Aanuf/fix for asym

src/nncf/openvino/optimized_functions/models.py

@@ -532,6 +532,9 @@ def _build_integer_quantization_model(
         min_values, max_values = opset.convert(min_values, ov.Type.f32), opset.convert(max_values, ov.Type.f32)
 
         if is_asym_mode:
+            zero = opset.constant(0.0, ov.Type.f32)
+            min_values = opset.minimum(zero, min_values)
+            max_values = opset.maximum(zero, max_values)
             levels = level_high - level_low + 1
             scale = divide_op(max_values - min_values, opset.constant(levels - 1, ov.Type.f32))
             scale = opset.select(opset.less(opset.abs(scale), eps), eps, scale)

src/nncf/quantization/algorithms/weight_compression/weight_lowering.py

@@ -317,6 +317,11 @@ def calculate_integer_quantization_params(
         level_high = 2**num_bits - 1
         min_values = fns.min(weight, axis=reduction_axes, keepdims=True)  # [a1, r, a2] -> [a1, 1, a2]
         max_values = fns.max(weight, axis=reduction_axes, keepdims=True)  # [a1, r, a2] -> [a1, 1, a2]
+
+        zero = fns.zeros_like(min_values)
+        min_values = fns.minimum(zero, min_values)
+        max_values = fns.maximum(zero, max_values)
+
         scale, zero_point = calculate_scale_zero_point(
             min_values, max_values, level_low, level_high, narrow_range=False
         )
@@ -365,6 +370,7 @@ def get_integer_quantization_error(
 
     decompressed_weight = integer_quantize_dequantize_weight(weight, config, reduction_axes)
     decompressed_weight = decompressed_weight.reshape(weight.shape)
+
     if reduction == "max_mean":
         diff = (decompressed_weight - weight) ** 2
         layer_err = fns.mean(diff, axis=reduction_axes)

tests/cross_fw/test_templates/template_test_weights_compression.py

@@ -209,27 +209,28 @@ def wrap_model(model, data) -> CompressionParams:
         ("mode", "all_layers", "ratio", "ref_ids"),
         (
             (SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, True, 1, [0, 1, 2, 3, 4]),
-            (SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, True, 0.8, [0, 3, 4]),
+            (SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, True, 0.8, [0, 1, 4]),
             (SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, True, 0.4, [0]),
             (SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, True, 0.2, []),
             (SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, False, 1, [0, 1, 2, 3]),
-            (SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, False, 0.8, [0, 1, 3]),
+            (SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, False, 0.8, [0, 1, 2]),
             (SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, False, 0.4, [0]),
             (SensitivityMetric.WEIGHT_QUANTIZATION_ERROR, False, 0.2, []),
             (SensitivityMetric.HESSIAN_INPUT_ACTIVATION, True, 0.8, [0, 1, 2]),
             (SensitivityMetric.HESSIAN_INPUT_ACTIVATION, False, 0.8, [0, 1, 2]),
-            (SensitivityMetric.MEAN_ACTIVATION_VARIANCE, True, 0.8, [0, 1, 2]),
+            (SensitivityMetric.MEAN_ACTIVATION_VARIANCE, True, 0.8, [0, 1, 4]),
             (SensitivityMetric.MEAN_ACTIVATION_VARIANCE, False, 0.8, [0, 1, 2]),
-            (SensitivityMetric.MAX_ACTIVATION_VARIANCE, True, 0.8, [0, 1, 2]),
+            (SensitivityMetric.MAX_ACTIVATION_VARIANCE, True, 0.8, [0, 1, 4]),
             (SensitivityMetric.MAX_ACTIVATION_VARIANCE, False, 0.8, [0, 1, 2]),
-            (SensitivityMetric.MEAN_ACTIVATION_MAGNITUDE, True, 0.8, [0, 1, 2]),
+            (SensitivityMetric.MEAN_ACTIVATION_MAGNITUDE, True, 0.8, [0, 1, 4]),
             (SensitivityMetric.MEAN_ACTIVATION_MAGNITUDE, False, 0.8, [0, 1, 2]),
         ),
     )
     @pytest.mark.parametrize("transpose_a", (False, True))
     def test_mixed_precision(self, mode, all_layers, ratio, ref_ids, transpose_a, mocker):
         model = self.get_sequential_matmul_model(transpose_a=transpose_a)
         input_shape = (4, 4) if transpose_a else (1, 4, 4)
+
         first = self.to_tensor(np.ones(input_shape, dtype=np.float32))
         second = self.to_tensor(np.arange(16, dtype=np.float32)).reshape(input_shape)
         dataset = Dataset([first, second], self.get_transform_func())
@@ -323,6 +324,7 @@ def test_scale_estimation(self, mocker, transpose_a, is_moe, check_sampling_acti
             reference = self.get_moe_scale_estimation_ref(check_sampling_activation_stats_flow)
         else:
             reference = self.get_scale_estimation_ref(check_sampling_activation_stats_flow)
+
         assert fns.allclose(Tensor(reference), computed_scale)
 
