Rework HWConfig (#3793)

### Changes

Replace JSON-based configuration loading with a Python-style HW Config.

### Reason for changes

Simplifies modification and maintenance by keeping configuration logic
in Python code instead of external JSON files.

### Tests

https://github.com/openvinotoolkit/nncf/actions/runs/20596514331
manual/job/post_training_quantization/767/
https://github.com/openvinotoolkit/nncf/actions/runs/20596794339

Author: AlexanderDokuchaev

src/nncf/common/hardware/config.py

@@ -0,0 +1,85 @@
+# Copyright (c) 2025 Intel Corporation
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#      http://www.apache.org/licenses/LICENSE-2.0
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from dataclasses import field
+from itertools import product
+from typing import Any
+
+from nncf.common.quantization.structs import QuantizationScheme
+from nncf.common.quantization.structs import QuantizerConfig
+from nncf.parameters import StrEnum
+
+SCALES = "scales"
+UNIFIED = "unified"
+ADJUST_PADDING = "adjust_padding"
+
+
+class Granularity(StrEnum):
+    PER_CHANNEL = "per_channel"
+    PER_TENSOR = "per_tensor"
+
+
+@dataclass(frozen=True, kw_only=True, slots=True)
+class QConfigSpace:
+    """
+    A class to represent the configuration space for quantization.
+
+    :param bits: Number of bits for quantization.
+    :param mode: Available quantization schemes.
+    :param granularity: Granularity options for quantization.
+    :param narrow_range: Indicates narrow range quantization.
+    :param signedness_to_force: Optional signedness enforcement.
+    """
+
+    bits: int
+    mode: tuple[QuantizationScheme, ...]
+    granularity: tuple[Granularity, ...]
+    narrow_range: tuple[bool, ...]
+    signedness_to_force: bool | None = None
+
+    def get_all_qconfigs(self) -> list[QuantizerConfig]:
+        """
+        Generate a list of all possible QuantizerConfig instances based on the current
+        settings of mode, granularity, narrow_range, and other parameters.
+
+        :return: A list of QuantizerConfig objects, each representing
+            a unique combination of the quantization parameters.
+        """
+        ret = []
+        for mode, granularity, narrow_range in product(self.mode, self.granularity, self.narrow_range):
+            ret.append(
+                QuantizerConfig(
+                    num_bits=self.bits,
+                    mode=mode,
+                    per_channel=granularity == Granularity.PER_CHANNEL,
+                    narrow_range=narrow_range,
+                    signedness_to_force=self.signedness_to_force,
+                )
+            )
+        return ret
+
+
+@dataclass(frozen=True, kw_only=True, slots=True)
+class OpDesc:
+    """
+    Represents the description of quantization schemes applicable for activations and weights of operation.
+
+    :param type: The type of the operation.
+    :param activations: A tuple containing the quantization configuration for the activations of the operation.
+    :param weights: A tuple containing the quantization configuration for the weights of the operation.
+    :param attributes: A dictionary of additional attributes.
+    """
+
+    type: str
+    activations: tuple[QConfigSpace, ...] = field(default_factory=tuple)
+    weights: tuple[QConfigSpace, ...] = field(default_factory=tuple)
+    attributes: dict[str, Any] = field(default_factory=dict)

src/nncf/common/hardware/configs/cpu.json

@@ -9,8 +9,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from nncf.parameters import StrEnum
 
-class HWConfigOpName:
+
+class HWConfigOpName(StrEnum):
     CONVOLUTION = "Convolution"
     DEPTHWISECONVOLUTION = "DepthWiseConvolution"
     MATMUL = "MatMul"

