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 14, 05:30 UTC
[Bugfix] Fix default enable for CUTLASS MLA on SM100 (#22738)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Aug 13, 2025 at 00:22 UTC
b1361c7273f60ca244e5425bdb7a9120057327fe
1 file changed
+6
-1
vllm/platforms/cuda.py
+6
-1
@@ -152,6 +152,9 @@ class CudaPlatformBase(Platform):
152
if cls.is_device_capability(100):
153
# Blackwell => Force CutlassMLA.
154
use_cutlass_mla = True
155
+ # TODO: This does not work, because the
156
+ # global_force_attn_backend_context_manager is not set.
157
+ # See vllm/attention/selector.py:_cached_get_attn_backend
158
envs.VLLM_ATTENTION_BACKEND = "CUTLASS_MLA"
159
else:
160
# Not Blackwell
@@ -217,7 +220,9 @@ class CudaPlatformBase(Platform):
220
if use_mla:
221
# TODO(lucas): refactor to be more concise
222
# we should probably consider factoring out V1 here
220
- if selected_backend == _Backend.CUTLASS_MLA:
223
+ if selected_backend == _Backend.CUTLASS_MLA or (
224
+ cls.is_device_capability(100) and selected_backend is None
225
+ and block_size == 128):
226
if use_v1:
227
logger.info_once("Using Cutlass MLA backend on V1 engine.")
228
return ("vllm.v1.attention.backends.mla."