[ONNX] Fix weight quantization for GroupConvolution (openvinotoolkit#…

…3126)

### Reason for changes

Not quantized weights for GroupConv

### Related tickets

158085

### Tests

ptq perf run 84

nncf/onnx/graph/metatypes/groups.py

@@ -10,6 +10,8 @@
 # limitations under the License.
 
 from nncf.onnx.graph.metatypes import onnx_metatypes
+from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXOpWithWeightsMetatype
+from nncf.onnx.graph.metatypes.onnx_metatypes import get_operator_metatypes
 
 QUANTIZE_AGNOSTIC_OPERATIONS = [
     onnx_metatypes.ONNXGlobalMaxPoolMetatype,
@@ -67,14 +69,19 @@
     onnx_metatypes.ONNXMinimumMetatype,
 ]
 
-
 CONSTANT_WEIGHT_LAYER_METATYPES = [
-    onnx_metatypes.ONNXConvolutionMetatype,
-    onnx_metatypes.ONNXDepthwiseConvolutionMetatype,
-    onnx_metatypes.ONNXConvolutionTransposeMetatype,
-    onnx_metatypes.ONNXEmbeddingMetatype,
+    metatype
+    for metatype in get_operator_metatypes()
+    if issubclass(metatype, ONNXOpWithWeightsMetatype) and metatype.weight_port_ids
 ]
 
+POSSIBLE_WEIGHT_LAYER_METATYPES = [
+    metatype
+    for metatype in get_operator_metatypes()
+    if issubclass(metatype, ONNXOpWithWeightsMetatype) and metatype.possible_weight_ports
+]
+
+OPERATIONS_WITH_WEIGHTS = list(set().union(CONSTANT_WEIGHT_LAYER_METATYPES, POSSIBLE_WEIGHT_LAYER_METATYPES))
 
 LINEAR_OPERATIONS = [
     onnx_metatypes.ONNXConvolutionMetatype,
@@ -124,11 +131,6 @@
     onnx_metatypes.ONNXMeanMetatype,
 ]
 
-OPERATIONS_WITH_WEIGHTS = [
-    *CONSTANT_WEIGHT_LAYER_METATYPES,
-    *MATMUL_METATYPES,
-]
-
 
 BATCH_NORMALIZATION_OPERATIONS = [
     onnx_metatypes.ONNXBatchNormMetatype,

nncf/onnx/graph/metatypes/onnx_metatypes.py

@@ -58,17 +58,17 @@ def determine_subtype(cls, model: onnx.ModelProto, node: onnx.NodeProto) -> Opti
 class ONNXOpWithWeightsMetatype(ONNXOpMetatype):
     """
     Metatype which could have weights.
-
-    :param weight_channel_axis: Axis for weight per-channel quantization, meaning the number of output filters.
-    :param weight_port_ids: Input ports of the node's weight.
-    If the value is None the weight_port_id should be determined dynamically.
-    :param bias_port_id: Input port of the node's bias.
-    If the value is None it means that the Metatype does not have bias.
+    :param weight_channel_axis: Axis for weight per-channel quantization.
+    :param weight_port_ids: Constant input ports of the node's weight. Defaults to an empty list.
+    :param bias_port_id: Input port of the node's bias. If the value is None,
+    it means that the Metatype does not have bias. Defaults to None.
+    :param possible_weight_ports: Input ports on which weight could be laid. Defaults to an empty list.
     """
 
     weight_channel_axis: int
-    weight_port_ids: Optional[List[int]] = None
+    weight_port_ids: List[int] = []
     bias_port_id: Optional[int] = None
+    possible_weight_ports: List[int] = []
 
 
 @ONNX_OPERATION_METATYPES.register(is_subtype=True)
@@ -131,19 +131,17 @@ class ONNXGemmMetatype(ONNXOpWithWeightsMetatype):
     op_names = ["Gemm"]
     hw_config_names = [HWConfigOpName.MATMUL]
     weight_channel_axis = -1  # For port_id=1
-    weight_port_ids = None
     bias_port_id = 2
     possible_weight_ports = [0, 1]
     output_channel_axis = -1
 
 
 @ONNX_OPERATION_METATYPES.register()
-class ONNXMatMulMetatype(ONNXOpMetatype):
+class ONNXMatMulMetatype(ONNXOpWithWeightsMetatype):
     name = "MatMulOp"
     op_names = ["MatMul"]
     hw_config_names = [HWConfigOpName.MATMUL]
     weight_channel_axis = -1  # For port_id=1
-    weight_port_ids = None
     bias_port_id = 2
     possible_weight_ports = [0, 1]
     output_channel_axis = -1
@@ -454,7 +452,7 @@ class ONNXReciprocalMetatype(ONNXOpMetatype):
 
