Update on "[quant][pt2e][xnnpack] XNNPACKQuantizer skip quantization for input and output to workaround histogram observer problem"

jerryzh168 · jerryzh168 · commit 44d31956231d · 2023-12-01T10:06:05.000-08:00
Summary:
att, this is because histogram observer does not work for a corner case in mobilebert (observing a scalar tensor of float32 max value)
because histc operator errors out when the value is larger than certain number

Test Plan:
python test/test_quantization.py -k test_mul_float32_max

Reviewers:

Subscribers:

Tasks:

Tags:

[ghstack-poisoned]
diff --git a/torch/ao/quantization/quantizer/xnnpack_quantizer_utils.py b/torch/ao/quantization/quantizer/xnnpack_quantizer_utils.py
@@ -13,7 +13,6 @@
     get_aten_graph_module,
 )
 from torch.ao.quantization.quantizer import (
-    FixedQParamsQuantizationSpec,
     QuantizationAnnotation,
     QuantizationSpec,
     QuantizationSpecBase,
@@ -617,9 +616,7 @@ def _annotate_adaptive_avg_pool2d(
     return annotated_partitions
 
 
-def _is_input_large_scalar(
-    node: Node, gm: torch.fx.GraphModule
-):
+def _is_input_large_scalar(node: Node, gm: torch.fx.GraphModule):
     """Check if input is a large scalar value. So that we can skip quantization for the node
     since histc op (in HistogramObserver) only works for values up to certain upper bound
     """
