v4.2 release. (#2587)

* Fix default cluster callback values to 1 to avoid profiler failure when these values are not set in command line.

* v4.2 release.
This commit is contained in:
Junkai-Wu
2025-08-23 06:11:24 +08:00
committed by GitHub
parent 11cad1f67b
commit a49a78ffef
351 changed files with 28182 additions and 2032 deletions

View File

@@ -546,6 +546,22 @@ class KernelScheduleType(enum.Enum):
Nvf4TmaWarpSpecialized1SmSm100 = enum_auto()
Nvf4TmaWarpSpecialized2SmSm100 = enum_auto()
# FP4 Ultra
BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs16Sm103 = enum_auto()
BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs16Sm103 = enum_auto()
BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs32Sm103 = enum_auto()
BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs32Sm103 = enum_auto()
BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs16Sm103DisablePrefetch = enum_auto()
BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs16Sm103DisablePrefetch = enum_auto()
BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs32Sm103DisablePrefetch = enum_auto()
BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs32Sm103DisablePrefetch = enum_auto()
BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs16Sm103TmaPrefetch = enum_auto()
BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs16Sm103TmaPrefetch = enum_auto()
BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs32Sm103TmaPrefetch = enum_auto()
BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs32Sm103TmaPrefetch = enum_auto()
Mxf8f6f4TmaWarpSpecializedCooperativeSm120 = enum_auto()
Mxf8f6f4TmaWarpSpecializedPingpongSm120 = enum_auto()
Nvf4TmaWarpSpecializedCooperativeSm120 = enum_auto()
@@ -603,6 +619,22 @@ KernelScheduleTag = {
KernelScheduleType.Nvf4TmaWarpSpecialized1SmSm100: 'cutlass::gemm::KernelTmaWarpSpecialized1SmNvf4Sm100',
KernelScheduleType.Nvf4TmaWarpSpecialized2SmSm100: 'cutlass::gemm::KernelTmaWarpSpecialized2SmNvf4Sm100',
# FP4 Ultra
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs16Sm103: 'cutlass::gemm::KernelTmaWarpSpecialized1SmBlockScaledMxNvf4UltraVs16Sm103',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs16Sm103: 'cutlass::gemm::KernelTmaWarpSpecialized2SmBlockScaledMxNvf4UltraVs16Sm103',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs32Sm103: 'cutlass::gemm::KernelTmaWarpSpecialized1SmBlockScaledMxNvf4UltraVs32Sm103',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs32Sm103: 'cutlass::gemm::KernelTmaWarpSpecialized2SmBlockScaledMxNvf4UltraVs32Sm103',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs16Sm103TmaPrefetch: 'cutlass::gemm::KernelTmaWarpSpecialized1SmBlockScaledMxNvf4UltraVs16Sm103TmaPrefetch',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs16Sm103TmaPrefetch: 'cutlass::gemm::KernelTmaWarpSpecialized2SmBlockScaledMxNvf4UltraVs16Sm103TmaPrefetch',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs32Sm103TmaPrefetch: 'cutlass::gemm::KernelTmaWarpSpecialized1SmBlockScaledMxNvf4UltraVs32Sm103TmaPrefetch',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs32Sm103TmaPrefetch: 'cutlass::gemm::KernelTmaWarpSpecialized2SmBlockScaledMxNvf4UltraVs32Sm103TmaPrefetch',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs16Sm103DisablePrefetch: 'cutlass::gemm::KernelTmaWarpSpecialized1SmBlockScaledMxNvf4UltraVs16Sm103DisablePrefetch',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs16Sm103DisablePrefetch: 'cutlass::gemm::KernelTmaWarpSpecialized2SmBlockScaledMxNvf4UltraVs16Sm103DisablePrefetch',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs32Sm103DisablePrefetch: 'cutlass::gemm::KernelTmaWarpSpecialized1SmBlockScaledMxNvf4UltraVs32Sm103DisablePrefetch',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs32Sm103DisablePrefetch: 'cutlass::gemm::KernelTmaWarpSpecialized2SmBlockScaledMxNvf4UltraVs32Sm103DisablePrefetch',
KernelScheduleType.PtrArrayTmaWarpSpecializedCooperative: 'cutlass::gemm::KernelPtrArrayTmaWarpSpecializedCooperative',
KernelScheduleType.PtrArrayTmaWarpSpecializedCooperativeFP8FastAccum: 'cutlass::gemm::KernelPtrArrayTmaWarpSpecializedCooperativeFP8FastAccum',
KernelScheduleType.PtrArrayTmaWarpSpecializedPingpong: 'cutlass::gemm::KernelPtrArrayTmaWarpSpecializedPingpong',
@@ -677,6 +709,21 @@ KernelScheduleSuffixes = {