 
 @ONNX_OPERATION_METATYPES.register(is_subtype=True)
-class ONNXEmbeddingMetatype(ONNXOpMetatype):
+class ONNXEmbeddingMetatype(ONNXOpWithWeightsMetatype):
     name = "EmbeddingOp"
     hw_config_names = [HWConfigOpName.EMBEDDING]
     weight_port_ids = [0]

nncf/onnx/graph/nncf_graph_builder.py

@@ -23,8 +23,8 @@
 from nncf.common.graph.operator_metatypes import InputNoopMetatype
 from nncf.common.graph.operator_metatypes import OutputNoopMetatype
 from nncf.onnx.graph.metatypes.groups import CONSTANT_WEIGHT_LAYER_METATYPES
-from nncf.onnx.graph.metatypes.groups import MATMUL_METATYPES
 from nncf.onnx.graph.metatypes.groups import OPERATIONS_WITH_BIAS
+from nncf.onnx.graph.metatypes.groups import POSSIBLE_WEIGHT_LAYER_METATYPES
 from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXGemmMetatype
 from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXOpMetatype
 from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXOpWithWeightsMetatype
@@ -95,7 +95,7 @@ def get_possible_weight_port_ids(metatype: ONNXOpMetatype) -> List[int]:
     :param metatype: Metatype.
     :return: Port ids.
     """
-    if metatype in MATMUL_METATYPES:
+    if metatype in POSSIBLE_WEIGHT_LAYER_METATYPES:
         return metatype.possible_weight_ports
     return []
 

nncf/quantization/algorithms/fast_bias_correction/onnx_backend.py

@@ -81,8 +81,7 @@ def get_bias_value(node: NNCFNode, nncf_graph: NNCFGraph, model: onnx.ModelProto
     @staticmethod
     def get_activation_port_ids_for_bias_node(node: NNCFNode) -> Tuple[int, int]:
         activation_port = 0
-
-        if hasattr(node.metatype, "possible_weight_ports"):
+        if node.metatype.possible_weight_ports:
             activation_ports = deepcopy(node.metatype.possible_weight_ports)
             for weight_port in node.layer_attributes.weight_attrs:
                 activation_ports.remove(weight_port)

tests/post_training/data/ptq_reference_data.yaml

@@ -41,7 +41,7 @@ torchvision/resnet18_backend_FX_TORCH:
 torchvision/mobilenet_v3_small_BC_backend_FP32:
   metric_value: 0.6766
 torchvision/mobilenet_v3_small_BC_backend_OV:
-  metric_value: 0.6669
+  metric_value: 0.6681
 torchvision/mobilenet_v3_small_BC_backend_ONNX:
   metric_value: 0.6679
 torchvision/mobilenet_v3_small_BC_backend_FX_TORCH:
@@ -103,7 +103,7 @@ timm/dpn68_backend_CUDA_TORCH:
 timm/dpn68_backend_FP32:
   metric_value: 0.76342
 timm/dpn68_backend_ONNX:
-  metric_value: 0.75906
+  metric_value: 0.7592
 timm/dpn68_backend_OV:
   metric_value: 0.75972
 timm/dpn68_backend_TORCH:
@@ -201,7 +201,7 @@ timm/regnetx_002_backend_CUDA_TORCH:
 timm/regnetx_002_backend_FP32:
   metric_value: 0.68756
 timm/regnetx_002_backend_ONNX:
-  metric_value: 0.6848
+  metric_value: 0.6854
 timm/regnetx_002_backend_OV:
   metric_value: 0.6852
 timm/regnetx_002_backend_TORCH:
@@ -211,7 +211,7 @@ timm/resnest14d_backend_CUDA_TORCH:
 timm/resnest14d_backend_FP32:
   metric_value: 0.75516
 timm/resnest14d_backend_ONNX:
-  metric_value: 0.75428
+  metric_value: 0.7538
 timm/resnest14d_backend_OV:
   metric_value: 0.75
 timm/resnest14d_backend_TORCH:
@@ -253,7 +253,7 @@ timm/visformer_small_backend_CUDA_TORCH:
 timm/visformer_small_backend_FP32:
   metric_value: 0.82098
 timm/visformer_small_backend_ONNX:
-  metric_value: 0.81562
+  metric_value: 0.8160
 timm/visformer_small_backend_OV:
   metric_value: 0.81674
 timm/visformer_small_backend_TORCH:

tests/post_training/data/wc_reference_data.yaml

@@ -7,40 +7,44 @@ tinyllama_data_aware_backend_OV:
   num_int4: 94
   num_int8: 124
 tinyllama_data_aware_awq_stateful_backend_OV:
-  metric_value: 0.85571
+  metric_value: 0.85616
   num_int4: 94
   num_int8: 124
 tinyllama_data_aware_awq_scale_estimation_backend_OV:
-  metric_value: 0.86355
+  metric_value: 0.85502
   num_int4: 94
   num_int8: 124
 tinyllama_data_aware_awq_scale_estimation_stateful_backend_OV:
-  metric_value: 0.86355
+  metric_value: 0.85502
   num_int4: 94
   num_int8: 124
 tinyllama_int8_data_free_backend_TORCH:
   metric_value: 0.95624
   num_int4: 0
   num_int8: 312
 tinyllama_data_aware_gptq_scale_estimation_stateful_backend_OV:
-  metric_value: 0.86697
+  metric_value: 0.86503
   num_int4: 94
   num_int8: 124
   metrics_xfail_reason: "Issue-148819"
 tinyllama_scale_estimation_per_channel_backend_OV:
-  metric_value: 0.80798
+  metric_value: 0.81389
   num_int4: 188
   num_int8: 124
 tinyllama_data_aware_lora_stateful_backend_OV:
   metric_value: 0.83446
   num_int4: 94
   num_int8: 500
 tinyllama_NF4_scale_estimation_stateful_per_channel_backend_OV:
-  metric_value: 0.87132
+  metric_value: 0.88663
   num_int4: 11
   num_int8: 290
   metrics_xfail_reason: "Issue-148819"
 tinyllama_awq_backup_mode_none_backend_OV:
-  metric_value: 0.85679
+  metric_value: 0.84783
   num_int4: 208
   num_int8: 0
+tinyllama_int4_data_free_backend_TORCH:
+  metric_value: 0.73873
+  num_int4: 114
+  num_int8: 84