[ONNX] Fix sporadic results in BC (#3081)

### Changes

1. This PR addresses an issue using `ONNXRuntime==1.19.2` where a tensor
used as both an input and output in a model shares the same memory. This
causes unexpected behavior: updating the input tensor inadvertently
modifies the statistics data due to memory overlap.
The issue was confirmed by calling
`np.shares_memory(input_data['image'], outputs['image'])`, which
returned `True`, indicating that the input and output tensors share
memory. After applying the proposed changes, the same check now returns
`False`, confirming that memory sharing is resolved.
To fix this, the `ONNXEngine` logic has been updated to create a copy of
any output tensor that is also used as a model input. This ensures that
the input tensor and statistics data remain independent, avoiding
unintended side effects.

2. Merge RawReducer and NoopReducer
3. Minor fixes (remove warnings + fix bug in BC)


### Reason for changes

Regression

### Related tickets

156025

### Tests

PTQ run 549

nncf/experimental/common/tensor_statistics/collectors.py

@@ -412,25 +412,17 @@ def __init__(self, tensor_collectors: List[TensorCollector]) -> None:
 ##################################################
 
 
-class NoopReducer(TensorReducerBase):
+class RawReducer(TensorReducerBase):
     def __init__(self):
         super().__init__(inplace=False)
 
     def get_inplace_fn(self) -> Optional[InplaceInsertionFNType]:
         return None
 
-    def _reduce_out_of_place(self, x: List[TensorType]) -> List[TensorType]:
+    def _reduce_out_of_place(self, x: List[Tensor]) -> List[Tensor]:
         return x
 
 
-class RawReducer(NoopReducer):
-    def __init__(self):
-        super().__init__()
-
-    def __call__(self, x: List[Tensor]):
-        return self._reduce_out_of_place(x)
-
-
 class ShapeReducer(TensorReducerBase):
     def __init__(self, inplace: bool = False):
         super().__init__(inplace=inplace)

nncf/onnx/engine.py

@@ -38,7 +38,13 @@ def infer(self, input_data: Dict[str, np.ndarray]) -> Dict[str, np.ndarray]:
         :param input_data: inputs for the model
         :return output_data: models outputs
         """
-        output_tensors = self.sess.run([], {k: v for k, v in input_data.items() if k in self.input_names})
+        output_tensors = self.sess.run([], input_data)
         model_outputs = self.sess.get_outputs()
 
-        return {output.name: tensor for tensor, output in zip(output_tensors, model_outputs)}
+        outputs_safe = {}
+        for tensor, output in zip(output_tensors, model_outputs):
+            # Workaround for https://github.com/microsoft/onnxruntime/issues/21922
+            # After fixing this copying should be removed
+            outputs_safe[output.name] = tensor.copy() if output.name in self.input_names else tensor
+
+        return outputs_safe

nncf/onnx/graph/model_transformer.py

@@ -49,7 +49,7 @@ class ONNXModelTransformer(ModelTransformer):
     def __init__(self, model: onnx.ModelProto):
         infered_model = onnx.shape_inference.infer_shapes(model)
         super().__init__(infered_model)
-        self.onnx_model_extractor = onnx.utils.Extractor(self._model)
+        self.onnx_model_extractor = onnx.utils.Extractor(infered_model)
 
     def _get_target_edge(
         self,

nncf/onnx/statistics/collectors.py

@@ -20,7 +20,6 @@
 from nncf.experimental.common.tensor_statistics.collectors import MeanReducer
 from nncf.experimental.common.tensor_statistics.collectors import MinReducer
 from nncf.experimental.common.tensor_statistics.collectors import NoopAggregator
-from nncf.experimental.common.tensor_statistics.collectors import NoopReducer
 from nncf.experimental.common.tensor_statistics.collectors import QuantileReducer
 from nncf.experimental.common.tensor_statistics.collectors import RawReducer
 from nncf.experimental.common.tensor_statistics.collectors import ShapeAggregator
@@ -48,7 +47,7 @@ def get_mean_statistic_collector(
         reducer = BatchMeanReducer(inplace)
     else:
         reducer = MeanPerChReducer(channel_axis=channel_axis, inplace=inplace)
-    noop_reducer = NoopReducer()
+    raw_reducer = RawReducer()
 
     kwargs = {
         "num_samples": num_samples,
@@ -60,7 +59,7 @@ def get_mean_statistic_collector(
 
     collector = TensorCollector(MeanTensorStatistic)
     collector.register_statistic_branch(MeanTensorStatistic.MEAN_STAT, reducer, aggregate_mean)
-    collector.register_statistic_branch(MeanTensorStatistic.SHAPE_STAT, noop_reducer, aggregate_shape)
+    collector.register_statistic_branch(MeanTensorStatistic.SHAPE_STAT, raw_reducer, aggregate_shape)
     return collector
 
