Fix (test): custom a2q quantizer for ort tests

tests/brevitas_ort/common.py

@@ -36,14 +36,22 @@
 FLOAT_TOLERANCE = 1e-6
 KERNEL_SIZE = 1  # keep float error during fake-quantization under control
 BIT_WIDTHS = range(2, 9)
+ACCUMULATOR_BIT_WIDTH_FOR_TESTS = 16
+
+
+# For testing purpose, we create a custom quantizer with a reduced bitwidth for the accumulator
+class A2QWeightQuantizerForTests(Int8AccumulatorAwareWeightQuant):
+    accumulator_bit_width = ACCUMULATOR_BIT_WIDTH_FOR_TESTS
+
+
 WBIOL_QUANTIZERS = {
     'asymmetric_per_tensor_float':
         (ShiftedUint8WeightPerTensorFloat, ShiftedUint8ActPerTensorFloat),
     'symmetric_per_tensor_float': (Int8WeightPerTensorFloat, Int8ActPerTensorFloat),
     'asymmetric_per_channel_float':
         (ShiftedUint8WeightPerChannelFloat, ShiftedUint8ActPerTensorFloat),
     'symmetric_per_channel_float': (Int8WeightPerChannelFloat, Int8ActPerTensorFloat),
-    'a2q': (Int8AccumulatorAwareWeightQuant, Int8ActPerTensorFloat),
+    'a2q': (A2QWeightQuantizerForTests, Int8ActPerTensorFloat),
     'symmetric_per_tensor_fixed_point': (Int8WeightPerTensorFixedPoint, Int8ActPerTensorFixedPoint),
     'symmetric_per_channel_fixed_point':
         (Int8WeightPerChannelFixedPoint, Int8ActPerTensorFixedPoint)}