Skip to content

Update linear -> conv2d int16 for Ethos #14763

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
merged 1 commit into from
Oct 4, 2025
+5 −15
Merged

Update linear -> conv2d int16 for Ethos #14763

Show file tree
Hide file tree
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
6 changes: 3 additions & 3 deletions backends/arm/operators/op_conv2d.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -182,11 +182,11 @@ def define_node(
acc_type = ts.DType.FP32

tosa_graph.addConst(
[1], output.dtype, [input_zp], name=f"{conv2d_output_name}_input_zp"
[1], inputs[0].dtype, [input_zp], name=f"{conv2d_output_name}_input_zp"
)
tosa_graph.addConst(
[1],
output.dtype,
inputs[1].dtype,
weight_zp,
name=f"{conv2d_output_name}_weight_zp",
)
Expand Down Expand Up @@ -269,7 +269,7 @@ def define_node(

# For quantized convolution, rescale the output value back to the same
# integer value domain of the next op. Otherwise return float32 output.
if inputs[0].dtype == ts.DType.INT8 or inputs[0].dtype == ts.DType.INT16:
if output.dtype == ts.DType.INT8 or output.dtype == ts.DType.INT16:
# Get scale_factor from input, weight, and output.
input_scale = input_qparams[0].get_scale_per_tensor() # type: ignore[possibly-undefined] # pyre-ignore [61]
per_channel_quant = input_qparams[1].per_channel # pyre-ignore [61]
Expand Down
14 changes: 2 additions & 12 deletions backends/arm/test/ops/test_linear.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,8 +8,6 @@

from typing import Tuple

import pytest

import torch
from executorch.backends.arm.quantizer.arm_quantizer import (
get_symmetric_a16w8_quantization_config,
Expand Down Expand Up @@ -313,12 +311,8 @@ def test_linear_16a8w_tosa_INT(test_data: torch.Tensor):
pipeline.run()


@common.parametrize("test_data", test_data_rank1_INT | test_data_rank4_INT)
@common.parametrize("test_data", test_data_all_16a8w)
@common.XfailIfNoCorstone300
@pytest.mark.xfail(
reason="Ethos-U55 A16W8 linear: int16 matmul not yet supported; pending backend support or linear->conv1x1 lowering. See: https://github.com/pytorch/executorch/issues/13947",
strict=False,
)
def test_linear_16a8w_u55_INT16(test_data: torch.Tensor):
"""Test linear operation with 16A8W quantization on U55 (16-bit activations, 8-bit weights)"""
test_data, out_features, has_bias, per_channel_quantization = test_data()
Expand Down Expand Up @@ -347,12 +341,8 @@ def test_linear_16a8w_u55_INT16(test_data: torch.Tensor):
pipeline.run()


@common.parametrize("test_data", test_data_rank1_INT | test_data_rank4_INT)
@common.parametrize("test_data", test_data_all_16a8w)
@common.XfailIfNoCorstone320
@pytest.mark.xfail(
reason="Ethos-U55 A16W8 linear: int16 matmul not yet supported; pending backend support or linear->conv1x1 lowering. See: https://github.com/pytorch/executorch/issues/13947",
strict=False,
)
def test_linear_16a8w_u85_INT16(test_data: torch.Tensor):
"""Test linear operation with 16A8W quantization on U85 (16-bit activations, 8-bit weights)"""
test_data, out_features, has_bias, per_channel_quantization = test_data()
Expand Down
Loading