[Performance] Move the contiguous to torch compile region (#13199)
This commit is contained in:
@@ -1481,6 +1481,8 @@ class MRotaryEmbedding(RotaryEmbedding):
|
|||||||
num_tokens = positions.shape[-1]
|
num_tokens = positions.shape[-1]
|
||||||
cos_sin = self.cos_sin_cache[positions]
|
cos_sin = self.cos_sin_cache[positions]
|
||||||
cos, sin = cos_sin.chunk(2, dim=-1)
|
cos, sin = cos_sin.chunk(2, dim=-1)
|
||||||
|
cos = cos.contiguous()
|
||||||
|
sin = sin.contiguous()
|
||||||
query_shape = query.shape
|
query_shape = query.shape
|
||||||
key_shape = key.shape
|
key_shape = key.shape
|
||||||
if positions.ndim == 2:
|
if positions.ndim == 2:
|
||||||
|
|||||||
Reference in New Issue
Block a user