[amd][gptoss] Perf gain because of block alignment (#28024)

Signed-off-by: Smit Kadvani <smit.kadvani@gmail.com>
Co-authored-by: Smit Shaileshbhai Kadvani <kadvani@meta.com>
This commit is contained in:
smit kadvani
2025-11-07 05:27:42 +00:00
committed by GitHub
co-authored by Smit Shaileshbhai Kadvani
parent c0a4b95d64
commit 11fd69dd54
2 changed files with 13 additions and 2 deletions
@@ -7,6 +7,7 @@ import torch
from vllm.logger import init_logger
from vllm.platforms import current_platform
from vllm.triton_utils import triton
from vllm.utils.torch_utils import direct_register_custom_op, is_torch_equal_or_newer
logger = init_logger(__name__)
@@ -99,6 +100,14 @@ def _can_support_mxfp4(
)
def get_padding_alignment():
return (
256
if triton.runtime.driver.active.get_current_target().arch in ("gfx950",)
else 128
)
def _dequant_mxfp4(
x: torch.Tensor, scale: torch.Tensor, float_dtype: torch.dtype
) -> torch.Tensor: