Tests' fixes

Author: ljaljushkin

nncf/quantization/algorithms/weight_compression/algorithm.py

@@ -160,18 +160,22 @@ def check_user_compression_configuration(
         msg = f"The ratio should be between 0 and 1, but ratio={ratio} is specified."
         raise nncf.ValidationError(msg)
 
-    for size in [
-        subset_size,
-        advanced_parameters.awq_params.subset_size,
-        advanced_parameters.scale_estimation_params.subset_size,
-        advanced_parameters.gptq_params.subset_size,
-        advanced_parameters.lora_correction_params.subset_size,
-    ]:
+    values_to_check = [subset_size]
+    ranks = []
+    if advanced_parameters:
+        values_to_check.extend[
+            advanced_parameters.awq_params.subset_size,
+            advanced_parameters.scale_estimation_params.subset_size,
+            advanced_parameters.gptq_params.subset_size,
+            advanced_parameters.lora_correction_params.subset_size,
+        ]
+        ranks = [advanced_parameters.lora_adapter_rank, advanced_parameters.lora_correction_params.adapter_rank]
+    for size in values_to_check:
         if size <= 0:
             msg = f"The subset_size value should be positive, but subset_size={size} is given."
             raise nncf.ValidationError(msg)
 
-    for rank in [advanced_parameters.lora_adapter_rank, advanced_parameters.lora_correction_params.adapter_rank]:
+    for rank in ranks:
         if rank <= 0:
             msg = f"The lora adapter rank should be positive, but rank={rank} is given."
             raise nncf.ValidationError(msg)

nncf/torch/model_transformer.py

@@ -9,6 +9,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import contextlib
 from collections import defaultdict
 from functools import partial
 from typing import Callable, Dict, List, Optional, Tuple
@@ -20,6 +21,7 @@
 from nncf.common.graph.model_transformer import ModelTransformer
 from nncf.common.graph.transformations.commands import TargetType
 from nncf.common.graph.transformations.commands import TransformationPriority
+from nncf.errors import InternalError
 from nncf.torch.extractor import extract_model
 from nncf.torch.graph.transformations.commands import ExtraCompressionModuleType
 from nncf.torch.graph.transformations.commands import PTBiasCorrectionCommand
@@ -80,15 +82,23 @@ def transform(self, transformation_layout: PTTransformationLayout) -> NNCFNetwor
             for command in transformation_layout.transformations:
                 compression_module = command.fn
                 if isinstance(compression_module, nn.Module):
-                    target_point = command.target_points[0]
-                    node_with_weight = graph.get_node_by_name(target_point.target_node_name)
-                    weight_node = get_const_node(node_with_weight, target_point.input_port_id, graph)
-                    if weight_node is None:
-                        weight_node = node_with_weight  # Decompression in DQ compression format is applied to const.
-                    const_data = get_const_data(weight_node, model)
-                    # Compression module and the corresponding layer may have a different device in multi-device setup
-                    # (e.g. when HF model was loaded with device_map='auto'). Need to align devices.
-                    compression_module.to(const_data.device)
+                    points = [command.target_point]
+                    if hasattr(command, "target_points"):
+                        points = command.target_points
+                    for target_point in points:
+                        target_node = graph.get_node_by_name(target_point.target_node_name)
+                        weight_node = None
+                        with contextlib.suppress(InternalError):
+                            weight_node = get_const_node(target_node, target_point.input_port_id, graph)
+                        if weight_node is None:
+                            weight_node = target_node  # Decompression in DQ compression format is applied to const
+                        const_data = None
+                        with contextlib.suppress(AttributeError):
+                            const_data = get_const_data(weight_node, model)
+                        if const_data is not None:
+                            # Compression module and the corresponding layer may have a different device in multi-device
+                            # setup (e.g. when HF model was loaded with device_map='auto'). Need to align devices.
+                            compression_module.to(const_data.device)
             model.nncf.rebuild_graph()
 
         return model