Skip to content

fix(causal): tolerate minor rho proxy overshoot #3

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
DaviBonetto merged 1 commit into from
Mar 17, 2026
Merged

fix(causal): tolerate minor rho proxy overshoot #3

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
14 changes: 12 additions & 2 deletions scripts/run_causal_intervention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -183,10 +183,20 @@ def hook(module, args, output):
)
state["printed"] = True

if torch.any(rho_tokens_after > target_rho_tensor + 1e-6):
# The rho proxy is approximate, so keep a small numerical margin.
rho_margin = 1e-3
max_rho_after = float(rho_tokens_after.max().item())
if max_rho_after > float(target_rho_tensor.item()) + rho_margin:
raise RuntimeError(
f"Layer {layer_idx} clamp failed: "
f"rho_after={rho_after['max']:.6f} > target={float(target_rho_tensor.item()):.6f}"
f"rho_after={max_rho_after:.6f} > "
f"target={float(target_rho_tensor.item()):.6f} "
f"(margin={rho_margin:.1e})"
)
if debug and not state["printed"] and max_rho_after > float(target_rho_tensor.item()):
print(
f"[layer {layer_idx}] rho_after overshoot "
f"{max_rho_after:.6f} within margin={rho_margin:.1e}"
)

dt_raw_clamped = dt_raw_work.clone()
Expand Down
Loading