[Bugfix] Fix SM121 (DGX Spark) exclusion from Marlin/CUTLASS FP8 paths (#35568)

Signed-off-by: Blake Ledden <blake@secondnaturecomputing.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Pavani Majety <pmajety@nvidia.com>
This commit is contained in:
Blake Ledden
2026-05-15 10:59:00 -07:00
committed by GitHub
co-authored by Claude Opus 4.6 Pavani Majety
parent f45c210885
commit 06d020bb6e
8 changed files with 19 additions and 14 deletions
@@ -202,7 +202,7 @@ struct cutlass_3x_gemm_sm120 {
sizeof(typename CollectiveEpilogue::SharedStorage))>,
KernelSchedule>::CollectiveOp;
using GemmKernel = enable_sm120_only<cutlass::gemm::kernel::GemmUniversal<
using GemmKernel = enable_sm120_family<cutlass::gemm::kernel::GemmUniversal<
Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue, void>>;
};
@@ -72,7 +72,7 @@ struct cutlass_3x_gemm_sm120_custom {
sizeof(typename CollectiveEpilogue::SharedStorage))>,
KernelSchedule, void>::CollectiveOp;
using GemmKernel = enable_sm120_only<cutlass::gemm::kernel::GemmUniversal<
using GemmKernel = enable_sm120_family<cutlass::gemm::kernel::GemmUniversal<
Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue, void>>;
};
@@ -15,11 +15,11 @@ SUPPORT_SM80 = False
for arch in sys.argv[1].split(","):
arch = arch[: arch.index(".") + 2].replace(".", "")
arch = int(arch)
# only SM89 and SM120 fully support
# mma.sync.aligned.m16n8k32.row.col.f32.e4m3.e4m3.f32.
# SM89 and the SM12x family (SM120 RTX 5090, SM121 DGX Spark GB10)
# fully support mma.sync.aligned.m16n8k32.row.col.f32.e4m3.e4m3.f32.
# SM90 and SM100 can use this PTX, but it’s simulated
# with FP16 MMA, so it cannot achieve any acceleration.
if arch in [89, 120]:
if arch == 89 or arch // 10 == 12:
SUPPORT_FP8 = True
if arch >= 80:
SUPPORT_SM80 = True
+2 -2
View File
@@ -448,8 +448,8 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* b_bias,
"FP8 only support Ada Lovelace or newer GPUs.");
TORCH_CHECK(
major_capability * 10 + minor_capability == 89 ||
major_capability * 10 + minor_capability == 120,
"Marlin W4A8-FP8 only support SM89 or SM120 device (It is slower than "
major_capability == 12,
"Marlin W4A8-FP8 only support SM89 or SM12x device (It is slower than "
"Marlin W4A16 on other devices).");
}
+3 -3
View File
@@ -15,11 +15,11 @@ SUPPORT_SM80 = False
for arch in sys.argv[1].split(","):
arch = arch[: arch.index(".") + 2].replace(".", "")
arch = int(arch)
# only SM89 and SM120 fully support
# mma.sync.aligned.m16n8k32.row.col.f32.e4m3.e4m3.f32.
# SM89 and the SM12x family (SM120 RTX 5090, SM121 DGX Spark GB10)
# fully support mma.sync.aligned.m16n8k32.row.col.f32.e4m3.e4m3.f32.
# SM90 and SM100 can use this PTX, but it’s simulated
# with FP16 MMA, so it cannot achieve any acceleration.
if arch in [89, 120]:
if arch == 89 or arch // 10 == 12:
SUPPORT_FP8 = True
if arch >= 80:
SUPPORT_SM80 = True