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

Jump to bottom

Fix (gptq): fix for depthwise act_order #688

Merged
merged 1 commit into from
Jul 28, 2023
Merged

Fix (gptq): fix for depthwise act_order #688

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
10 changes: 6 additions & 4 deletions src/brevitas/graph/gptq.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -381,16 +381,18 @@ def single_layer_update(self, percdamp=.01):
if self.groups > 1:
# In case of depthwise convs, each weight matrix interacts with only
# part of the input values, thus with only one of the hessian matrix
for ii in range(self.groups):
for ii, perm in enumerate(permutation_list):
weight_block[ii, i:] -= error[ii] * h_inv_block[ii, i, i:]
# We need to update the original weights
weight[ii, perm[i1:i2][i:]] = weight_block[ii, i:].to(dtype)
else:
perm = permutation_list[0]
weight_block[:, i:] -= error.unsqueeze(1).matmul(
h_inv_block[0, i, i:].unsqueeze(0))
# We need to update the original weights
weight[:, perm[i1:i2][i:]] = weight_block[:, i:].to(dtype)
error_block[:, i] = error

# We need to update the original weights
weight[:, perm[i1:i2][i:]] = weight_block[:, i:].to(dtype)

if self.groups > 1:
# In case of depthwise convs, each weight matrix interacts with only
# part of the input values, thus with only one of the hessian matrix
Expand Down
Loading