Fix tests

Author: daniil-lyakhov

nncf/experimental/quantization/quantizer.py

@@ -23,6 +23,10 @@
 
 
 class IntDtype(Enum):
+    """
+    Enum of possible integer types.
+    """
+
     INT8 = "INT8"
     UINT8 = "UINT8"
 
@@ -112,7 +116,7 @@ def transform_prior_quantization(self, model: TModel) -> TModel:
         """
 
     @abstractmethod
-    def get_quantization_setup(self, model: TModel, nncf_graph: NNCFGraph) -> ExtendedFXQuantizerSetup:
+    def get_quantization_setup(self, model: TModel, nncf_graph: NNCFGraph) -> SingleConfigQuantizerSetup:
         """
         Builds SingleConfigQuantizerSetup for the given model.
 

nncf/experimental/torch/fx/quantization/quantizer/openvino_adapter.py

@@ -12,7 +12,7 @@
 import torch.fx
 
 from nncf.common.graph.graph import NNCFGraph
-from nncf.experimental.quantization.quantizer import ExtendedFXQuantizerSetup
+from nncf.common.quantization.quantizer_setup import SingleConfigQuantizerSetup
 from nncf.experimental.quantization.quantizer import Quantizer
 from nncf.experimental.torch.fx.quantization.quantizer.openvino_quantizer import OpenVINOQuantizer
 
@@ -28,5 +28,5 @@ def __init__(self, quantizer: OpenVINOQuantizer):
     def transform_prior_quantization(self, model: torch.fx.GraphModule) -> torch.fx.GraphModule:
         return self._quantizer.transform_for_annotation(model)
 
-    def get_quantization_setup(self, model: torch.fx.GraphModule, nncf_graph: NNCFGraph) -> ExtendedFXQuantizerSetup:
+    def get_quantization_setup(self, model: torch.fx.GraphModule, nncf_graph: NNCFGraph) -> SingleConfigQuantizerSetup:
         return self._quantizer.get_nncf_quantization_setup(model, nncf_graph)

nncf/experimental/torch/fx/quantization/quantizer/openvino_quantizer.py

@@ -36,8 +36,6 @@
 from nncf.common.quantization.quantizer_setup import SingleConfigQuantizerSetup
 from nncf.common.quantization.structs import QuantizationScheme
 from nncf.common.utils.api_marker import api
-from nncf.experimental.quantization.quantizer import ExtendedFXQuantizerSetup
-from nncf.experimental.quantization.quantizer import IntDtype
 from nncf.experimental.torch.fx.nncf_graph_builder import GraphConverter
 from nncf.experimental.torch.fx.node_utils import get_graph_node_by_name
 from nncf.quantization.advanced_parameters import FP8QuantizationParameters
@@ -137,16 +135,9 @@ def set_ignored_scope(
 
     def get_nncf_quantization_setup(
         self, model: torch.fx.GraphModule, nncf_graph: NNCFGraph
-    ) -> ExtendedFXQuantizerSetup:
+    ) -> SingleConfigQuantizerSetup:
         self._min_max_algo._set_backend_entity(model)
-        base_setup = self._min_max_algo.find_quantization_setup(model, nncf_graph)
-        dtype_map = {}
-        for id_, qp in base_setup.quantization_points.items():
-            dtype_map[id_] = None if qp.qconfig.mode == QuantizationScheme.SYMMETRIC else IntDtype.UINT8.value
-
-        state = base_setup.get_state()
-        state[ExtendedFXQuantizerSetup.QUANTIZER_DTYPE_NAME] = dtype_map
-        return ExtendedFXQuantizerSetup.from_state(state)
+        return self._min_max_algo.find_quantization_setup(model, nncf_graph)
 
     def annotate(self, model: torch.fx.GraphModule) -> torch.fx.GraphModule:
         """

nncf/quantization/algorithms/min_max/algorithm.py

@@ -718,7 +718,7 @@ def _get_quantizer_setup(
 
     def _add_weight_quantization_target_point(
         self, quantization_point: SingleConfigQuantizationPoint, nncf_graph: NNCFGraph
-    ) -> TargetPoint:
+    ) -> list[TargetPoint]:
         """
         Adds weight quantization target point to the set of existing points.
 
