Skip to content

Arm backend: Change QAT weight observer #11787

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
Jun 18, 2025
Merged

Arm backend: Change QAT weight observer #11787

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
23 changes: 12 additions & 11 deletions backends/arm/quantizer/arm_quantizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,7 +38,6 @@
HistogramObserver,
MinMaxObserver,
MovingAverageMinMaxObserver,
MovingAveragePerChannelMinMaxObserver,
ObserverOrFakeQuantizeConstructor,
PerChannelMinMaxObserver,
PlaceholderObserver,
Expand Down Expand Up @@ -95,24 +94,26 @@ def get_symmetric_quantization_config(
**extra_args,
),
)

# Setup quantization config for weights
weight_qscheme = (
torch.per_channel_symmetric if is_per_channel else torch.per_tensor_symmetric
)
weight_observer_or_fake_quant_ctr: ObserverOrFakeQuantizeConstructor = (
MinMaxObserver
)
# Determine the right observer/fake-quant constructor
if is_qat:
# TODO: qat + per channel?
weight_observer_or_fake_quant_ctr = FusedMovingAvgObsFakeQuantize
elif is_per_channel:
weight_observer_or_fake_quant_ctr = PerChannelMinMaxObserver
# 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}

extra_args: Dict[str, Any] = {"eps": 2**-12}
if is_qat:
if weight_qscheme == torch.per_tensor_symmetric:
extra_args["observer"] = MovingAverageMinMaxObserver
else:
extra_args["observer"] = MovingAveragePerChannelMinMaxObserver # type: ignore[dict-item]
weight_quantization_spec = QuantizationSpec(
dtype=torch.int8,
quant_min=weight_qmin,
Expand Down
Loading