[Torch] AWQ support for WeightCompression (#3279)

### Changes

Add support of AWQ (`AWQAlgoAlgoBackend`) for Torch:

- Put common awq tests in test template
- Add torch test case for WC conformance


### Reason for changes

To have better compression results with AWQ support for Torch

### Related tickets

160668

### Tests
WC run -
https://github.com/openvinotoolkit/nncf/actions/runs/13368073134

Author: kshpv

nncf/quantization/algorithms/weight_compression/algorithm.py

@@ -257,6 +257,16 @@ def __init__(
         criterion_cls = MIXED_PRECISION_CRITERIA.get(self._sensitivity_metric)
         self._mixed_precision_algo = criterion_cls(primary_config, self._ratio, self._subset_size)
         self._statistics_path = self._advanced_parameters.statistics_path
+
+        if self._awq:
+            awq_params = self._advanced_parameters.awq_params
+            self.awq_algo = AWQ(
+                awq_params.subset_size,
+                awq_params.percent_to_apply,
+                awq_params.alpha_min,
+                awq_params.alpha_max,
+                awq_params.steps,
+            )
         if self._gptq:
             gptq_params = self._advanced_parameters.gptq_params
             self._gptq_algo = GPTQ(
@@ -586,26 +596,12 @@ def apply(
             nodes_to_compress = list(
                 filter(lambda node: node.node_name not in nodes_names_to_exclude, nodes_to_compress)
             )
-
         if self._awq:
-            awq_params = self._advanced_parameters.awq_params
-            awq_algo = AWQ(
-                model,
-                self._backend_entity.name_to_node_mapping,
-                all_weight_params,
-                nodes_to_compress,
-                statistics,
-                awq_params.subset_size,
-                awq_params.percent_to_apply,
-                awq_params.alpha_min,
-                awq_params.alpha_max,
-                awq_params.steps,
-            )
-            awq_algo.apply(model, graph)
+            self.awq_algo.apply(model, graph, all_weight_params, nodes_to_compress, statistics, self._backend_entity)
             # After applying AWQ we need to update statistics since AWQ alters the activations
-            statistics = awq_algo.update_statistics(statistics)
+            statistics = self.awq_algo.update_statistics(statistics)
             # del is used to prematurely mark non-necessary data as free for garbage collection
-            del awq_algo
+            del self.awq_algo
 
         scales = {}
         zero_points = {}

nncf/quantization/algorithms/weight_compression/awq.py

@@ -11,10 +11,9 @@
 
 from copy import deepcopy
 from dataclasses import dataclass
-from typing import Any, Dict, List, Optional, TypeVar
+from typing import Dict, List, Optional, TypeVar
 
 import nncf
-from nncf import Dataset
 from nncf import nncf_logger
 from nncf.common.factory import ModelTransformerFactory
 from nncf.common.graph.graph import NNCFGraph
@@ -29,6 +28,7 @@
 from nncf.parameters import CompressWeightsMode
 from nncf.quantization.algorithms.algorithm import Algorithm
 from nncf.quantization.algorithms.weight_compression.activation_stats import process_stats
+from nncf.quantization.algorithms.weight_compression.backend import WeightCompressionAlgoBackend
 from nncf.quantization.algorithms.weight_compression.config import WeightCompressionParameters
 from nncf.quantization.algorithms.weight_compression.weight_lowering import calculate_nf4_scale
 from nncf.quantization.algorithms.weight_compression.weight_lowering import do_nf4_dequantization
@@ -61,34 +61,20 @@ class AWQ(Algorithm):
 
