[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
+5 -1
View File
@@ -746,7 +746,8 @@ def marlin_moe_generate_valid_test_cases():
for sub_case in inner_combinations:
if (
sub_case[0] == scalar_types.float8_e4m3fn
and current_platform.get_device_capability() not in [89, 120]
and not current_platform.is_device_capability(89)
and not current_platform.is_device_capability_family(120)
):
continue
@@ -897,6 +898,7 @@ class MarlinMoEWeightData:
marlin_moe_generate_valid_test_cases(),
)
@pytest.mark.skipif(current_platform.is_rocm(), reason="Skip for rocm")
@pytest.mark.usefixtures("default_vllm_config")
def test_fused_marlin_moe(
a_type: ScalarType,
b_type: ScalarType,
@@ -1009,6 +1011,7 @@ def test_fused_marlin_moe(
@pytest.mark.flaky(reruns=2)
@pytest.mark.skipif(current_platform.is_rocm(), reason="Skip for rocm")
@pytest.mark.usefixtures("default_vllm_config")
@pytest.mark.parametrize("m", [1, 256])
def test_fused_marlin_moe_with_bias(m):
set_random_seed(0)
@@ -1081,6 +1084,7 @@ def test_fused_marlin_moe_with_bias(m):
@pytest.mark.flaky(reruns=2)
@pytest.mark.skipif(current_platform.is_rocm(), reason="Skip for rocm")
@pytest.mark.usefixtures("default_vllm_config")
@pytest.mark.parametrize("m", [1, 64, 256])
@pytest.mark.parametrize("n,k", [(1024, 1024), (2048, 2048)])
@pytest.mark.parametrize("e,topk", [(8, 2), (64, 4)])
@@ -381,7 +381,8 @@ def marlin_generate_valid_test_cases():
for sub_case in inner_combinations:
if (
sub_case[0] == scalar_types.float8_e4m3fn
and current_platform.get_device_capability() not in [89, 120]
and not current_platform.is_device_capability(89)
and not current_platform.is_device_capability_family(120)
):
continue
args = sub_case + (size_m, size_n, size_k) + case[4:]