 

nncf/openvino/statistics/collectors.py

@@ -24,7 +24,6 @@
 from nncf.experimental.common.tensor_statistics.collectors import MeanVarianceReducer
 from nncf.experimental.common.tensor_statistics.collectors import MinReducer
 from nncf.experimental.common.tensor_statistics.collectors import NoopAggregator
-from nncf.experimental.common.tensor_statistics.collectors import NoopReducer
 from nncf.experimental.common.tensor_statistics.collectors import QuantileReducer
 from nncf.experimental.common.tensor_statistics.collectors import RawReducer
 from nncf.experimental.common.tensor_statistics.collectors import ShapeAggregator
@@ -128,7 +127,7 @@ def get_mean_statistic_collector(
         reducer = OVBatchMeanReducer(inplace)
     else:
         reducer = OVMeanPerChanelReducer(channel_axis=channel_axis, inplace=inplace)
-    noop_reducer = NoopReducer()
+    raw_reducer = RawReducer()
 
     kwargs = {
         "num_samples": num_samples,
@@ -139,7 +138,7 @@ def get_mean_statistic_collector(
 
     collector = TensorCollector(MeanTensorStatistic)
     collector.register_statistic_branch(MeanTensorStatistic.MEAN_STAT, reducer, aggregate_mean)
-    collector.register_statistic_branch(MeanTensorStatistic.SHAPE_STAT, noop_reducer, aggregate_shape)
+    collector.register_statistic_branch(MeanTensorStatistic.SHAPE_STAT, raw_reducer, aggregate_shape)
     return collector
 
 

nncf/quantization/algorithms/bias_correction/algorithm.py

@@ -442,7 +442,7 @@ def _get_bias_shift_magnitude(current_bias_value: Tensor, updated_bias_value: Te
         """
         bias_shift_magnitude = fns.max(
             fns.abs(
-                (updated_bias_value - current_bias_value) / (current_bias_value + fns.finfo(current_bias_value).min)
+                (updated_bias_value - current_bias_value) / (current_bias_value + fns.finfo(current_bias_value).eps)
             )
         )
         return bias_shift_magnitude

nncf/quantization/algorithms/weight_compression/backend.py

@@ -20,7 +20,7 @@
 from nncf.common.graph.transformations.commands import TargetType
 from nncf.common.tensor_statistics.collectors import TensorStatisticCollectorBase
 from nncf.experimental.common.tensor_statistics.collectors import HAWQAggregator
-from nncf.experimental.common.tensor_statistics.collectors import NoopReducer
+from nncf.experimental.common.tensor_statistics.collectors import RawReducer
 from nncf.experimental.common.tensor_statistics.collectors import TensorCollector
 from nncf.experimental.common.tensor_statistics.statistics import HessianTensorStatistic
 from nncf.quantization.algorithms.weight_compression.config import WeightCompressionParameters
@@ -252,7 +252,7 @@ def scale_insertion_command(source_node, next_nodes, source_node_output_port, sc
 class MixedPrecisionAlgoBackend(ABC):
     @staticmethod
     def hawq_statistic_collector(subset_size: Optional[int] = None) -> TensorCollector:
-        reducer = NoopReducer()
+        reducer = RawReducer()
         aggregator = HAWQAggregator(num_samples=subset_size)
         collector = TensorCollector(HessianTensorStatistic)
         collector.register_statistic_branch(HessianTensorStatistic.HESSIAN_INPUT_ACTIVATION_STATS, reducer, aggregator)

nncf/quantization/fake_quantize.py

@@ -122,11 +122,13 @@ def tune_range(
         fval = -left_border * s
         qval = fns.round(fval)
 
-    ra = fns.where(qval < level_high, qval / (qval - level_high) * right_border, left_border)
     with warnings.catch_warnings():
         # If `qval` is 0 `rb` will equal `right_border`, and we don't want to show an unnecessary division by 0 warning
+        # The same for (qval - level_high)
         warnings.simplefilter("ignore")
+        ra_then_result = qval / (qval - level_high) * right_border
         rb_then_result = (qval - level_high) / qval * left_border
+    ra = fns.where(qval < level_high, ra_then_result, left_border)
     rb = fns.where(qval > 0.0, rb_then_result, right_border)
 
     range_a = right_border - ra

nncf/torch/tensor_statistics/collectors.py

@@ -27,7 +27,6 @@
 from nncf.experimental.common.tensor_statistics.collectors import MinAggregator
 from nncf.experimental.common.tensor_statistics.collectors import MinReducer
 from nncf.experimental.common.tensor_statistics.collectors import NoopAggregator
-from nncf.experimental.common.tensor_statistics.collectors import NoopReducer
 from nncf.experimental.common.tensor_statistics.collectors import PercentileAggregator
 from nncf.experimental.common.tensor_statistics.collectors import QuantileReducer
 from nncf.experimental.common.tensor_statistics.collectors import RawReducer
@@ -246,7 +245,7 @@ def _get_collection_without_reduction(
     :return: Target statistic collector.
     """
     tensor_collector = TensorCollector(statistic_cls)
-    reducer = NoopReducer()
+    reducer = RawReducer()
     aggregation_axes = list(set(list(aggregation_axes) + [dim + 1 for dim in reduction_axes]))
     aggregator = aggregator_cls(
         aggregation_axes=aggregation_axes,
@@ -311,7 +310,7 @@ def get_mean_statistic_collector(
         reducer = BatchMeanReducer()
     else:
         reducer = MeanPerChReducer(channel_axis=channel_axis)
-    noop_reducer = NoopReducer()
+    raw_reducer = RawReducer()
 