     def __init__(
         self,
-        model: TModel,
-        name_to_node_mapping: Dict[str, Any],
-        all_weight_params: List[WeightCompressionParameters],
-        nodes_to_compress: List[NNCFNode],
-        statistics: Dict[str, WCTensorStatistic],
         subset_size: int = 32,
-        percent_to_apply=0.002,
-        alpha_min=0.0,
-        alpha_max=1.0,
-        steps=100,
+        percent_to_apply: float = 0.002,
+        alpha_min: float = 0.0,
+        alpha_max: float = 1.0,
+        steps: int = 100,
     ):
         """
-        :param model: Model for applying algorithm.
-        :param name_to_node_mapping: Name to node mapping for updating node weights.
-        :param all_weight_params: List of all weight parameters.
-        :param nodes_to_compress: List of nodes for processing.
-        :param statistics: Input activation statistics for each node.
         :param subset_size: The number of samples for AWQ.
         :param percent_to_apply: The percent of outliers for correction.
         :param alpha_min: Minimum value of smoothness parameter for grid search.
         :param alpha_max: Maximal value of smoothness parameter for grid search.
         :param steps: The number of the steps in grid search.
         """
         super().__init__()
-        self.name_to_node_mapping = name_to_node_mapping
-        self._all_weight_params = all_weight_params
-        self._nodes_to_compress = nodes_to_compress
-        self._statistics = statistics
         self._subset_size = subset_size
         self._percent_to_apply = percent_to_apply
         self._alpha_min = alpha_min
@@ -98,44 +84,54 @@ def __init__(
         self._patterns = None
         self._scale_per_target_node = {}
 
-        self._set_backend_entity(model)
-
     @property
     def available_backends(self) -> List[BackendType]:
-        return [BackendType.OPENVINO]
+        return [BackendType.OPENVINO, BackendType.TORCH]
 
-    def _set_backend_entity(self, model: TModel) -> None:
+    def _set_backend_entity(
+        self, model: TModel, wc_backend_entity: Optional[WeightCompressionAlgoBackend] = None
+    ) -> None:
         """
         Creates a helper class with a backed-specific logic of the algorithm.
 
         :param model: Backend-specific input model.
+        :param wc_backend_entity: Weight compression algorithm backend.
         """
         model_backend = get_backend(model)
         if model_backend == BackendType.OPENVINO:
             from nncf.quantization.algorithms.weight_compression.openvino_backend import OVAWQAlgoAlgoBackend
 
-            self._backend_entity = OVAWQAlgoAlgoBackend(model, self.name_to_node_mapping)
-            self._patterns = self._backend_entity.get_awq_patterns()
+            self._backend_entity = OVAWQAlgoAlgoBackend(model, wc_backend_entity.name_to_node_mapping)
+        elif model_backend == BackendType.TORCH:
+            from nncf.quantization.algorithms.weight_compression.torch_backend import PTAWQAlgoAlgoBackend
+
+            self._backend_entity = PTAWQAlgoAlgoBackend()
+
         else:
             msg = f"Cannot return backend-specific AWQ entity because {model_backend.value} is not supported!"
             raise nncf.UnsupportedBackendError(msg)
+        self._patterns = self._backend_entity.get_awq_patterns()
 
     def apply(
         self,
         model: TModel,
         graph: NNCFGraph,
-        statistic_points: Optional[StatisticPointsContainer] = None,
-        dataset: Optional[Dataset] = None,
+        all_weight_params: List[WeightCompressionParameters],
+        nodes_to_compress: List[NNCFNode],
+        statistics: Dict[str, WCTensorStatistic],
+        wc_backend_entity: Optional[WeightCompressionAlgoBackend] = None,
     ) -> TModel:
         """
         Applies the algorithm to the model.
-
         :param model: Model for applying algorithm.
         :param graph: Model graph.
-        :param statistic_points: Statistic points with collected statistics values.
-        :param dataset: A representative dataset for the calibration process.
+        :param all_weight_params: List of all weight parameters.
+        :param nodes_to_compress: List of nodes for processing.
+        :param statistics: Input activation statistics for each node.
+        :param wc_backend_entity: Weight compression algorithm backend.
         :return: A resulting model.
         """
+        self._set_backend_entity(model, wc_backend_entity)
         matches = []
 
         inference_nncf_graph = transform_to_inference_graph(deepcopy(graph), [], [], [], [])
@@ -151,7 +147,7 @@ def apply(
         model_transformer = ModelTransformerFactory.create(model, inplace=True)
 
         awq_data = {}
-        name_mapping = {wp.weight_name: idx for idx, wp in enumerate(self._all_weight_params)}
+        name_mapping = {wp.weight_name: idx for idx, wp in enumerate(all_weight_params)}
 
         for match in matches:
             nncf_node = graph.get_node_by_key(match[-1])
@@ -166,11 +162,11 @@ def apply(
             if target_node_names[-1] not in name_mapping:
                 continue
 
