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

[Bugfix] Fix empty (nullptr) channelwise scales when loading wNa16 using compressed tensors #6798

Merged
merged 1 commit into from
Jul 25, 2024
+8 −3
Merged

[Bugfix] Fix empty (nullptr) channelwise scales when loading wNa16 using compressed tensors #6798

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
11 changes: 8 additions & 3 deletions ...model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,7 +54,12 @@ def create_weights(self, layer: torch.nn.Module, input_size: int,
output_size_per_partition = sum(output_partition_sizes)

# If group_size is -1, we are in channelwise case.
group_size = input_size if self.group_size == -1 else self.group_size
channelwise = (self.group_size == -1)
group_size = input_size if channelwise else self.group_size
row_parallel = (input_size != input_size_per_partition)
# In the case of channelwise quantization, we need to replicate the
# scales across all gpus.
partition_scales = (row_parallel and not channelwise)

verify_marlin_supports_shape(
output_size_per_partition=output_size_per_partition,
Expand All @@ -65,8 +70,8 @@ def create_weights(self, layer: torch.nn.Module, input_size: int,
weight_scale_dim = None
scales_and_zp_size = input_size // group_size

if (input_size != input_size_per_partition
and self.group_size is not None):
if partition_scales:
assert input_size_per_partition % group_size == 0
weight_scale_dim = 1
scales_and_zp_size = input_size_per_partition // group_size

Expand Down
Loading