[Performance] Reduce DeepGEMM N dim restriction from 128 to 64 multiplier (#28687)

Signed-off-by: Alexander Matveev <amatveev@redhat.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Alexander Matveev
2025-11-19 15:47:13 -08:00
committed by GitHub
co-authored by mgoin
parent 8e38e99829
commit 3aaa94ac99
3 changed files with 36 additions and 6 deletions
+9 -2
View File
@@ -365,11 +365,18 @@ def should_use_deepgemm_for_fp8_linear(
):
if supports_deep_gemm is None:
supports_deep_gemm = is_deep_gemm_supported()
# Verify DeepGEMM N/K dims requirements
# NOTE: Also synchronized with test_w8a8_block_fp8_deep_gemm_matmul
# test inside kernels/quatization/test_block_fp8.py
N_MULTIPLE = 64
K_MULTIPLE = 128
return (
supports_deep_gemm
and output_dtype == torch.bfloat16
and weight.shape[0] % 128 == 0
and weight.shape[1] % 128 == 0
and weight.shape[0] % N_MULTIPLE == 0
and weight.shape[1] % K_MULTIPLE == 0
)