src/nncf/common/hardware/configs/gpu.json

@@ -0,0 +1,135 @@
+# Copyright (c) 2025 Intel Corporation
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#      http://www.apache.org/licenses/LICENSE-2.0
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from nncf.common.hardware.defines import Granularity
+from nncf.common.hardware.defines import QConfigSpace
+from nncf.common.quantization.structs import QuantizationScheme
+
+q8_a_sym = QConfigSpace(
+    bits=8,
+    mode=(QuantizationScheme.SYMMETRIC,),
+    granularity=(Granularity.PER_TENSOR,),
+    narrow_range=(False,),
+)
+
+q8_a = QConfigSpace(
+    bits=8,
+    mode=(QuantizationScheme.SYMMETRIC, QuantizationScheme.ASYMMETRIC),
+    granularity=(Granularity.PER_TENSOR,),
+    narrow_range=(False,),
+)
+
+q8_a_ch = QConfigSpace(
+    bits=8,
+    mode=(QuantizationScheme.SYMMETRIC, QuantizationScheme.ASYMMETRIC),
+    granularity=(Granularity.PER_CHANNEL, Granularity.PER_TENSOR),
+    narrow_range=(False,),
+)
+
+q8_w_sym = QConfigSpace(
+    bits=8,
+    mode=(QuantizationScheme.SYMMETRIC,),
+    granularity=(Granularity.PER_CHANNEL, Granularity.PER_TENSOR),
+    narrow_range=(True,),
+    signedness_to_force=True,
+)
+
+q8_w_sym_any_nr = QConfigSpace(
+    bits=8,
+    mode=(QuantizationScheme.SYMMETRIC,),
+    granularity=(Granularity.PER_CHANNEL, Granularity.PER_TENSOR),
+    narrow_range=(True, False),
+    signedness_to_force=True,
+)
+
+q8_w_asym = QConfigSpace(
+    bits=8,
+    mode=(QuantizationScheme.ASYMMETRIC,),
+    granularity=(Granularity.PER_CHANNEL, Granularity.PER_TENSOR),
+    narrow_range=(False,),
+)
+
+q16_a_sym = QConfigSpace(
+    bits=16,
+    mode=(QuantizationScheme.SYMMETRIC,),
+    granularity=(Granularity.PER_TENSOR,),
+    narrow_range=(False,),
+)
+
+q16_a = QConfigSpace(
+    bits=16,
+    mode=(QuantizationScheme.SYMMETRIC, QuantizationScheme.ASYMMETRIC),
+    granularity=(Granularity.PER_TENSOR,),
+    narrow_range=(False,),
+)
+
+q16_a_ch = QConfigSpace(
+    bits=16,
+    mode=(QuantizationScheme.SYMMETRIC, QuantizationScheme.ASYMMETRIC),
+    granularity=(Granularity.PER_CHANNEL, Granularity.PER_TENSOR),
+    narrow_range=(False,),
+)
+
+q16_w_sym = QConfigSpace(
+    bits=16,
+    mode=(QuantizationScheme.SYMMETRIC,),
+    granularity=(Granularity.PER_CHANNEL, Granularity.PER_TENSOR),
+    narrow_range=(True,),
+    signedness_to_force=True,
+)
+
+q16_w_sym_any_nr = QConfigSpace(
+    bits=16,
+    mode=(QuantizationScheme.SYMMETRIC,),
+    granularity=(Granularity.PER_CHANNEL, Granularity.PER_TENSOR),
+    narrow_range=(True, False),
+    signedness_to_force=True,
+)
+
+q16_w_asym = QConfigSpace(
+    bits=16,
+    mode=(QuantizationScheme.ASYMMETRIC,),
+    granularity=(Granularity.PER_CHANNEL, Granularity.PER_TENSOR),
+    narrow_range=(False,),
+)
+
+q4_tn = QConfigSpace(
+    bits=4,
+    mode=(QuantizationScheme.SYMMETRIC,),
+    granularity=(Granularity.PER_TENSOR,),
+    narrow_range=(False,),
+)
+q4_ch = QConfigSpace(
+    bits=4,
+    mode=(QuantizationScheme.SYMMETRIC,),
+    granularity=(Granularity.PER_CHANNEL,),
+    narrow_range=(False,),
+)
+q4_w = QConfigSpace(
+    bits=4,
+    mode=(QuantizationScheme.SYMMETRIC,),
+    granularity=(Granularity.PER_CHANNEL, Granularity.PER_TENSOR),
+    narrow_range=(False,),
+)
+q2_ch = QConfigSpace(
+    bits=2,
+    mode=(QuantizationScheme.SYMMETRIC,),
+    granularity=(Granularity.PER_CHANNEL,),
+    narrow_range=(False,),
+)
+
+q2_w = QConfigSpace(
+    bits=2,
+    mode=(QuantizationScheme.SYMMETRIC,),
+    granularity=(Granularity.PER_CHANNEL, Granularity.PER_TENSOR),
+    narrow_range=(False,),
+)

src/nncf/common/hardware/configs/npu.json

@@ -0,0 +1,10 @@
+# Copyright (c) 2025 Intel Corporation
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#      http://www.apache.org/licenses/LICENSE-2.0
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.