[Kernel] Optional Q-norm in fused DSv4 MLA epilogue; group_size=32 for packed FP8 quant (#56215)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Yongye Zhu
2026-09-10 00:35:59 -07:00
committed by GitHub
co-authored by Claude Opus 5
parent 9e25706560
commit be1cb9834b
6 changed files with 267 additions and 148 deletions
@@ -80,6 +80,11 @@ def test_per_token_group_quant_fp8(
# Larger shapes with padding
(127, 7168, 128),
(253, 640, 128),
(1, 32, 32),
(3, 64, 32),
(7, 160, 32),
(32, 512, 32),
(127, 4096, 32),
],
)
@pytest.mark.parametrize("poisoned_scales", [False, True])