     kwargs = {
         "num_samples": num_samples,
@@ -322,7 +321,7 @@ def get_mean_statistic_collector(
 
     collector = TensorCollector(MeanTensorStatistic)
     collector.register_statistic_branch(MeanTensorStatistic.MEAN_STAT, reducer, aggregate_mean)
-    collector.register_statistic_branch(MeanTensorStatistic.SHAPE_STAT, noop_reducer, aggregate_shape)
+    collector.register_statistic_branch(MeanTensorStatistic.SHAPE_STAT, raw_reducer, aggregate_shape)
     return collector
 
 

tests/common/experimental/test_reducers_and_aggregators.py

@@ -31,7 +31,6 @@
 from nncf.experimental.common.tensor_statistics.collectors import MedianNoOutliersAggregator
 from nncf.experimental.common.tensor_statistics.collectors import MinAggregator
 from nncf.experimental.common.tensor_statistics.collectors import NoopAggregator
-from nncf.experimental.common.tensor_statistics.collectors import NoopReducer
 from nncf.experimental.common.tensor_statistics.collectors import PercentileAggregator
 from nncf.experimental.common.tensor_statistics.collectors import RawReducer
 from nncf.experimental.common.tensor_statistics.collectors import ShapeAggregator
@@ -173,19 +172,19 @@ def squeeze_tensor(self, ref_tensor: List[Any], axes: Optional[Tuple[int]] = Non
     def cast_tensor(self, tensor, dtype: Dtype):
         pass
 
-    @pytest.mark.parametrize("reducer_cls", [NoopReducer, RawReducer])
+    @pytest.mark.parametrize("reducer_cls", [RawReducer])
     @pytest.mark.parametrize("input_data", [np.arange(24).reshape((1, 2, 3, 4)), np.array([])])
     def test_other_reducers(self, reducer_cls, input_data):
         reducer = reducer_cls()
         tensor_data = self.get_nncf_tensor(input_data)
         reduced_input = reducer([tensor_data])
-        if reducer_cls == NoopReducer and tensor_data.isempty():
+        if tensor_data.isempty():
             assert reduced_input is None
         else:
             assert len(reduced_input) == 1
             assert fns.allclose(reduced_input[0], tensor_data)
 
-    @pytest.mark.parametrize("reducer_cls", [NoopReducer, RawReducer, ShapeReducer])
+    @pytest.mark.parametrize("reducer_cls", [RawReducer, ShapeReducer])
     def test_other_reducers_name_hash_equal(self, reducer_cls):
         reducers_instances = [reducer_cls() for _ in range(2)]
         assert hash(reducers_instances[0]) == hash(reducers_instances[1])

tests/common/quantization/test_tune_range.py

@@ -11,13 +11,21 @@
 import warnings
 
 import numpy as np
+import pytest
 
 from nncf.quantization.fake_quantize import tune_range
 from nncf.tensor import Tensor
 
 
-def test_tune_range_zero_division_warning():
+@pytest.mark.parametrize(
+    "params",
+    (
+        (Tensor(np.array([0.0])), Tensor(np.array([1.0])), 8, False),
+        (Tensor(np.array([-1.0])), Tensor(np.array([0.0])), 8, False),
+    ),
+)
+def test_tune_range_zero_division_warning(params):
     with warnings.catch_warnings(record=True) as w:
         # Calling tune_range should not raise a warning
-        tune_range(Tensor(np.array([0.0])), Tensor(np.array([1.0])), 8, False)
+        tune_range(*params)
         assert len(w) == 0

tests/cross_fw/test_templates/test_weights_compression_backends.py

@@ -18,7 +18,7 @@
 from nncf.experimental.common.tensor_statistics.collectors import MeanAbsMaxReducer
 from nncf.experimental.common.tensor_statistics.collectors import MeanAggregator
 from nncf.experimental.common.tensor_statistics.collectors import MeanVarianceReducer
-from nncf.experimental.common.tensor_statistics.collectors import NoopReducer
+from nncf.experimental.common.tensor_statistics.collectors import RawReducer
 from nncf.experimental.common.tensor_statistics.collectors import TensorCollector
 
 
@@ -58,7 +58,7 @@ def check_reducer(self, collector: TensorCollector, expected_reducer_type):
     @pytest.mark.parametrize(
         "algo_func, aggregator_type, reducer_type",
         [
-            ("get_hawq_with_backend", HAWQAggregator, NoopReducer),
+            ("get_hawq_with_backend", HAWQAggregator, RawReducer),
             ("get_mean_variance_with_backend", MeanAggregator, MeanVarianceReducer),
             ("get_max_variance_with_backend", MeanAggregator, MaxVarianceReducer),
             ("get_mean_max_with_backend", MeanAggregator, MeanAbsMaxReducer),