C3x SM100 FP8 blockwise - Pad activation scales to multiple of 4 (#57377)

Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
Co-authored-by: Codex <noreply@openai.com>
This commit is contained in:
Thien Tran
2026-09-18 08:39:27 +00:00
committed by GitHub
co-authored by Codex
parent 95d723700b
commit 8da75d6190
3 changed files with 39 additions and 6 deletions
+7 -2
View File
@@ -162,11 +162,16 @@ def test_w8a8_block_fp8_matmul(M, N, K, block_size, out_dtype, seed):
@pytest.mark.skipif(
not current_platform.is_cuda(), reason="CUTLASS only supported on CUDA platform."
)
@pytest.mark.parametrize(
# 65/66/67 cover all M%4 residue classes above the SM100 swapAB
# threshold (m <= 64); 1026 crosses multiple 128-row SF atoms.
"M",
[32, 65, 66, 67, 1026],
)
@torch.inference_mode()
def test_w8a8_block_fp8_cutlass_matmul():
def test_w8a8_block_fp8_cutlass_matmul(M):
# Test simple case where weight.shape % 128 != 0,
# like in DSV3 kv_a_proj_with_mqa
M = 32
N = 576
K = 7168
block_size = [128, 128]