-            weight_params = self._all_weight_params[name_mapping[target_node_names[-1]]]
+            weight_params = all_weight_params[name_mapping[target_node_names[-1]]]
 
             if weight_params.compression_config.num_bits != 4:
                 continue
-            target_node = self._nodes_to_compress[name_mapping[target_node_names[-1]]]
+            target_node = nodes_to_compress[name_mapping[target_node_names[-1]]]
 
             # avoid matching different patterns for the same node
             if target_node.node_name in awq_data:
@@ -182,7 +178,7 @@ def apply(
                 merge_node_names = []
                 for weight_op_friendly_name, _ in self._backend_entity.get_weight_names_and_port_ids(nncf_node, graph):
                     merge_node_names.append(weight_op_friendly_name)
-                merge_node = self._nodes_to_compress[name_mapping[merge_node_names[-1]]]
+                merge_node = nodes_to_compress[name_mapping[merge_node_names[-1]]]
             else:  # pattern Act->MatMul or Act->Multiply->MatMul
                 merge_node = nncf_node
 
@@ -204,7 +200,7 @@ def apply(
 
             config = wp.compression_config
 
-            s, X = process_stats(self._statistics[k], self._subset_size)
+            s, X = process_stats(statistics[k], self._subset_size)
 
             top_k = max(int(s.shape[0] * self._percent_to_apply), 1)
             topk_idxs = fns.argsort(-s)[:top_k]

nncf/quantization/algorithms/weight_compression/torch_backend.py

@@ -19,6 +19,7 @@
 from nncf.common.graph.graph import NNCFNode
 from nncf.common.graph.operator_metatypes import CONST_NOOP_METATYPES
 from nncf.common.graph.operator_metatypes import OperatorMetatype
+from nncf.common.graph.patterns import GraphPattern
 from nncf.common.graph.transformations.commands import TargetType
 from nncf.common.graph.transformations.layout import TransformationLayout
 from nncf.common.tensor_statistics.statistic_point import StatisticPoint
@@ -35,6 +36,9 @@
 from nncf.experimental.common.tensor_statistics.statistics import MeanVarianceTensorStatistic
 from nncf.experimental.common.tensor_statistics.statistics import WCTensorStatistic
 from nncf.parameters import CompressWeightsMode
+from nncf.quantization.algorithms.smooth_quant.torch_backend import SQMultiply
+from nncf.quantization.algorithms.weight_compression.awq_patterns import get_awq_patterns
+from nncf.quantization.algorithms.weight_compression.backend import AWQAlgoBackend
 from nncf.quantization.algorithms.weight_compression.backend import MixedPrecisionAlgoBackend
 from nncf.quantization.algorithms.weight_compression.backend import WeightCompressionAlgoBackend
 from nncf.quantization.algorithms.weight_compression.config import WeightCompressionParameters
@@ -44,6 +48,8 @@
 from nncf.tensor.definitions import TensorDataType
 from nncf.torch.dynamic_graph.scope import Scope
 from nncf.torch.graph import operator_metatypes as om
+from nncf.torch.graph.operator_metatypes import PTMulMetatype
+from nncf.torch.graph.pattern_operations import ATOMIC_ACTIVATIONS_OPERATIONS
 from nncf.torch.graph.transformations.commands import PTSharedFnInsertionCommand
 from nncf.torch.graph.transformations.commands import PTTargetPoint
 from nncf.torch.model_graph_manager import find_const_node_in_constant_subgraph
@@ -52,6 +58,7 @@
 from nncf.torch.model_graph_manager import get_module_by_name
 from nncf.torch.model_graph_manager import split_const_name
 from nncf.torch.model_transformer import PTModelTransformer
+from nncf.torch.model_transformer import update_parameter
 from nncf.torch.nncf_network import NNCFNetwork
 from nncf.torch.quantization.layers import INT4AsymmetricWeightsDecompressor
 from nncf.torch.quantization.layers import INT4SymmetricWeightsDecompressor
@@ -202,12 +209,12 @@ def get_weight_shape(node_with_weight: NNCFNode, weight_port_id: int, graph: NNC
     def set_weight(
         self, node_with_weight: NNCFNode, weight_port_id: int, model: torch.nn.Module, graph: NNCFGraph, weight: Tensor
     ):
-        pass
+        update_parameter(node_with_weight.node_name, "weight", weight.data, model)
 
