Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Adding Support for Non-Parametric Layernorm #1332

Closed
wants to merge 6 commits into from
Closed

Adding Support for Non-Parametric Layernorm #1332

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
a6a002f
Added non_parametric_layernorm
aflah02 Dec 16, 2024
3d8af16
Added non_parametric_layernorm to args
aflah02 Dec 16, 2024
cfc73c2
Create 7B.yml\
aflah02 Dec 16, 2024
8d52c92
Create 1B.yml
aflah02 Dec 16, 2024
0af74c0
Merge branch 'EleutherAI:main' into olmo-support
aflah02 Jan 4, 2025
b0fed23
Delete configs/OLMo directory
aflah02 Jan 4, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 6 additions & 0 deletions megatron/model/norms.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,6 +33,12 @@ def get_norm(neox_args):
norm = MixedFusedLayerNorm
else:
norm = LayerNorm
elif neox_args.norm == "non_parametric_layernorm":
eps = neox_args.layernorm_epsilon
if neox_args.layernorm_fusion:
raise ValueError(f"neox_args.layernorm_fusion not supported for non_parametric_layernorm")
else:
norm = LayerNorm(elementwise_affine=False, bias=False)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This leads to an error due to initializing torch.nn.LayerNorm without normalized_shape.
Also, get_norm should not return an initialized object.
#1338 resolves this by creating a separate NonParametricLayernorm class.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks!! I'll close this
I totally forgot you need to return a non-initialized object

elif neox_args.norm == "scalenorm":
eps = neox_args.scalenorm_epsilon
norm = ScaleNorm
Expand Down
4 changes: 2 additions & 2 deletions megatron/neox_arguments/neox_args.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -168,10 +168,10 @@ class NeoXArgsModel(NeoXArgsTemplate):
"""

norm: Literal[
"layernorm", "rmsnorm", "scalenorm", "te_rmsnorm", "te_layernorm"
"layernorm", "rmsnorm", "non_parametric_layernorm", "scalenorm", "te_rmsnorm", "te_layernorm"
] = "layernorm"
"""
Normalization layer to use. Choose from "layernorm", "rmsnorm", "scalenorm", "te_rmsnorm", "te_layernorm".
Normalization layer to use. Choose from "layernorm", "rmsnorm", "non_parametric_layernorm", "scalenorm", "te_rmsnorm", "te_layernorm".
"""

layernorm_fusion: bool = False
Expand Down