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
[Log] Warning Once for Cutlass MLA (#23137)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Aug 19, 2025 at 02:24 UTC
90bbe0a5adc0b9636e84ed3ad058d5853158f4bb
1 file changed
+3
-3
vllm/v1/attention/backends/mla/cutlass_mla.py
+3
-3
@@ -115,7 +115,7 @@ class CutlassMLAImpl(MLACommonImpl[MLACommonMetadata]):
115
self._use_old_cutlass_mla = False
116
force_old_cutlass = os.environ.get("FORCE_OLD_CUTLASS_MLA", None)
117
if force_old_cutlass:
118
- logger.warning("Forcing old cutlass mla kernel")
118
+ logger.warning_once("Forcing old cutlass mla kernel")
119
self._use_old_cutlass_mla = True
120
121
# TODO: Currently, num_kv_splits is limited to 16 to avoid hanging
@@ -123,8 +123,8 @@ class CutlassMLAImpl(MLACommonImpl[MLACommonMetadata]):
123
# FORCE_NUM_KV_SPLITS=1
124
force_num_kv_splits = os.environ.get("FORCE_NUM_KV_SPLITS", None)
125
if force_num_kv_splits:
126
- logger.warning("Forcing num_kv_splits to %d",
127
- int(force_num_kv_splits))
126
+ logger.warning_once("Forcing num_kv_splits to %d",
127
+ int(force_num_kv_splits))
128
self._num_kv_splits = int(force_num_kv_splits)
129
else:
130
self._num_kv_splits = -1 # => Auto-detect