[Bugfix] Add contiguous call inside rope kernel wrapper (#17091)

Signed-off-by: 苏政渊 <suzhengyuan@moonshot.cn>
Co-authored-by: 苏政渊 <suzhengyuan@moonshot.cn>
This commit is contained in:
Zhengyuan Su (苏政渊)
2025-04-28 19:24:07 -07:00
committed by GitHub
co-authored by 苏政渊
parent 165cb56329
commit 17eb306fcc
2 changed files with 17 additions and 7 deletions
+14 -3
View File
@@ -158,8 +158,13 @@ def rotary_embedding(
cos_sin_cache: torch.Tensor,
is_neox: bool,
) -> None:
torch.ops._C.rotary_embedding(positions, query, key, head_size,
cos_sin_cache, is_neox)
# TODO: Remove this contiguous call when the kernel is updated to support tensor slices
query_contiguous = query.contiguous()
key_contiguous = key.contiguous()
torch.ops._C.rotary_embedding(positions, query_contiguous, key_contiguous,
head_size, cos_sin_cache, is_neox)
query.copy_(query_contiguous)
key.copy_(key_contiguous)
def batched_rotary_embedding(positions: torch.Tensor, query: torch.Tensor,
@@ -167,9 +172,15 @@ def batched_rotary_embedding(positions: torch.Tensor, query: torch.Tensor,
cos_sin_cache: torch.Tensor, is_neox: bool,
rot_dim: int,
cos_sin_cache_offsets: torch.Tensor) -> None:
torch.ops._C.batched_rotary_embedding(positions, query, key, head_size,
# TODO: Remove this contiguous call when the kernel is updated to support tensor slices
query_contiguous = query.contiguous()
key_contiguous = key.contiguous()
torch.ops._C.batched_rotary_embedding(positions, query_contiguous,
key_contiguous, head_size,
cos_sin_cache, is_neox, rot_dim,
cos_sin_cache_offsets)
query.copy_(query_contiguous)
key.copy_(key_contiguous)
# layer norm ops