[Perf][MoE] Skip padded work in SM100 FP4 DeepGEMM experts (#59044)

Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
This commit is contained in:
Tyler Michael Smith
2026-10-16 16:20:52 -04:00
committed by GitHub
co-authored by Codex Robert Shaw
parent cc2e4fecf2
commit 68ebea789a
2 changed files with 27 additions and 8 deletions
@@ -116,7 +116,7 @@ def test_silu_mul_fp8_quant_deep_gemm(T: int, N: int):
@pytest.mark.parametrize("scale_format", ["packed", "float32", "ceil_ue8m0"])
@pytest.mark.parametrize("hidden_size", [256, 2048])
@pytest.mark.parametrize("clamp_limit", [None, 10.0])
@pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA graph test")
@pytest.mark.skipif(not current_platform.is_cuda_alike(), reason="CUDA/ROCm graph test")
def test_silu_quant_skips_expert_padding_on_graph_replay(
counts: list[int],
group_size: int,
@@ -135,7 +135,10 @@ def test_silu_quant_skips_expert_padding_on_graph_replay(
x *= 16
clean_x = x.clone()
ends = torch.empty(len(counts), device=DEVICE, dtype=torch.int32)
output = torch.empty((capacity, hidden_size), device=DEVICE, dtype=FLOAT8_DTYPE)
output_dtype = (
FLOAT8_DTYPE if scale_format == "packed" else current_platform.fp8_dtype()
)
output = torch.empty((capacity, hidden_size), device=DEVICE, dtype=output_dtype)
def set_counts(values):
positions = []
@@ -187,6 +190,7 @@ def test_silu_quant_skips_expert_padding_on_graph_replay(
output.view(torch.uint8).fill_(91)
graph.replay()
reference_q, reference_s = quantize(x)
assert actual_q.dtype == reference_q.dtype
torch.testing.assert_close(
actual_q.view(torch.uint8)[live],
reference_q.view(torch.uint8)[live],