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] Optimize Log for FP8MOE (#25709)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Sep 30, 2025 at 22:18 UTC
1244948885e364b3103aa3b515ce66fa4d9955fa
1 file changed
+4
-4
vllm/model_executor/layers/quantization/fp8.py
+4
-4
@@ -467,7 +467,8 @@ class Fp8MoEMethod(FusedMoEMethodBase):
467
logger.info_once("DeepGemm disabled: FlashInfer MOE is"
468
" enabled.")
469
elif (is_deep_gemm_supported()):
470
- logger.info_once("Using DeepGemm kernels for Fp8MoEMethod.")
470
+ logger.debug_once(
471
+ "DeepGemm kernels available for Fp8MoEMethod.")
472
self.allow_deep_gemm = True
473
else:
474
logger.warning_once(
@@ -481,9 +482,8 @@ class Fp8MoEMethod(FusedMoEMethodBase):
482
elif (current_platform.is_cuda()
483
and current_platform.is_device_capability(100)
484
and not self.flashinfer_moe_backend):
484
- logger.info_once(
485
- "Using CutlassBlockScaledGroupedGemm kernels for Fp8 MOE "
486
- "on SM100.")
485
+ logger.debug_once(
486
+ "CutlassBlockScaledGroupedGemm available for Fp8MoEMethod.")
487
self.allow_cutlass_block_scaled_grouped_gemm = True
488
489
def create_weights(self, layer: Module, num_experts: int, hidden_size: int,