Comments

Author: daniil-lyakhov

src/nncf/experimental/quantization/structs.py

@@ -9,23 +9,17 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Any, Optional
+from typing import Any, Literal, Optional
 
+import nncf
 from nncf.common.quantization.structs import QuantizationScheme
 from nncf.common.quantization.structs import QuantizerConfig
 from nncf.config.schemata.defaults import QUANTIZATION_BITS
 from nncf.config.schemata.defaults import QUANTIZATION_NARROW_RANGE
 from nncf.config.schemata.defaults import QUANTIZATION_PER_CHANNEL
-from nncf.parameters import StrEnum
+from nncf.tensor.definitions import TensorDataType
 
-
-class IntDtype(StrEnum):
-    """
-    Enum of possible integer types.
-    """
-
-    INT8 = "INT8"
-    UINT8 = "UINT8"
+IntDtype = Literal[TensorDataType.int8, TensorDataType.uint8]
 
 
 class ExtendedQuantizerConfig(QuantizerConfig):
@@ -40,7 +34,7 @@ def __init__(
         signedness_to_force: Optional[bool] = None,
         per_channel: bool = QUANTIZATION_PER_CHANNEL,
         narrow_range: bool = QUANTIZATION_NARROW_RANGE,
-        dest_dtype: IntDtype = IntDtype.INT8,
+        dest_dtype: IntDtype = TensorDataType.int8,
     ):
         """
         :param num_bits: Bitwidth of the quantization.
@@ -54,6 +48,9 @@ def __init__(
         :param dest_dtype: Target integer data type for quantized values.
         """
         super().__init__(num_bits, mode, signedness_to_force, per_channel, narrow_range)
+        if dest_dtype not in [TensorDataType.int8, TensorDataType.uint8]:
+            msg = f"Quantization configurations with dest_dtype=={dest_dtype} are not supported."
+            raise nncf.ParameterNotSupportedError(msg)
         self.dest_dtype = dest_dtype
 
     def __str__(self) -> str:

src/nncf/experimental/torch/fx/quantization/quantizer/torch_ao_adapter.py

@@ -31,8 +31,8 @@
 from nncf.common.quantization.structs import QuantizationScheme as QuantizationMode
 from nncf.experimental.quantization.quantizer import Quantizer
 from nncf.experimental.quantization.structs import ExtendedQuantizerConfig
-from nncf.experimental.quantization.structs import IntDtype
 from nncf.experimental.torch.fx.nncf_graph_builder import GraphConverter
+from nncf.tensor.definitions import TensorDataType
 
 EdgeOrNode = Union[tuple[torch.fx.Node, torch.fx.Node]]
 
@@ -160,7 +160,7 @@ def get_quantizer_config_from_annotated_model(annotated: torch.fx.GraphModule) -
                 msg = f"Unknown qscheme: {qspec.qscheme}"
                 raise nncf.InternalError(msg)
 
-            dtype = IntDtype.INT8 if qspec.dtype is torch.int8 else IntDtype.UINT8
+            dtype = TensorDataType.int8 if qspec.dtype is torch.int8 else TensorDataType.uint8
             mode = (
                 QuantizationMode.SYMMETRIC
                 if qspec.qscheme in [torch.per_channel_symmetric, torch.per_tensor_symmetric]

