[HotFix]Fix dtype mismatch in nsa indexer on AMD device (#17518)
This commit is contained in:
@@ -186,7 +186,7 @@ class Indexer(MultiPlatformOp):
|
|||||||
self.hidden_size,
|
self.hidden_size,
|
||||||
self.n_heads,
|
self.n_heads,
|
||||||
bias=False,
|
bias=False,
|
||||||
params_dtype=torch.bfloat16,
|
params_dtype=torch.bfloat16 if _is_cuda else torch.float32,
|
||||||
prefix=add_prefix("weights_proj", prefix),
|
prefix=add_prefix("weights_proj", prefix),
|
||||||
)
|
)
|
||||||
self.k_norm = LayerNorm(self.head_dim, dtype=torch.float32)
|
self.k_norm = LayerNorm(self.head_dim, dtype=torch.float32)
|
||||||
|
|||||||
Reference in New Issue
Block a user