Skip to content

Commit 13b3a75

Browse files
authored
[KERNELS] Fix swizzling contiguity (#7582)
Because `(ab)^{-1] = b^{-1}a^{-1}`.
1 parent 984b694 commit 13b3a75

File tree

1 file changed

+2
-2
lines changed
  • python/triton_kernels/triton_kernels/tensor_details/layout_details

1 file changed

+2
-2
lines changed

python/triton_kernels/triton_kernels/tensor_details/layout_details/hopper_scale.py

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -34,7 +34,7 @@ def swizzle_data(self, data):
3434
data = data.flatten(-3, -2)
3535
assert data.shape[-2] == M // 32
3636
assert data.shape[-1] == K * 32
37-
data = data.transpose(-1, -2).contiguous()
37+
data = data.contiguous().transpose(-1, -2)
3838
return data
3939

4040
def unswizzle_data(self, data):
@@ -46,7 +46,7 @@ def unswizzle_data(self, data):
4646
perm = list(range(b)) + [b + p for p in perm]
4747
data = data.permute(*perm)
4848
data = data.reshape(*batch, M * 32, K // 32)
49-
data = data.transpose(-1, -2).contiguous()
49+
data = data.contiguous().transpose(-1, -2)
5050
return data
5151

5252
def swizzle_block_shape(self, block_shape):

0 commit comments

Comments
 (0)