src/nncf/quantization/algorithms/min_max/torch_fx_backend.py

@@ -28,7 +28,6 @@
 from nncf.experimental.common.tensor_statistics.collectors import REDUCERS_MAP
 from nncf.experimental.common.tensor_statistics.collectors import TensorReducerBase
 from nncf.experimental.quantization.structs import ExtendedQuantizerConfig
-from nncf.experimental.quantization.structs import IntDtype
 from nncf.experimental.torch.fx.commands import FXApplyTransformationCommand
 from nncf.experimental.torch.fx.model_utils import get_target_point
 from nncf.experimental.torch.fx.transformations import qdq_insertion_transformation_builder
@@ -38,6 +37,7 @@
 from nncf.quantization.fake_quantize import FakeConvertParameters
 from nncf.quantization.fake_quantize import FakeQuantizeParameters
 from nncf.quantization.range_estimator import StatisticsType
+from nncf.tensor.definitions import TensorDataType
 from nncf.torch.graph.graph import PTNNCFGraph
 from nncf.torch.graph.graph import PTTargetPoint
 from nncf.torch.graph.operator_metatypes import ELEMENTWISE_OPERATIONS
@@ -199,20 +199,20 @@ def _create_quantizer(
         if isinstance(quantizer_config, ExtendedQuantizerConfig):
             dtype = quantizer_config.dest_dtype
         elif quantizer_config.mode != QuantizationScheme.SYMMETRIC:
-            dtype = IntDtype.UINT8
+            dtype = TensorDataType.uint8
         else:
             dtype = (
-                IntDtype.INT8
+                TensorDataType.int8
                 if quantizer_config.signedness_to_force or torch.any(parameters.input_low.data < 0.0)
-                else IntDtype.UINT8
+                else TensorDataType.uint8
             )
 
         if per_channel:
             observer = torch.ao.quantization.observer.PerChannelMinMaxObserver
         else:
             observer = torch.ao.quantization.observer.MinMaxObserver
 
-        if dtype is IntDtype.INT8:
+        if dtype is TensorDataType.int8:
             level_high = 127
             level_low = -128
         else:
@@ -241,7 +241,7 @@ def _create_quantizer(
             observer=observer,
             quant_max=level_high,
             quant_min=level_low,
-            dtype=torch.qint8 if dtype is IntDtype.INT8 else torch.quint8,
+            dtype=torch.qint8 if dtype is TensorDataType.int8 else torch.quint8,
             qscheme=qscheme,
             eps=1e-16,
         )

src/nncf/tensor/definitions.py

@@ -38,6 +38,10 @@ class TensorDataType(StrEnum):
     Enum representing the different tensor data types.
     """
 
+    @staticmethod
+    def _generate_next_value_(name, start, count, last_values):
+        return name.lower()
+
     float16 = auto()
     bfloat16 = auto()
     float32 = auto()

tests/common/experimental/test_structs.py

@@ -0,0 +1,38 @@
+# 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.
+
+import pytest
+
+import nncf
+from nncf.experimental.quantization.structs import ExtendedQuantizerConfig
+from nncf.tensor.definitions import TensorDataType
+
+
+@pytest.mark.parametrize(
+    "dest_dtype",
+    [
+        TensorDataType.float16,
+        TensorDataType.bfloat16,
+        TensorDataType.float32,
+        TensorDataType.float64,
+        TensorDataType.f8e4m3,
+        TensorDataType.f8e5m2,
+        TensorDataType.nf4,
+        TensorDataType.int32,
+        TensorDataType.int64,
+        TensorDataType.uint4,
+        TensorDataType.int4,
+        None,
+    ],
+)
+def test_extended_q_config_non_supported_dest_dtype(dest_dtype):
+    with pytest.raises(nncf.ParameterNotSupportedError):
+        ExtendedQuantizerConfig(dest_dtype=dest_dtype)

tests/torch2/conftest.py

@@ -50,6 +50,7 @@ def pytest_configure(config: Config) -> None:
     regen_dot = config.getoption("--regen-ref-data", False)
     if regen_dot:
         os.environ["NNCF_TEST_REGEN_DOT"] = "1"
+        os.environ["NNCF_TEST_REGEN_JSON"] = "1"
 
     nncf_debug = config.getoption("--nncf-debug", False)
     if nncf_debug:

tests/torch2/fx/test_calculation_quantizer_params.py

@@ -26,6 +26,7 @@
 from nncf.quantization.algorithms.min_max.torch_fx_backend import FXMinMaxAlgoBackend
 from nncf.quantization.fake_quantize import calculate_quantizer_parameters
 from nncf.tensor import Tensor
+from nncf.tensor.definitions import TensorDataType
 
 INPUT_SHAPE = (2, 3, 4, 5)
 
@@ -79,7 +80,7 @@ class CaseQuantParams:
 