@@ -728,7 +728,7 @@ def _add_weight_quantization_target_point(
         weight_quantization_target_points = self._get_weight_quantization_target_points(quantization_point, nncf_graph)
         for weight_quantization_target_point in weight_quantization_target_points:
             self._quantization_target_points_to_qconfig[weight_quantization_target_point] = quantization_point.qconfig
-        return weight_quantization_target_point
+        return weight_quantization_target_points
 
     def _add_activation_quantization_target_point(
         self, quantization_point: SingleConfigQuantizationPoint, nncf_graph: NNCFGraph
@@ -854,17 +854,19 @@ def fill_quantization_target_points(
         if isinstance(quantizer_setup, ExtendedQuantizerSetup):
             extra_params = quantizer_setup.get_extra_params()
         else:
-            extra_params = collections.defaultdict(lambda: collections.defaultdict(None))
+            extra_params = collections.defaultdict(lambda: collections.defaultdict(lambda: None))
 
         for id_, quantization_point in quantization_points:
             if quantization_point.is_weight_quantization_point():
-                tp = self._add_weight_quantization_target_point(quantization_point, nncf_graph)
+                tps = self._add_weight_quantization_target_point(quantization_point, nncf_graph)
             elif quantization_point.is_activation_quantization_point():
                 tp = self._add_activation_quantization_target_point(quantization_point, nncf_graph)
+                tps = [tp]
             else:
                 msg = "Incorrect quantization point"
                 raise nncf.InternalError(msg)
-            self._extra_params[tp] = extra_params[id_]
+            for tp in tps:
+                self._extra_params[tp] = extra_params[id_]
 
         return self._quantization_target_points_to_qconfig, self._unified_scale_groups, self._extra_params
 

nncf/quantization/algorithms/min_max/onnx_backend.py

@@ -158,7 +158,7 @@ def create_unified_scales_quantizers_insertion_commands(
     ) -> list[ONNXQuantizerInsertionCommand]:
         return [
             ONNXMinMaxAlgoBackend.create_quantizer_insertion_command(
-                nncf_graph, target_point, quantizer_config, parameters
+                nncf_graph, target_point, quantizer_config, parameters, extra_params
             )
             for target_point in target_points
         ]

nncf/quantization/algorithms/min_max/torch_fx_backend.py

@@ -200,15 +200,13 @@ def _create_quantizer(
         eps = 1e-16
         if dtype is None:
             if quantizer_config.mode != QuantizationScheme.SYMMETRIC:
-                msg = (
-                    "Wrong usage of dtype parameter: it should be specified for the QuantizationScheme.ASYMMETRIC mode"
+                dtype = IntDtype.UINT8
+            else:
+                dtype = (
+                    IntDtype.INT8
+                    if quantizer_config.signedness_to_force or torch.any(parameters.input_low.data < 0.0)
+                    else IntDtype.UINT8
                 )
-                raise nncf.InternalError(msg)
-            dtype = (
-                IntDtype.INT8
-                if quantizer_config.signedness_to_force or torch.any(parameters.input_low.data < 0.0)
-                else IntDtype.UINT8
-            )
 
         if per_channel:
             observer = torch.ao.quantization.observer.PerChannelMinMaxObserver
@@ -237,6 +235,10 @@ def _create_quantizer(
         scale, zero_point = get_scale_zp_from_input_low_input_high(
             level_low, level_high, parameters.input_low.data, parameters.input_high.data
         )
+
+        scale = scale.view(-1)
+        zero_point = zero_point.view(-1)
+
         fakequantizer = FakeQuantize(
             observer=observer,
             quant_max=level_high,

tests/cross_fw/test_templates/test_ptq_params.py

@@ -10,6 +10,7 @@
 # limitations under the License.
 from abc import abstractmethod
 from collections import Counter
+from collections import defaultdict
 from copy import deepcopy
 
 import pytest
@@ -361,6 +362,7 @@ def test_unified_scales_command_creation(self, mocker):
 
         algo._quantization_target_points_to_qconfig = q_tp_vs_qcf
         algo._unified_scale_groups = [unified_scales_group]
+        algo._extra_params = defaultdict(lambda: defaultdict(lambda: None))
 
         mock_transformer = mocker.MagicMock()
         mocker.patch(
@@ -407,9 +409,9 @@ def test_empty_statistics(self, mode, mocker):
 
         dummy_tp = {target_point: QuantizerConfig()}
         if mode == "target_point":
-            dummy_tps = (dummy_tp, {})
+            dummy_tps = (dummy_tp, {}, defaultdict(lambda: None))
         else:
-            dummy_tps = ({}, ((target_point,),))
+            dummy_tps = ({}, ((target_point,),), defaultdict(lambda: None))
         stat_points.add_statistic_point(
             StatisticPoint(target_point, DummyMinMaxTensorCollector(None, None), algo._algorithm_key)
         )

