Skip to content

Remove jumprelu_threshold validation rule #32

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

Merged
curt-tigges merged 1 commit into from
Jun 2, 2025
Merged

Remove jumprelu_threshold validation rule #32

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
1 change: 0 additions & 1 deletion clt/config/clt_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,7 +40,6 @@ def __post_init__(self):
assert self.num_features > 0, "Number of features must be positive"
assert self.num_layers > 0, "Number of layers must be positive"
assert self.d_model > 0, "Model dimension must be positive"
assert self.jumprelu_threshold > 0, "JumpReLU threshold must be positive"
valid_norm_methods = ["auto", "estimated_mean_std", "none"]
assert (
self.normalization_method in valid_norm_methods
Expand Down
4 changes: 4 additions & 0 deletions clt/models/theta.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,6 +48,10 @@ def __init__(
self.rank = dist.get_rank(process_group)

if self.config.activation_fn == "jumprelu":
if self.config.jumprelu_threshold == 0:
logger.warning(
f"Rank {self.rank}: jumprelu_threshold is 0, expecting to load log_threshold from checkpoint."
)
initial_threshold_val = torch.ones(
config.num_layers, config.num_features, device=self.device, dtype=self.dtype
) * torch.log(torch.tensor(config.jumprelu_threshold, device=self.device, dtype=self.dtype))
Expand Down
Loading