[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:
co-authored by
Claude Opus 5
parent
9e25706560
commit
be1cb9834b
@@ -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])
|
||||
|
||||
Reference in New Issue
Block a user