     @staticmethod

tests/onnx/quantization/test_weights_compression.py

@@ -553,40 +553,40 @@ def get_scale_estimation_ref(check_sampling_activation_stats_flow):
         return (
             np.array(
                 [
-                    [[0.473328]],
-                    [[0.929023]],
-                    [[1.446527]],
-                    [[1.920595]],
-                    [[2.517054]],
-                    [[3.030102]],
-                    [[3.584279]],
-                    [[4.043509]],
-                    [[4.620008]],
-                    [[5.165322]],
-                    [[5.710637]],
-                    [[6.122581]],
-                    [[6.655914]],
-                    [[7.237174]],
-                    [[7.722580]],
+                    [[0.47332805]],
+                    [[1.0]],
+                    [[1.4732642]],
+                    [[2.0380495]],
+                    [[2.6054149]],
+                    [[3.0301015]],
+                    [[3.679056]],
+                    [[4.175322]],
+                    [[4.700384]],
+                    [[5.2552223]],
+                    [[5.8100615]],
+                    [[6.3083715]],
+                    [[6.858295]],
+                    [[7.4082184]],
+                    [[7.722581]],
                     [[8.255914]],
                 ]
             ).T,
             np.array(
                 [
                     [[0.47344488]],
-                    [[0.9287766]],
-                    [[1.4463282]],
-                    [[1.920052]],
-                    [[2.5167778]],
+                    [[1.0]],
+                    [[1.5450557]],
+                    [[2.0380037]],
+                    [[2.6055446]],
                     [[3.02987]],
-                    [[3.5842714]],
-                    [[4.0429296]],
-                    [[4.619769]],
-                    [[5.165224]],
-                    [[5.7106786]],
-                    [[6.121212]],
-                    [[6.654546]],
-                    [[7.2366524]],
+                    [[3.679132]],
+                    [[4.1754694]],
+                    [[4.7001443]],
+                    [[5.2551227]],
+                    [[5.810101]],
+                    [[6.308658]],
+                    [[6.8587303]],
+                    [[7.4]],
                     [[7.7212124]],
                     [[8.254545]],
                 ]
@@ -601,44 +601,44 @@ def get_moe_scale_estimation_ref(check_sampling_activation_stats_flow):
                     [
                         [
                             [
-                                7.5732,
-                                7.4667,
-                                7.4667,
-                                7.4667,
-                                7.4667,
-                                7.2602,
-                                7.4667,
-                                7.4667,
-                                7.4667,
-                                7.4667,
-                                7.3083,
-                                7.8467,
-                                7.2233,
-                                7.2715,
-                                7.4205,
-                                7.4667,
+                                7.573249,
+                                7.58195,
+                                7.6,
+                                7.6666665,
+                                7.1209445,
+                                7.260152,
+                                7.866667,
+                                7.9333334,
+                                8.0,
+                                8.066667,
+                                8.528544,
+                                8.659291,
+                                8.879055,
+                                8.469787,
+                                8.4,
+                                8.364824,
                             ]
                         ]
                     ],
                     [
                         [
                             [
-                                14.8205,
-                                14.9032,
-                                14.9858,
-                                15.0685,
-                                15.1512,
-                                14.3400,
-                                14.4173,
-                                14.4945,
-                                14.5718,
-                                14.6491,
-                                14.7264,
-                                14.8037,
-                                14.8810,
-                                14.9583,
-                                15.0355,
-                                15.1128,
+                                16.0,
+                                16.089771,
+                                16.179543,
+                                16.269318,
+                                16.359089,
+                                16.44886,
+                                16.538631,
+                                16.628407,
+                                16.718176,
+                                16.80795,
+                                16.89772,
+                                16.987492,
+                                15.812495,
+                                15.89516,
+                                15.977826,
+                                16.060493,
                             ]
                         ]
                     ],
@@ -650,43 +650,43 @@ def get_moe_scale_estimation_ref(check_sampling_activation_stats_flow):
                         [
                             [
                                 7.575118,
-                                7.4666667,
-                                7.4666667,
-                                7.4666667,
-                                7.4666667,
+                                7.5841107,
+                                7.6,
+                                7.6666665,
+                                7.112954,
                                 7.254837,
-                                7.4666667,
-                                7.4666667,
-                                7.4666667,
-                                7.4666667,
-                                7.495066,
+                                7.866667,
+                                7.9333334,
+                                8.0,
+                                8.066667,
+                                8.531546,
                                 7.850108,
-                                7.219489,
-                                7.2685375,
-                                7.418597,
-                                7.4666667,
+                                8.887045,
+                                8.468656,
+                                8.4,
+                                8.361673,
                             ]
                         ]
                     ],
                     [
                         [
                             [
-                                14.820066,
-                                14.902746,
-                                14.985427,
-                                15.068108,
-                                15.150787,
-                                14.3391285,
-                                14.416424,
-                                14.493721,
-                                14.571016,
-                                14.648311,
-                                14.725608,
-                                14.802904,
-                                14.8801985,
-                                14.957496,
-                                15.034791,
-                                15.112087,
+                                16.0,
+                                16.089788,
+                                16.17958,
+                                16.269371,
+                                16.359161,
+                                16.448954,
+                                16.538742,
+                                16.628534,
+                                16.718325,
+                                16.808115,
+                                16.897905,
+                                16.987696,
+                                15.812232,
+                                15.894914,
+                                15.977593,
+                                16.060274,
                             ]
                         ]
                     ],

