Skip to content

Arm backend: Add support for QAT+per-channel combo #13511

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
zingo merged 1 commit into from
Aug 19, 2025
+15 −9
Merged

Arm backend: Add support for QAT+per-channel combo #13511

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
10 changes: 6 additions & 4 deletions backends/arm/quantizer/arm_quantizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -101,18 +101,20 @@ def get_symmetric_quantization_config(
weight_observer_or_fake_quant_ctr: ObserverOrFakeQuantizeConstructor = (
MinMaxObserver
)

# Determine the right observer/fake-quant constructor
if is_qat:
# Set plain fake-quant with true min/max
weight_observer_or_fake_quant_ctr = FakeQuantize
if is_per_channel:
weight_observer_or_fake_quant_ctr = PerChannelMinMaxObserver
Copy link
Contributor

@digantdesai digantdesai Aug 20, 2025

Choose a reason for hiding this comment

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

Don't we need a FakeQuantize variant of this observer with qat?

Copy link
Collaborator Author

@oscarandersson8218 oscarandersson8218 Aug 20, 2025

Choose a reason for hiding this comment

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

Yes, you're right. Will look into a fix and proper testing of QAT.

Copy link
Collaborator Author

@oscarandersson8218 oscarandersson8218 Sep 5, 2025

Choose a reason for hiding this comment

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

Fixed here #14001

else:
# Set plain fake-quant with true min/max
weight_observer_or_fake_quant_ctr = FakeQuantize
else:
# PTQ: set min/max observer
weight_observer_or_fake_quant_ctr = (
PerChannelMinMaxObserver if is_per_channel else MinMaxObserver
)

extra_args = {"eps": 2**-12}

weight_quantization_spec = QuantizationSpec(
dtype=torch.int8,
quant_min=weight_qmin,
Expand Down
14 changes: 9 additions & 5 deletions backends/arm/test/misc/test_bn_relu_folding_qat.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,13 +40,17 @@ def forward(self, x: torch.Tensor):


models = {
"conv_bn_relu": ConvModule(batch_norm=True),
"conv_relu": ConvModule(batch_norm=False),
# name : (model, is_per_channel)
"conv_bn_relu_per_channel": (ConvModule(batch_norm=True), True),
"conv_relu_per_channel": (ConvModule(batch_norm=False), True),
"conv_bn_relu_per_tensor": (ConvModule(batch_norm=True), False),
"conv_relu_per_tensor": (ConvModule(batch_norm=False), False),
}


@common.parametrize("model", models)
def test_qat_tosa_INT(model: torch.nn.Module):
@common.parametrize("test_data", models)
def test_qat_tosa_INT(test_data):
model, per_channel = test_data
pipeline = TosaPipelineINT[input_t1](model, model.test_data, [], [], qtol=1)
tosa_version = conftest.get_option("tosa_version")
tosa_profiles = {
Expand All @@ -59,7 +63,7 @@ def test_qat_tosa_INT(model: torch.nn.Module):
Quantize(
quantizer=quantizer,
quantization_config=get_symmetric_quantization_config(
is_qat=True, is_per_channel=False
is_qat=True, is_per_channel=per_channel
),
is_qat=True,
),
Expand Down
Loading