     def insert_adapters(
         self, wc_params: WeightCompressionParameters, lora_A: Tensor, lora_B: Tensor, int8_lora: bool
     ) -> None:
-        pass
+        raise NotImplementedError()
 
     @staticmethod
     def get_filter_fn_for_statistics(activation_port_id: int, algorithm_key: str) -> Callable[[StatisticPoint], bool]:
@@ -320,6 +327,39 @@ def transform_model(
         return transformed_model
 
 
+class PTAWQAlgoAlgoBackend(AWQAlgoBackend, PTWeightCompressionAlgoBackend):
+    @staticmethod
+    def get_awq_patterns():
+        return get_awq_patterns(
+            PTWeightCompressionAlgoBackend.MATMUL_METATYPES,
+            PTMulMetatype,
+            ATOMIC_ACTIVATIONS_OPERATIONS[GraphPattern.METATYPE_ATTR],
+        )
+
+    @staticmethod
+    def scale_insertion_command(
+        source_node: NNCFNode,
+        next_nodes,
+        source_output_port_id: int,
+        scale: torch.Tensor,
+    ) -> PTSharedFnInsertionCommand:
+        input_port_id = 0
+        target_points = []
+        for node in next_nodes:
+            target_points.append(
+                PTTargetPoint(
+                    PTWeightCompressionAlgoBackend.TARGET_TYPE_TO_PT_INS_TYPE_MAP[TargetType.PRE_LAYER_OPERATION],
+                    node.node_name,
+                    input_port_id=input_port_id,
+                )
+            )
+
+        sq_multiply = SQMultiply(scale.shape)
+        sq_multiply.scale = scale
+        scale_node_name = f"{source_node.node_name}/awq_mul"
+        return PTSharedFnInsertionCommand(target_points, sq_multiply, scale_node_name)
+
+
 class PTMixedPrecisionAlgoBackend(MixedPrecisionAlgoBackend, PTWeightCompressionAlgoBackend):
     @staticmethod
     def mean_variance_statistic_collector(

nncf/quantization/quantize_model.py

@@ -518,11 +518,7 @@ def compress_weights(
             msg = "Torch backend does not support NF4 and E2M1 modes for weight compression."
             raise nncf.ParameterNotSupportedError(msg)
 
-        options = {
-            "awq": awq,
-            "gptq": gptq,
-            "lora_correction": lora_correction,
-        }
+        options = {"gptq": gptq, "lora_correction": lora_correction}
         unsupported_options = [name for name, value in options.items() if value is not None]
         if unsupported_options:
             msg = f"Torch backend does not support {', '.join(unsupported_options)} option(s). Set them to None."

nncf/torch/model_transformer.py

@@ -14,7 +14,6 @@
 from typing import Callable, Dict, List, Optional, Tuple
 
 import torch
-from torch import Tensor
 from torch import nn
 from torch.nn.parameter import Parameter
 
@@ -242,7 +241,7 @@ def _apply_weights_update_transformations(
         return model
 
 
-def update_parameter(target_node_name: str, parameter_name: str, new_value: Tensor, model: NNCFNetwork) -> None:
+def update_parameter(target_node_name: str, parameter_name: str, new_value: torch.Tensor, model: NNCFNetwork) -> None:
     """
     Update parameter for target module.