Arm backend: Correct per channel axis for transpose conv (pytorch#17842)

Adjust per-channel weight axis for conv_transpose2d based on group(s).

The corrected axis is propagated to:
- QuantizationSpec
- QAT fake-quant constructors wrapped in PartialWrapper
- Non-QAT observer/fake-quant constructors (via with_args)



cc @digantdesai @SS-JIA @freddan80 @per @zingo @oscarandersson8218
@Sebastian-Larsson @robell

Signed-off-by: Måns Nilsson <mans.nilsson@arm.com>

Author: mansnils

backends/arm/quantizer/quantization_annotator.py

@@ -9,6 +9,7 @@
 
 """
 
+import functools
 import logging
 import operator
 from dataclasses import dataclass, replace
@@ -22,6 +23,7 @@
 from torch._subclasses import FakeTensor
 
 from torch.fx import Node
+from torchao.quantization.pt2e import PartialWrapper
 from torchao.quantization.pt2e.quantizer import (
     annotate_input_qspec_map,
     annotate_output_qspec,
@@ -85,21 +87,51 @@ def _as_list(x):
 
 def _adjust_weight_qspec_for_conv_transpose(node: Node, weight_qspec):
     if (
-        node.target == torch.ops.aten.conv_transpose2d.input
-        and isinstance(weight_qspec, QuantizationSpec)
-        and weight_qspec.qscheme == torch.per_channel_symmetric
-        and weight_qspec.ch_axis != 1
+        node.target != torch.ops.aten.conv_transpose2d.input
+        or not isinstance(weight_qspec, QuantizationSpec)
+        or weight_qspec.qscheme != torch.per_channel_symmetric
     ):
-        return QuantizationSpec(
-            dtype=weight_qspec.dtype,
-            observer_or_fake_quant_ctr=weight_qspec.observer_or_fake_quant_ctr,
-            quant_min=weight_qspec.quant_min,
-            quant_max=weight_qspec.quant_max,
-            qscheme=weight_qspec.qscheme,
-            ch_axis=1,
-            is_dynamic=weight_qspec.is_dynamic,
+        return weight_qspec
+
+    # For now skip axis adjustment for a8w4 per-channel configs (int4 weights).
+    if weight_qspec.quant_min == -7 and weight_qspec.quant_max == 7:
+        return weight_qspec
+
+    groups = 1
+    if len(node.args) > 6 and isinstance(node.args[6], int):
+        groups = node.args[6]
+    expected_axis = 0 if groups != 1 else 1
+    if weight_qspec.ch_axis == expected_axis:
+        return weight_qspec
+
+    observer_or_fake_quant_ctr = weight_qspec.observer_or_fake_quant_ctr
+    # TorchAO PT2e QAT commonly represents the ctor as PartialWrapper(partial(...)).
+    # Rebuild it to update ch_axis while preserving callable_args.
+    if isinstance(observer_or_fake_quant_ctr, PartialWrapper):
+        original_callable_args = dict(observer_or_fake_quant_ctr.callable_args)
+        base_partial = observer_or_fake_quant_ctr.p
+        if isinstance(base_partial, functools.partial):
+            base_keywords = dict(base_partial.keywords or {})
+            base_keywords["ch_axis"] = expected_axis
+            observer_or_fake_quant_ctr = PartialWrapper(
+                functools.partial(base_partial.func, **base_keywords)
+            )
+            observer_or_fake_quant_ctr.callable_args = original_callable_args
+    # Non-QAT observer/fake-quant constructors can be updated via with_args.
+    elif hasattr(observer_or_fake_quant_ctr, "with_args"):
+        observer_or_fake_quant_ctr = observer_or_fake_quant_ctr.with_args(
+            ch_axis=expected_axis
         )
-    return weight_qspec
+
+    return QuantizationSpec(
+        dtype=weight_qspec.dtype,
+        observer_or_fake_quant_ctr=observer_or_fake_quant_ctr,
+        quant_min=weight_qspec.quant_min,
+        quant_max=weight_qspec.quant_max,
+        qscheme=weight_qspec.qscheme,
+        ch_axis=expected_axis,
+        is_dynamic=weight_qspec.is_dynamic,
+    )
 
 
 def _is_ok_for_quantization(

backends/arm/test/ops/test_transpose_conv2d.py

@@ -11,16 +11,21 @@
 from executorch.backends.arm.quantizer.arm_quantizer import (
     get_symmetric_a16w8_quantization_config,
     get_symmetric_a8w4_quantization_config,
+    get_symmetric_quantization_config,
+    TOSAQuantizer,
 )
 from executorch.backends.arm.test import common
 from executorch.backends.arm.test.tester.test_pipeline import (
     EthosU55PipelineINT,
     EthosU85PipelineINT,
     OpNotSupportedPipeline,
+    QuantizationPipeline,
     TosaPipelineFP,
     TosaPipelineINT,
     VgfPipeline,
 )
+from executorch.backends.arm.tosa.specification import TosaSpecification
+from executorch.backends.test.harness.stages.quantize import Quantize
 
 aten_op = "torch.ops.aten.conv_transpose2d.input"
 exir_op = "executorch_exir_dialects_edge__ops_aten_convolution_default"  # No edge transpoe conv
@@ -94,6 +99,21 @@ def forward(self, x):
     for q in [True, False]
 }
 
+test_data_QAT = {
+    "qat_basic": lambda: (
+        TransposeConv2d(
+            in_channels=16,
+            out_channels=4,
+            kernel_size=4,
+            stride=2,
+            padding=1,
+            groups=1,
+        ),
+        True,
+        True,
+    ),
+}
+
 u55_supported_test_data_INT = {
     k: v
     for k, v in test_data_INT.items()
@@ -150,6 +170,29 @@ def test_conv_transpose2d_tosa_INT(test_data):
     pipeline.run()
 
 
+@common.parametrize("test_data", test_data_QAT)
+def test_conv_transpose2d_tosa_INT_qat_per_channel_quantization_pipeline(test_data):
+    model, is_per_channel, is_qat = test_data()
+    inputs = model.get_inputs()
+    quantizer = TOSAQuantizer(TosaSpecification.create_from_string("TOSA-1.0+INT"))
+    quantizer.set_global(
+        get_symmetric_quantization_config(
+            is_per_channel=is_per_channel,
+            is_qat=is_qat,
+        )
+    )
+    pipeline = QuantizationPipeline[input_t](model, inputs, quantizer)
+    pipeline.change_args(
+        "quantize",
+        Quantize(
+            quantizer,
+            quantization_config=quantizer.global_config,
+            is_qat=is_qat,
+        ),
+    )
+    pipeline.run()
+
+
 _a8w4_transpose_conv_xfails = {
     k: "per-channel int4 weight quantization is not supported for transpose conv yet."
     for k in test_data_INT