kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-01-14 06:27:45 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

disable sm103 moe kernel

Browse Source 
Signed-off-by: Xiwen Yu <13230610+VALLIS-NERIA@users.noreply.github.com>
This commit is contained in:
Xiwen Yu 2025-09-01 10:47:58 +08:00
parent a765ee4d21
commit 14154ec1d3
1 changed files with 4 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_gemm_template_dispatch_tma_ws.h
View File

@ -102,22 +102,22 @@ void dispatchMoeGemmSelectBiasTmaWarpSpecialized(TmaWarpSpecializedGroupedGemmIn
TLLM_THROW("Please recompile with support for hopper by passing 90-real as an arch to build_wheel.py.");
}
#endif
#ifndef COMPILE_BLACKWELL_SM103_TMA_GROUPED_GEMMS
// #ifndef COMPILE_BLACKWELL_SM103_TMA_GROUPED_GEMMS
else if constexpr (Arch::kMinComputeCapability == 103)
{
static bool first_time = true;
if (first_time)
{
TLLM_LOG_WARNING(
"Falling back to sm100f version. For best performance please recompile with support for blackwell by "
"passing 103-real as an arch to build_wheel.py.");
"Falling back to sm100f version due to a bug in cutlass." /*"For best performance please recompile with support for blackwell by "
"passing 103-real as an arch to build_wheel.py."*/);
first_time = false;
}
return dispatchMoeGemmSelectBiasTmaWarpSpecialized<cutlass::arch::Sm100, T, WeightType, OutputType, EpilogueTag,
FUSION, TileShape, ClusterShape>(
hopper_input, num_experts, multi_processor_count, stream, occupancy, workspace_size);
}
#endif
// #endif
#ifndef COMPILE_BLACKWELL_TMA_GROUPED_GEMMS
else if constexpr (Arch::kMinComputeCapability >= 100 && Arch::kMinComputeCapability < 120)
{