Avoid squeeze for seq_len equal to one. (#2821)

### Changes

Changed squeeze logic for statistics postprocessing. Some models has
internal processing which leads to variational seq_len and in this case
we have statistics with different dimension if one off them has seq_len
equal to one.

### Reason for changes

Bug with AWQ/SE for mixtral-8x7b-v0.1

### Related tickets



### Tests


tests/openvino/native/quantization/test_weights_compression.py::test_one_dimentional_samples

nncf/quantization/algorithms/weight_compression/algorithm.py

@@ -506,6 +506,31 @@ def input_filter_func(point):
         self._fp_inputs[input_id] = input_fp
         return self._fp_inputs[input_id]
 
+    @staticmethod
+    def _get_dynamic_shape(x: List[Tensor]):
+        """
+        Compute common shape for set of tensors.
+        For example: return [-1, 10] for tensors with shapes [[1, 10], [5, 10], [100, 10]]
+            or [-1, 10] for tensors with shapes [[1, 1, 10], [1, 1, 10], [1, 100, 10]]
+        :param x: (List[Tensor]): Set of tensors.
+
+        :return: resulting shape with -1 for dimension with dynamic axis,
+                 common size for dimension with static axis if size > 1 else None.
+        """
+        if len(x) == 0:
+            return []
+        res = list(x[0].shape)
+        sz = len(res)
+
+        for i in x:
+            i_shape = i.shape
+            for j in range(sz):
+                if i_shape[j] != res[j]:
+                    res[j] = -1
+        res = [i for i in res if i != 1]
+
+        return res
+
     def _get_activations(
         self, dataset: Dataset, subset_size: int, nodes_to_compress: List[NNCFNode], graph: NNCFGraph, model: TModel
     ) -> Dict[str, List[Tensor]]:
@@ -561,7 +586,9 @@ def _get_activations(
         for node_name, output_id in _collected_stat_inputs_map.items():
             act_node_name, output_port_id = output_id
             x_fp = self._get_fp_inputs(statistic_container, node_name=act_node_name, port_id=output_port_id)
-            x_fp = [i.squeeze() for i in x_fp]  # List[tensor(seq_length, hidden_dim)]
+
+            d_shape = self._get_dynamic_shape(x_fp)
+            x_fp = [i.reshape(d_shape) for i in x_fp]  # List[tensor(seq_length, hidden_dim)]
             activations[node_name] = x_fp
 
             for shared_node_name in act_vs_shared_node_names_mapping[act_node_name]:

tests/openvino/native/models.py

@@ -950,7 +950,7 @@ def get_weights(weights_data, is_int8, name):
             return (qw - zp) * scale
 
     def _create_ov_model(self, is_int8=False):
-        input_node = opset.parameter([8, 8], name="Input_1")
+        input_node = opset.parameter([-1, 8], name="Input_1")
 
         weights_data1 = np.arange(0, 64).reshape(8, 8)
         weights_data1[:] = 2.0

tests/openvino/native/quantization/test_weights_compression.py

@@ -1018,6 +1018,21 @@ def test_call_max_var_criterion_with_dataset_gptq_neg_group_size(mode):
             assert op.get_shape() == [sz, 1]
 
 
+@pytest.mark.parametrize("mode", INT4_MODES)
+def test_one_dimentional_samples(mode):
+    model = AWQMatmulModel().ov_model
+    sz = 8
+    n_samples = 10
+    dataset = Dataset([np.ones([i + 1, sz]) for i in range(n_samples)])
+
+    compressed_model = compress_weights(model, mode=mode, ratio=1.0, group_size=-1, dataset=dataset, awq=True)
+
+    for op in compressed_model.get_ordered_ops():
+        op_name = op.get_friendly_name()
+        if op.get_type_name() == "Constant" and ("/zero_point" in op_name or "/scale" in op_name):
+            assert op.get_shape() == [sz, 1]
+
+
 def get_shape_for_second_input(op_with_weights: ov.Node) -> List[int]:
     return list(op_with_weights.inputs()[1].get_shape())