vikarti.anatra/blt
1
0
Fork 0
mirror of https://github.com/facebookresearch/blt.git synced 2025-02-22 13:02:14 +00:00
Code Issues Actions 8 Packages Projects Releases Wiki Activity

comment + black
Some checks failed
Lint with Black / lint (push) Has been cancelled
Details
Lint with isort / lint (push) Has been cancelled
Details

Browse source
This commit is contained in:
Srini Iyer 2025-02-06 22:14:20 +00:00
parent 30f82211c4
commit ba922695b3
2 changed files with 8 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

8
bytelatent/distributed.py
View file

@ -463,7 +463,13 @@ def parallelize_model(
raise ValueError(f"Invalid fsdp_type: {distributed_args.fsdp_type}")
if distributed_args.selective_activation_checkpointing:
for module in [model.global_transformer, model.local_encoder, model.local_decoder]:
# only works for blt models
# assuming that entropy models will not use checkpointing
for module in [
model.global_transformer,
model.local_encoder,
model.local_decoder,
]:
for i in range(len(module.layers)):
module.layers[i] = checkpoint_wrapper(
module.layers[i],

2
bytelatent/model/local_models.py
View file

@ -179,7 +179,7 @@ class LocalModelBase(nn.Module):
)
if self.patch_embedding_projection is not None:
patch_emb_std = self.dim_patch_emb ** (-0.5)
patch_emb_std = self.dim_patch_emb ** (-0.5)
nn.init.trunc_normal_(
self.patch_embedding_projection.weight,
mean=0.0,