use const graph for nncf.quantize

nncf/quantization/algorithms/fast_bias_correction/algorithm.py

@@ -155,7 +155,9 @@ def apply(
             output_fp = self._get_fp_outputs(statistic_points, out_node_name)
 
             extracted_model = self._extract_submodel(model_transformer, node_name)
-
+            if extracted_model is None:
+                nncf_logger.debug(f"Skipping node {node_name} because cant extract submodel")
+                continue
             sub_input_name, sub_output_name = self._backend_entity.get_sub_input_output_names(extracted_model)
 
             channel_axis = node.metatype.output_channel_axis

nncf/quantization/algorithms/hyperparameter_tuner/algorithm.py

@@ -345,7 +345,7 @@ def _prepare_pipeline_step(
 
         :param step_index: Zero-based index of pipeline step that should be prepared.
         :param step_model: A model.
-        :param step_graph: A graph assosiated with a model.
+        :param step_graph: A graph associated with a model.
         :param step_combinations: Combinations that change parameters only for the step_index-th pipeline step.
         """
         # Create a separate pipeline for each combination

nncf/quantization/algorithms/min_max/algorithm.py

@@ -12,7 +12,7 @@
 import collections
 import dataclasses
 from copy import deepcopy
-from typing import Any, Dict, List, Optional, OrderedDict, Set, TypeVar, Union
+from typing import Any, Dict, List, Optional, OrderedDict, Set, Tuple, TypeVar, Union
 
 import numpy as np
 
@@ -618,7 +618,7 @@ def _get_activation_quantization_target_point(
 
     def _get_quantization_target_points(
         self, model: TModel, nncf_graph: NNCFGraph
-    ) -> OrderedDict[TargetPoint, QuantizerConfig]:
+    ) -> Tuple[OrderedDict[TargetPoint, QuantizerConfig], List[List[TargetPoint]]]:
         """
         Returns Quantization Target Points.
         In the Compression Pipeline logic NNCF assumes that the compression pipeline works only on the single model.
@@ -988,7 +988,7 @@ def _is_node_after_producers(node):
                     quantizer_setup.discard(fq_2_q_key, True)
                     continue
 
-                # In the case of the two quantizers without the brancking after them,
+                # In the case of the two quantizers without the branching after them,
                 # it needs to check that all quantizers follows after producer nodes.
                 if _is_node_after_producers(fq_1_producer) and _is_node_after_producers(fq_2_producer):
                     fq_1_prod_shape = np.prod(nncf_graph.get_output_edges(fq_1_producer)[0].tensor_shape)

nncf/quantization/algorithms/min_max/backend.py

@@ -45,7 +45,7 @@ def mat_mul_metatypes(self) -> List[OperatorMetatype]:
     @abstractmethod
     def post_processing_metatypes(self) -> List[OperatorMetatype]:
         """
-        Property for the backend-specific post-processing metatypes (NonMaximumSupression, TopK, etc.).
+        Property for the backend-specific post-processing metatypes (NonMaximumSuppression, TopK, etc.).
         """
 
     @property

nncf/quantization/algorithms/min_max/torch_backend.py

@@ -80,18 +80,18 @@ def read_variable_metatypes(self) -> List[OperatorMetatype]:
 
     @property
     def conv_metatypes(self) -> List[OperatorMetatype]:
-        return [om.PTModuleConv1dMetatype, om.PTModuleConv2dMetatype, om.PTModuleConv3dMetatype]
+        return [om.PTConv1dMetatype, om.PTConv2dMetatype, om.PTConv3dMetatype]
 
     @property
     def overflow_fix_metatypes(self) -> List[OperatorMetatype]:
         return [
-            om.PTModuleConv1dMetatype,
-            om.PTModuleConv2dMetatype,
-            om.PTModuleConv3dMetatype,
-            om.PTModuleLinearMetatype,
-            om.PTModuleConvTranspose1dMetatype,
-            om.PTModuleConvTranspose2dMetatype,
-            om.PTModuleConvTranspose3dMetatype,
+            om.PTConv1dMetatype,
+            om.PTConv2dMetatype,
+            om.PTConv3dMetatype,
+            om.PTLinearMetatype,
+            om.PTConvTranspose1dMetatype,
+            om.PTConvTranspose2dMetatype,
+            om.PTConvTranspose3dMetatype,
         ]
 
     @property
@@ -210,7 +210,7 @@ def get_statistic_collector(
 
     @staticmethod
     def get_weight_tensor_port_ids(node: NNCFNode) -> List[Optional[int]]:
-        return [None]
+        return node.metatype.weight_port_ids
 