tests/cross_fw/test_templates/test_unified_scales.py

@@ -47,5 +47,5 @@ def test_unified_groups(
         nncf_graph = NNCFGraphFactory.create(backend_model)
         algo = MinMaxQuantization()
         algo._set_backend_entity(backend_model)
-        _, groups = algo._get_quantization_target_points(backend_model, nncf_graph)
+        _, groups, _ = algo._get_quantization_target_points(backend_model, nncf_graph)
         assert [[target.target_node_name for target in groups] for groups in groups] == unified_group

tests/post_training/data/ptq_reference_data.yaml

@@ -43,6 +43,11 @@ torchvision/resnet18_backend_FX_TORCH:
     error_message: "Openvino Model Files Not Found!"
     message: "Issue-166847"
 torchvision/resnet18_backend_CUDA_FX_TORCH:
+  metric_value: 0.6946
+  exception_xfail_reason:
+    type: "FileNotFoundError"
+    error_message: "Openvino Model Files Not Found!"
+    message: "Issue-166847"
 torchvision/resnet18_backend_OV_QUANTIZER_NNCF:
   metric_value: 0.6946
 torchvision/resnet18_backend_OV_QUANTIZER_AO:
@@ -51,10 +56,6 @@ torchvision/resnet18_backend_X86_QUANTIZER_NNCF:
   metric_value: 0.6946
 torchvision/resnet18_backend_X86_QUANTIZER_AO:
   metric_value: 0.6946
-  exception_xfail_reason:
-    type: "FileNotFoundError"
-    error_message: "Openvino Model Files Not Found!"
-    message: "Issue-166847"
 torchvision/mobilenet_v3_small_BC_backend_FP32:
   metric_value: 0.6766
 torchvision/mobilenet_v3_small_BC_backend_OV:
@@ -92,6 +93,11 @@ torchvision/vit_b_16_backend_FX_TORCH:
     error_message: "Openvino Model Files Not Found!"
     message: "Issue-166847"
 torchvision/vit_b_16_backend_CUDA_FX_TORCH:
+  metric_value: 0.80922
+  exception_xfail_reason:
+    type: "FileNotFoundError"
+    error_message: "Openvino Model Files Not Found!"
+    message: "Issue-166847"
 torchvision/vit_b_16_backend_OV_QUANTIZER_NNCF:
   metric_value: 0.80922
 torchvision/vit_b_16_backend_OV_QUANTIZER_AO:
@@ -100,10 +106,6 @@ torchvision/vit_b_16_backend_X86_QUANTIZER_NNCF:
   metric_value: 0.80922
 torchvision/vit_b_16_backend_X86_QUANTIZER_AO:
   metric_value: 0.80922
-  exception_xfail_reason:
-    type: "FileNotFoundError"
-    error_message: "Openvino Model Files Not Found!"
-    message: "Issue-166847"
 torchvision/swin_v2_s_backend_FP32:
   metric_value: 0.83712
 torchvision/swin_v2_s_backend_OV:
@@ -115,6 +117,11 @@ torchvision/swin_v2_s_backend_FX_TORCH:
     error_message: "Openvino Model Files Not Found!"
     message: "Issue-166847"
 torchvision/swin_v2_s_backend_CUDA_FX_TORCH:
+  metric_value: 0.8360
+  exception_xfail_reason:
+    type: "FileNotFoundError"
+    error_message: "Openvino Model Files Not Found!"
+    message: "Issue-166847"
 torchvision/swin_v2_s_backend_OV_QUANTIZER_NNCF:
   metric_value: 0.8360
 torchvision/swin_v2_s_backend_OV_QUANTIZER_AO:
@@ -123,10 +130,6 @@ torchvision/swin_v2_s_backend_X86_QUANTIZER_NNCF:
   metric_value: 0.8360
 torchvision/swin_v2_s_backend_X86_QUANTIZER_AO:
   metric_value: 0.8360
-  exception_xfail_reason:
-    type: "FileNotFoundError"
-    error_message: "Openvino Model Files Not Found!"
-    message: "Issue-166847"
 timm/crossvit_9_240_backend_CUDA_TORCH:
   metric_value: 0.7275
 timm/crossvit_9_240_backend_FP32:

tests/post_training/model_scope.py

@@ -90,9 +90,8 @@
         "compression_params": {
             "subset_size": 2,
         },
-        "backends": [
-            BackendType.FX_TORCH,
-            BackendType.CUDA_FX_TORCH,
+        "backends": FX_BACKENDS
+        + [
             BackendType.TORCH,
             BackendType.CUDA_TORCH,
             BackendType.OV,