Skip to content

[ET-VK][qconv] Fix depthwise weight_sums sum dimension #17504

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
meta-codesync merged 3 commits into from
Feb 20, 2026
Merged

[ET-VK][qconv] Fix depthwise weight_sums sum dimension #17504

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
eccc975
[ET-VK][qconv] Fix depthwise weight_sums sum dimension
Feb 17, 2026
a2b9fcf
Update on "[ET-VK][qconv] Fix depthwise weight_sums sum dimension"
Feb 18, 2026
8bf97f2
Update on "[ET-VK][qconv] Fix depthwise weight_sums sum dimension"
Feb 19, 2026
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: 11 additions & 3 deletions backends/vulkan/patterns/quantized_convolution.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -215,9 +215,17 @@ def make_q8ta_conv2d_custom_op(
with graph_module.graph.inserting_before(first_graph_node):
qweight_tensor_name = utils.get_tensor_name(ep, match.weight_node)
# Pre-compute the weight sums which are needed to apply activation zero point
# when using integer accumulation. For the reshaped 2D weight matrix (IC_per_group * H * W, OC),
# sum over dimension 0 to get sums per output channel
sum_per_output_channel = weight_tensor.sum(dim=1).to(torch.int32).contiguous()
# when using integer accumulation. Sum all weight elements per output channel.
if is_depthwise_conv:
# weight_tensor shape is (H, W, OC); sum over spatial dims (H, W)
sum_per_output_channel = (
weight_tensor.sum(dim=(0, 1)).to(torch.int32).contiguous()
)
else:
# weight_tensor shape is (OC, H*W*IC_per_group); sum over dim 1
sum_per_output_channel = (
weight_tensor.sum(dim=1).to(torch.int32).contiguous()
)
sums_name = qweight_tensor_name + "_sums"
# Sanitize the name
sums_name = sums_name.replace(".", "_")
Expand Down
Loading