[Kernel] Vectorized fp32 moe_sum reduction and support any topk (#46643)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
Michael Goin
2026-06-25 14:02:28 -07:00
committed by GitHub
co-authored by Claude
parent 2a6f8f0c05
commit e8c24a7695
2 changed files with 175 additions and 49 deletions
+16 -4
View File
@@ -1243,15 +1243,27 @@ def test_batched_moe_align_block_size_opcheck():
)
# topk=8 covers topk > 4; k=511 covers the non-vectorized scalar path. The
# layouts exercise contiguous input plus the two non-contiguous cases: a
# transpose (strided hidden -> scalar gather) and a topk-slice (hidden still
# contiguous -> vectorized).
@pytest.mark.parametrize("m", [1, 33, 222])
@pytest.mark.parametrize("topk", TOP_KS)
@pytest.mark.parametrize("topk", [*TOP_KS, 8])
@pytest.mark.parametrize("k", [128, 511, 1024])
@pytest.mark.parametrize("dtype", [torch.float32, torch.bfloat16])
def test_moe_sum(m: int, topk: int, k: int, dtype: torch.dtype):
input = torch.randn((m, topk, k), device="cuda", dtype=dtype)
@pytest.mark.parametrize("layout", ["contig", "transpose", "slice"])
def test_moe_sum(m: int, topk: int, k: int, dtype: torch.dtype, layout: str):
if layout == "transpose":
input = torch.randn((m, k, topk), device="cuda", dtype=dtype).transpose(1, 2)
elif layout == "slice":
input = torch.randn((m, 2 * topk, k), device="cuda", dtype=dtype)[:, ::2, :]
else:
input = torch.randn((m, topk, k), device="cuda", dtype=dtype)
assert input.is_contiguous() == (layout == "contig")
actual = torch.empty((m, k), device="cuda", dtype=dtype)
expected = input.sum(dim=1)
# Reduction accumulates in fp32.
expected = input.float().sum(dim=1).to(dtype)
torch.ops._moe_C.moe_sum(input, actual)
torch.testing.assert_close(actual, expected, atol=2e-2, rtol=0)