Skip to content

[WIP]Adds _weight_int8pack_mm pass for woq-int8 #3061

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
wants to merge 3 commits into
base: main
Choose a base branch
from
Open

[WIP]Adds _weight_int8pack_mm pass for woq-int8 #3061

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
f60e57a
Adds _weight_int8pack_mm pass for woq-int8
xiaowangintel Sep 25, 2025
b5572b8
Adds _weight_int8pack_mm pass for woq-int8
xiaowangintel Sep 25, 2025
39d2971
Update plain_layout.py
xiaowangintel Sep 29, 2025
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
22 changes: 15 additions & 7 deletions torchao/dtypes/uintx/plain_layout.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -253,16 +253,24 @@ def _linear_fp_act_int8_weight_impl(input_tensor, weight_tensor, bias):
# return torch.ops.aten._weight_int8pack_mm(input_tensor.contiguous(), w_vals_int8_t, weight_tensor.tensor_impl.scale)

# per channel int8 weight only quantizated mm
w_vals_int8_t = weight_tensor.tensor_impl.int_data.t()
w_vals_int8 = weight_tensor.tensor_impl.int_data
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

this is a code path for int8 cuda as well I think, changing it has a risk of perf regressions

also this is the older stack, I'd suggest to migrate first, WIP here for int8 + plain layout: #3038

scale = weight_tensor.tensor_impl.scale
m = torch.mm(
input_tensor.reshape(-1, input_tensor.shape[-1]),
w_vals_int8_t.to(input_tensor.dtype),
)
y = m * scale.to(m.dtype)
try:
y = torch.ops.aten._weight_int8pack_mm(
input_tensor.reshape(-1, input_tensor.shape[-1]),
w_vals_int8,
scale.to(input_tensor.dtype),
)
except Exception:
w_vals_int8_t = w_vals_int8.t()
m = torch.mm(
input_tensor.reshape(-1, input_tensor.shape[-1]),
w_vals_int8_t.to(input_tensor.dtype),
)
y = m * scale.to(m.dtype)
y = y.reshape(*input_tensor.shape[:-1], y.shape[-1])
if bias is not None:
y += bias.to(m.dtype)
y += bias.to(input_tensor.dtype)
return y


Expand Down
Loading