[Bugfix][Quantization] Fix MXFP8 startup crash on layers below mm_mxfp8 shape limits (#54223)

Signed-off-by: samuelkim7 <samuelmwkim@gmail.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Samuel Kim
2026-09-25 02:54:11 +00:00
committed by GitHub
co-authored by mgoin
parent 6936e77ba4
commit f55419c43c
16 changed files with 221 additions and 57 deletions
@@ -408,7 +408,9 @@ def test_fp8_block_marlin_padded_round_trip(shape):
_gpu_marlin_unsupported() or not is_fp8_marlin_supported(),
reason="FP8 Marlin is not supported on this GPU type.",
)
@pytest.mark.parametrize("shape", [(200, 288), (4640, 512)])
@pytest.mark.parametrize(
"shape", [(32, 256), (96, 256), (130, 256), (200, 288), (4640, 512)]
)
def test_mxfp8_marlin_padded_round_trip(shape):
"""MXFP8 exercises the e8m0 scale path, where padded 0.0 scales clamp to
2^-127 instead of zero and must still contribute nothing."""