlatformBase(Platform):
 
             if vllm_config.attention_config.backend is None:
                 # Default case
-                if cls.is_device_capability(100):
-                    # Blackwell => Force CutlassMLA.
+                if cls.is_device_capability(100) and not use_sparse:
+                    # Blackwell => Force CutlassMLA (unless sparse, i.e. DSv3.2).
                     use_cutlass_mla = True
                     # Set the backend in AttentionConfig so it's used during
                     # backend selection