A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 11, 21:15 UTC
[BugFix][DeepSeek-V3.2] Fix backend selection logic for Blackwell (#30195)
Lucas Wilkinson committed
Dec 7, 2025 at 10:53 UTC
0044c4038c571e7fb38acb008256f925dfe515f1
1 file changed
+2
-2
vllm/platforms/cuda.py
+2
-2
@@ -182,8 +182,8 @@ class CudaPlatformBase(Platform):
182
183
if vllm_config.attention_config.backend is None:
184
# Default case
185
- if cls.is_device_capability(100):
186
- # Blackwell => Force CutlassMLA.
185
+ if cls.is_device_capability(100) and not use_sparse:
186
+ # Blackwell => Force CutlassMLA (unless sparse, i.e. DSv3.2).
187
use_cutlass_mla = True
188
# Set the backend in AttentionConfig so it's used during
189
# backend selection