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

Fix: Removed errors for n-dimensional gradients. #65

Open
wants to merge 2 commits into
base: master
Choose a base branch
from
Open

Fix: Removed errors for n-dimensional gradients. #65

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9248f9c
added support for 4D gradient matrices
Aug 1, 2024
d7c80a4
merged changes from upstream
gslama12 Oct 16, 2024
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
19 changes: 15 additions & 4 deletions galore_torch/galore_projector.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,8 +8,16 @@ def __init__(self, rank, verbose=False, update_proj_gap=200, scale=1.0, proj_typ
self.scale = scale
self.ortho_matrix = None
self.proj_type = proj_type
self._original_shape = None

def project(self, full_rank_grad, iter):
# Reshape Nd tensor to 2d
self._original_shape = full_rank_grad.shape
if len(self._original_shape) > 2:
full_rank_grad = full_rank_grad.view(self._original_shape[0], -1) # Flatten to 2d
elif len(self._original_shape) == 1:
full_rank_grad = full_rank_grad.view(1, -1) # Reshape 1d to 2d

if self.proj_type == 'std':
if full_rank_grad.shape[0] >= full_rank_grad.shape[1]:
if self.ortho_matrix is None or iter % self.update_proj_gap == 0:
Expand Down Expand Up @@ -61,10 +69,10 @@ def project_back(self, low_rank_grad):
elif self.proj_type == 'full':
full_rank_grad = torch.matmul(self.ortho_matrix[0].to(low_rank_grad.device.type), low_rank_grad) @ self.ortho_matrix[1].to(low_rank_grad.device.type)


full_rank_grad = full_rank_grad.view(self._original_shape) # Restore original dimensions from 2d tensor
return full_rank_grad * self.scale


# svd decomposition
def get_orthogonal_matrix(self, weights, rank, type):
module_params = weights
Expand All @@ -77,17 +85,20 @@ def get_orthogonal_matrix(self, weights, rank, type):
else:
float_data = True
matrix = module_params.data

U, s, Vh = torch.linalg.svd(matrix, full_matrices = False)

#make the smaller matrix always to be orthogonal matrix
if type=='right':
A = U[:, :rank] @ torch.diag(s[:rank])
B = Vh[:rank, :]

if not float_data:
B = B.to(original_device).type(original_type)
return B
elif type=='left':
A = U[:, :rank]
B = torch.diag(s[:rank]) @ Vh[:rank, :]
if not float_data:
A = A.to(original_device).type(original_type)
return A
Expand Down