[Perf] Create TMA-aligned input scale tensor for DeepGemm on Hopper (#32619)
Signed-off-by: Xin Yang <xyangx@amazon.com>
This commit is contained in:
@@ -340,6 +340,11 @@ def _align(x: int, y: int) -> int:
|
||||
return cdiv(x, y) * y
|
||||
|
||||
|
||||
# Taken from https://github.com/deepseek-ai/DeepGEMM/blob/v2.1.1/csrc/utils/math.hpp#L19
|
||||
def get_tma_aligned_size(x: int, element_size: int):
|
||||
return _align(x, 16 // element_size)
|
||||
|
||||
|
||||
DEFAULT_BLOCK_SIZE = [128, 128]
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user