KernelScheduleType.Nvf4TmaWarpSpecialized1SmSm100: '_o_vs16_1sm',
KernelScheduleType.Nvf4TmaWarpSpecialized2SmSm100: '_o_vs16_2sm',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs16Sm103: '_o_vs16_1sm',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs16Sm103: '_o_vs16_2sm',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs32Sm103: '_o_vs32_1sm',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs32Sm103: '_o_vs32_2sm',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs16Sm103DisablePrefetch: '_o_vs16_1sm_nopf',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs16Sm103DisablePrefetch: '_o_vs16_2sm_nopf',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs32Sm103DisablePrefetch: '_o_vs32_1sm_nopf',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs32Sm103DisablePrefetch: '_o_vs32_2sm_nopf',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs16Sm103TmaPrefetch: '_o_vs16_1sm_tmapf',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs16Sm103TmaPrefetch: '_o_vs16_2sm_tmapf',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized1SmVs32Sm103TmaPrefetch: '_o_vs32_1sm_tmapf',
KernelScheduleType.BlockScaledMxNvf4UltraTmaWarpSpecialized2SmVs32Sm103TmaPrefetch: '_o_vs32_2sm_tmapf',
KernelScheduleType.PtrArrayTmaWarpSpecializedCooperative: '_warpspecialized_cooperative',
KernelScheduleType.PtrArrayTmaWarpSpecializedCooperativeFP8FastAccum: '_warpspecialized_cooperative_fp8_fastaccum',
KernelScheduleType.PtrArrayTmaWarpSpecializedPingpong: '_warpspecialized_pingpong',
@@ -713,8 +760,12 @@ class EpilogueScheduleType(enum.Enum):
PtrArrayNoSmemWarpSpecialized = enum_auto()
NoSmemWarpSpecialized1Sm = enum_auto()
NoSmemWarpSpecialized2Sm = enum_auto()
FastF32NoSmemWarpSpecialized1Sm = enum_auto()
FastF32NoSmemWarpSpecialized2Sm = enum_auto()
PtrArrayNoSmemWarpSpecialized1Sm = enum_auto()
PtrArrayNoSmemWarpSpecialized2Sm = enum_auto()
PtrArrayFastF32NoSmemWarpSpecialized1Sm = enum_auto()
PtrArrayFastF32NoSmemWarpSpecialized2Sm = enum_auto()
TmaWarpSpecialized = enum_auto()
TmaWarpSpecializedCooperative = enum_auto()
TmaWarpSpecialized1Sm = enum_auto()
@@ -732,8 +783,12 @@ EpilogueScheduleTag = {
EpilogueScheduleType.PtrArrayNoSmemWarpSpecialized: 'cutlass::epilogue::PtrArrayNoSmemWarpSpecialized',
EpilogueScheduleType.NoSmemWarpSpecialized1Sm: 'cutlass::epilogue::NoSmemWarpSpecialized1Sm',
EpilogueScheduleType.NoSmemWarpSpecialized2Sm: 'cutlass::epilogue::NoSmemWarpSpecialized2Sm',
EpilogueScheduleType.FastF32NoSmemWarpSpecialized1Sm: 'cutlass::epilogue::FastF32NoSmemWarpSpecialized1Sm',
EpilogueScheduleType.FastF32NoSmemWarpSpecialized2Sm: 'cutlass::epilogue::FastF32NoSmemWarpSpecialized2Sm',
EpilogueScheduleType.PtrArrayNoSmemWarpSpecialized1Sm: 'cutlass::epilogue::PtrArrayNoSmemWarpSpecialized1Sm',
EpilogueScheduleType.PtrArrayNoSmemWarpSpecialized2Sm: 'cutlass::epilogue::PtrArrayNoSmemWarpSpecialized2Sm',
EpilogueScheduleType.PtrArrayFastF32NoSmemWarpSpecialized1Sm: 'cutlass::epilogue::PtrArrayFastF32NoSmemWarpSpecialized1Sm',
EpilogueScheduleType.PtrArrayFastF32NoSmemWarpSpecialized2Sm: 'cutlass::epilogue::PtrArrayFastF32NoSmemWarpSpecialized2Sm',
EpilogueScheduleType.TmaWarpSpecialized: 'cutlass::epilogue::TmaWarpSpecialized',
EpilogueScheduleType.TmaWarpSpecializedCooperative: 'cutlass::epilogue::TmaWarpSpecializedCooperative',
EpilogueScheduleType.TmaWarpSpecialized1Sm: 'cutlass::epilogue::TmaWarpSpecialized1Sm',
@@ -752,8 +807,12 @@ EpilogueScheduleSuffixes = {
EpilogueScheduleType.PtrArrayNoSmemWarpSpecialized: '_epi_nosmem',
EpilogueScheduleType.NoSmemWarpSpecialized1Sm: '_epi_nosmem',
EpilogueScheduleType.NoSmemWarpSpecialized2Sm: '_epi_nosmem',
EpilogueScheduleType.FastF32NoSmemWarpSpecialized1Sm: '_epi_nosmem_fastf32',
EpilogueScheduleType.FastF32NoSmemWarpSpecialized2Sm: '_epi_nosmem_fastf32',
EpilogueScheduleType.PtrArrayNoSmemWarpSpecialized1Sm: '_epi_nosmem',
EpilogueScheduleType.PtrArrayNoSmemWarpSpecialized2Sm: '_epi_nosmem',
EpilogueScheduleType.PtrArrayFastF32NoSmemWarpSpecialized1Sm: '_epi_nosmem_fastf32',
EpilogueScheduleType.PtrArrayFastF32NoSmemWarpSpecialized2Sm: '_epi_nosmem_fastf32',
EpilogueScheduleType.TmaWarpSpecialized: '_epi_tma',
EpilogueScheduleType.TmaWarpSpecializedCooperative: '_epi_tma',
EpilogueScheduleType.TmaWarpSpecialized1Sm: '',