deftdawg/bitsandbytes-rocm
1
0
Fork 0
forked from mrq/bitsandbytes-rocm
Code Pull Requests Activity

add device and dtype parameters to StableEmbedding

Browse Source
This commit is contained in:
Victor Nova 2022-11-04 14:05:30 -07:00
parent 1efb87d89d
commit 62d39a237c
No known key found for this signature in database
1 changed files with 9 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
bitsandbytes/nn/modules.py
View File

@ -38,6 +38,8 @@ class StableEmbedding(torch.nn.Embedding):
scale_grad_by_freq: bool = False,
sparse: bool = False,
_weight: Optional[Tensor] = None,
device=None,
dtype=None,
) -> None:
super(StableEmbedding, self).__init__(
num_embeddings,
@ -48,8 +50,10 @@ class StableEmbedding(torch.nn.Embedding):
scale_grad_by_freq,
sparse,
_weight,
device,
dtype,
)
self.norm = torch.nn.LayerNorm(embedding_dim)
self.norm = torch.nn.LayerNorm(embedding_dim, device=device)
GlobalOptimManager.get_instance().register_module_override(
self, "weight", {"optim_bits": 32}
)
@ -81,7 +85,10 @@ class StableEmbedding(torch.nn.Embedding):
self.sparse,
)
return self.norm(emb)
# always apply layer norm in full precision
emb = emb.to(torch.get_default_dtype())
return self.norm(emb).to(self.weight.dtype)
class Embedding(torch.nn.Embedding):