[BugFix][DeepSeek-V3.2] Fix backend selection logic for Blackwell (#30195)

0044c403 · Lucas Wilkinson · GitHub · b952f4d3 · 0044c403
Unverified Commit 0044c403 authored Dec 07, 2025 by Lucas Wilkinson Committed by GitHub Dec 07, 2025
Show whitespace changes
Inline Side-by-side

Showing with 2 additions and 2 deletions

vllm/platforms/cuda.py vllm/platforms/cuda.py +2 -2

No files found.
--- 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