Skip to content

Document intentional zero-init of RMSNorm scale #481

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

Open
meemoprasad wants to merge 4 commits into
base: master
Choose a base branch
from
Open

Document intentional zero-init of RMSNorm scale #481

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
b6f1be2
Fix RMSNorm scale zero-initialization
meemoprasad Sep 2, 2026
93fb1a7
Initialize scale parameter to zeros
meemoprasad Oct 21, 2026
ea196c8
Document intentional zero-init of RMSNorm scale (flax)
meemoprasad Oct 21, 2026
81a1b77
Document intentional zero-init of RMSNorm scale (torch)
meemoprasad Oct 21, 2026
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
8 changes: 7 additions & 1 deletion src/timesfm/flax/normalization.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,7 +27,13 @@


class RMSNorm(nnx.Module):
"""RMS normalization."""
"""RMS normalization.

The scale is initialized to zeros on purpose: a freshly constructed layer
outputs zeros, so at init the residual connection carries the signal.
Loading a checkpoint overwrites the scale. See
`test_zero_scale_produces_zeros` in tests/test_torch_layers.py.
"""

__data__ = ("scale",)

Expand Down
8 changes: 7 additions & 1 deletion src/timesfm/torch/normalization.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,7 +19,13 @@


class RMSNorm(nn.Module):
"""RMS normalization."""
"""RMS normalization.

The scale is initialized to zeros on purpose: a freshly constructed layer
outputs zeros, so at init the residual connection carries the signal.
Loading a checkpoint overwrites the scale. See
`test_zero_scale_produces_zeros` in tests/test_torch_layers.py.
"""

def __init__(
self,
Expand Down
Loading