tests/openvino/native/data/2024.1/reference_scales/IntegerModel_compressed_weights_int4_asym.json

@@ -3,7 +3,7 @@
         "compressed_weight": [
             [
                 [
-                    7.0,
+                    8.0,
                     0.0,
                     13.0
                 ]
@@ -17,8 +17,8 @@
             ],
             [
                 [
-                    10.0,
-                    1.0,
+                    12.0,
+                    4.0,
                     0.0
                 ]
             ],
@@ -79,7 +79,7 @@
         "scale": [
             [
                 [
-                    0.040008544921875
+                    0.046630859375
                 ]
             ],
             [
@@ -89,7 +89,7 @@
             ],
             [
                 [
-                    0.041839599609375
+                    0.0545654296875
                 ]
             ],
             [

tests/openvino/native/data/2024.1/reference_scales/IntegerModel_compressed_weights_int8_asym.json

@@ -2,18 +2,18 @@
     "matmul_2_data": {
         "compressed_weight": [
             [
-                116,
+                136,
                 0,
-                213
+                219
             ],
             [
                 255,
                 61,
                 0
             ],
             [
-                177,
-                10,
+                196,
+                67,
                 0
             ],
             [
@@ -54,13 +54,13 @@
         ],
         "scale": [
             [
-                0.002353668212890625
+                0.002742767333984375
             ],
             [
                 0.00583648681640625
             ],
             [
-                0.002460479736328125
+                0.0032100677490234375
             ],
             [
                 0.0029277801513671875