increase build memory

Signed-off-by: Xiwen Yu <13230610+VALLIS-NERIA@users.noreply.github.com>
This commit is contained in:
Xiwen Yu 2025-09-08 11:28:39 +08:00
parent d42201e235
commit caea58aba4
2 changed files with 5 additions and 11 deletions

View File

@ -141,14 +141,8 @@ void tma_warp_specialized_generic_moe_gemm_kernelLauncher(TmaWarpSpecializedGrou
#ifndef COMPILE_BLACKWELL_SM103_TMA_GROUPED_GEMMS
else if constexpr (ArchTag::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.");
first_time = false;
}
// fallback sm100f logic is done in dispatchMoeGemmFinalDispatchTmaWarpSpecialized
TLLM_THROW("Please recompile with support for blackwell by passing 103-real as an arch to build_wheel.py.");
}
#endif
#ifndef COMPILE_BLACKWELL_SM120_TMA_GROUPED_GEMMS

View File

@ -112,7 +112,7 @@ def BUILD_CONFIGS = [
(WHEEL_EXTRA_ARGS) : "--extra-cmake-vars WARNING_IS_ERROR=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_CUDA_HOST_COMPILER=clang -DCMAKE_LINKER_TYPE=LLD",
(TARNAME) : "llvm-TensorRT-LLM-GH200.tar.gz",
(WHEEL_ARCHS): "90-real;100-real;103-real;120-real",
(BUILD_JOBS_FOR_CONFIG): "4", // TODO: Remove after fix the build OOM issue on SBSA
// (BUILD_JOBS_FOR_CONFIG): "4", // TODO: Remove after fix the build OOM issue on SBSA
],
]
@ -131,8 +131,8 @@ def globalVars = [
// TODO: Move common variables to an unified location
BUILD_CORES_REQUEST = "8"
BUILD_CORES_LIMIT = "8"
BUILD_MEMORY_REQUEST = "48Gi"
BUILD_MEMORY_LIMIT = "64Gi"
BUILD_MEMORY_REQUEST = "64Gi"
BUILD_MEMORY_LIMIT = "80Gi"
BUILD_JOBS = "8"
TESTER_CORES = "12"