     @staticmethod
     def get_weight_name(nncf_graph: NNCFGraph, target_point: PTTargetPoint) -> str:

nncf/quantization/algorithms/pipeline.py

@@ -36,7 +36,7 @@ def collect_statistics(
 
     :param statistic_points: Statistic points that need to be collected.
     :param model: A model.
-    :param graph: A graph assosiated with a model.
+    :param graph: A graph associated with a model.
     :param dataset: A dataset.
     :return: Collected statistics.
     """
@@ -105,7 +105,7 @@ def run_step(
         :param step_index: Zero-based index of the pipeline step that should be executed
         :param step_statistics: Statistics required to execute a pipeline step.
         :param model: A model to which a pipeline step will be applied.
-        :param graph: A graph assosiated with a model.
+        :param graph: A graph associated with a model.
         :return: The updated model after executing the pipeline step.
         """
         current_model = model
@@ -134,7 +134,7 @@ def run_from_step(
         :param model: This is the model after the (start_step_index - 1)-th pipeline
             step, or the initial model if start_step_index is 0.
         :param dataset: A dataset that holds the data items for pipeline steps.
-        :param graph: A graph assosiated with a model.
+        :param graph: A graph associated with a model.
         :param start_step_index: Zero-based pipeline step index from which the pipeline
             should be executed.
         :param step_index_to_statistics: A mapping from pipeline step index to statistics
@@ -175,7 +175,7 @@ def get_statistic_points_for_step(
 
         :param step_index: Zero-based index of the pipeline step.
         :param model: A model.
-        :param graph: A graph assosiated with a model.
+        :param graph: A graph associated with a model.
         :return: Statistics that should be collected to execute `step_index`-th pipeline step.
         """
         container = StatisticPointsContainer()

nncf/quantization/algorithms/smooth_quant/torch_backend.py

@@ -31,6 +31,7 @@
 from nncf.torch.graph.transformations.command_creation import create_command_to_update_weight
 from nncf.torch.graph.transformations.commands import PTSharedFnInsertionCommand
 from nncf.torch.graph.transformations.commands import PTTargetPoint
+from nncf.torch.model_analyzer import get_const
 from nncf.torch.nncf_network import NNCFNetwork
 from nncf.torch.quantization.default_quantization import DEFAULT_PT_QUANT_TRAIT_TO_OP_DICT
 from nncf.torch.tensor_statistics.collectors import PTAbsMaxReducer
@@ -53,14 +54,14 @@ class PTSmoothQuantAlgoBackend(SmoothQuantAlgoBackend):
     @property
     def convolution_metatypes(self) -> List[OperatorMetatype]:
         return [
-            om.PTModuleConv1dMetatype,
-            om.PTModuleConv2dMetatype,
-            om.PTModuleConv3dMetatype,
+            om.PTConv1dMetatype,
+            om.PTConv2dMetatype,
+            om.PTConv3dMetatype,
         ]
 
     @property
     def matmul_metatypes(self) -> List[OperatorMetatype]:
-        return [om.PTModuleLinearMetatype]
+        return [om.PTLinearMetatype]
 
     @property
     def quantize_agnostic_metatypes(self) -> List[OperatorMetatype]:
@@ -103,10 +104,10 @@ def get_abs_max_channel_collector(
 
     @staticmethod
     def get_weight_value(node_with_weight: NNCFNode, model: NNCFNetwork) -> Tensor:
-        node_module = model.nncf.get_containing_module(node_with_weight.node_name)
-        if node_module.weight is None:
-            raise RuntimeError(f"{node_module} module has no .weight attribute.")
-        return Tensor(node_module.weight.data)
+        data = get_const(node_with_weight, node_with_weight.metatype.weight_port_ids[0], model)
+        if data is None:
+            raise RuntimeError(f"{node_with_weight.node_name} node has no weights.")
+        return Tensor(data)
 
     @staticmethod
     def get_weight_tensor_port_id(node: NNCFNode) -> int:
@@ -136,7 +137,7 @@ def scale_insertion_command(
 
     @staticmethod
     def get_activation_channel_axis(node: NNCFNode, port_id: int) -> int:
-        if node.metatype == om.PTModuleLinearMetatype:
+        if node.metatype == om.PTLinearMetatype:
             return -1
         # TODO: Add activation axis calculation when MatMul will be supported
         return 1

nncf/torch/dynamic_graph/wrappers.py

@@ -278,8 +278,15 @@ def wrap_parameters(model: torch.nn.Module):
 
     :param model: A model.
     """
+    from nncf.torch.external_hook import EXTERNAL_OP_STORAGE_PREFIX
+    from nncf.torch.quantization.external_quantizer import EXTERNAL_QUANTIZERS_STORAGE_PREFIX
+
+    ignored_prefixes = [EXTERNAL_QUANTIZERS_STORAGE_PREFIX, EXTERNAL_OP_STORAGE_PREFIX]
+
     ctx = get_current_context()
     for name, param in model.named_parameters():
+        if any(name.startswith(ignore_prefix) for ignore_prefix in ignored_prefixes):
+            continue
         is_reused = name in ctx.reused_parameters
         tt = TracedParameter.from_torch_parameter(param, name, is_reused)
         ctx.register_traced_tensor(tt)

nncf/torch/external_hook.py

@@ -14,6 +14,7 @@
 from nncf.torch.dynamic_graph.context import TracingContext
 
 EXTERNAL_OP_STORAGE_NAME = "external_op"
+EXTERNAL_OP_STORAGE_PREFIX = "_nncf." + EXTERNAL_OP_STORAGE_NAME
 
 
 class ExternalOpCallHook: