From 0044c4038c571e7fb38acb008256f925dfe515f1 Mon Sep 17 00:00:00 2001 From: Lucas Wilkinson Date: Sun, 7 Dec 2025 10:53:51 -0500 Subject: [PATCH] [BugFix][DeepSeek-V3.2] Fix backend selection logic for Blackwell (#30195) --- vllm/platforms/cuda.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/vllm/platforms/cuda.py b/vllm/platforms/cuda.py index 39101c4314..915392a412 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