Skip to content

Run weighted_rigid_align's Kabsch SVD in float32 (fix bf16 crash) #283

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
dguo8412 wants to merge 1 commit into
base: production
Choose a base branch
from
Open

Run weighted_rigid_align's Kabsch SVD in float32 (fix bf16 crash) #283

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
6 changes: 4 additions & 2 deletions src/foundry/utils/alignment.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,7 +57,9 @@ def weighted_rigid_align(
# Computation of the covariance matrix
C = torch.einsum("bji,bjk->bik", w_resolved[..., None] * X_gt_resolved, X_resolved)

U, S, V = torch.linalg.svd(C)
# SVD has no bf16/half CUDA kernel and is numerically sensitive; compute the
# rotation in float32 and cast it back to the input dtype below.
U, S, V = torch.linalg.svd(C.float())

R = U @ V
B, _, _ = X_L.shape
Expand All @@ -71,7 +73,7 @@ def weighted_rigid_align(

det = torch.linalg.det(R)
F[..., -1, -1] = torch.sign(det)
R = U @ F @ V
R = (U @ F @ V).to(X_gt_L.dtype)

X_gt_L = X_gt_L - u_X_gt.unsqueeze(-2)
X_align_L = X_gt_L @ R + u_X.unsqueeze(-2)
Expand Down