 
 @pytest.mark.parametrize("case_to_test", SYM_CASES)
-@pytest.mark.parametrize("dtype", [IntDtype.UINT8, IntDtype.INT8])
+@pytest.mark.parametrize("dtype", [TensorDataType.uint8, TensorDataType.int8])
 def test_quantizer_params_sym(case_to_test: CaseQuantParams, dtype: Optional[IntDtype]):
     per_ch = case_to_test.per_channel
     narrow_range = case_to_test.narrow_range
@@ -97,7 +98,7 @@ def test_quantizer_params_sym(case_to_test: CaseQuantParams, dtype: Optional[Int
     quantizer = _get_quantizer(case_to_test, qconfig)
     assert quantizer.qscheme is torch.per_channel_symmetric if case_to_test.per_channel else torch.per_tensor_symmetric
 
-    signed = signedness_to_force or dtype is IntDtype.INT8
+    signed = signedness_to_force or dtype is TensorDataType.int8
     if signed:
         assert torch.allclose(quantizer.zero_point, torch.tensor(0, dtype=torch.int8))
     else:
@@ -380,7 +381,7 @@ def test_quantizer_params_sym_nr(case_to_test: CaseQuantParams, ref_signed: bool
 
 
 @pytest.mark.parametrize("case_to_test,ref_zp", ASYM_CASES)
-@pytest.mark.parametrize("dtype", [IntDtype.UINT8, IntDtype.INT8])
+@pytest.mark.parametrize("dtype", [TensorDataType.uint8, TensorDataType.int8])
 def test_quantizer_params_asym(case_to_test: CaseQuantParams, ref_zp: Union[int, list[int]], dtype: Optional[IntDtype]):
     per_ch = case_to_test.per_channel
     narrow_range = case_to_test.narrow_range
@@ -397,7 +398,7 @@ def test_quantizer_params_asym(case_to_test: CaseQuantParams, ref_zp: Union[int,
     quantizer = _get_quantizer(case_to_test, qconfig)
     assert quantizer.qscheme is torch.per_channel_affine if case_to_test.per_channel else torch.per_tensor_affine
 
-    signed = dtype is IntDtype.INT8
+    signed = dtype is TensorDataType.int8
     ref_zp = torch.tensor(ref_zp)
     if not signed:
         ref_zp += 127 if narrow_range else 128

tests/torch2/fx/test_quantizer.py

@@ -36,14 +36,14 @@
 import nncf
 from nncf.common.graph import NNCFGraph
 from nncf.common.utils.os import safe_open
-from nncf.experimental.quantization.structs import IntDtype
 from nncf.experimental.torch.fx import quantize_pt2e
 from nncf.experimental.torch.fx.nncf_graph_builder import GraphConverter
 from nncf.experimental.torch.fx.node_utils import get_graph_node_by_name
 from nncf.experimental.torch.fx.quantization.quantizer.openvino_adapter import OpenVINOQuantizerAdapter
 from nncf.experimental.torch.fx.quantization.quantizer.openvino_quantizer import OpenVINOQuantizer
 from nncf.experimental.torch.fx.quantization.quantizer.torch_ao_adapter import TorchAOQuantizerAdapter
 from nncf.experimental.torch.fx.quantization.quantizer.torch_ao_adapter import _get_edge_or_node_to_qspec
+from nncf.tensor.definitions import TensorDataType
 from tests.cross_fw.shared.nx_graph import compare_nx_graph_with_reference
 from tests.cross_fw.shared.paths import TEST_ROOT
 from tests.torch import test_models
@@ -256,6 +256,8 @@ def _normalize_qsetup_state(setup: dict[str, Any]) -> None:
     for qp in setup["quantization_points"].values():
         sorted_dq = sorted(qp[dq_key])
         qconfig = qp["qconfig"].copy()
+        if "dest_dtype" in qconfig:
+            qconfig["dest_dtype"] = "INT8" if qconfig["dest_dtype"] is TensorDataType.int8 else "UINT8"
         sorted_qps[f"{tuple(sorted_dq)}_{qp['qip_class']}"] = qconfig
     setup["quantization_points"] = sorted_qps
 
@@ -285,7 +287,9 @@ def _normalize_nncf_graph(nncf_graph: NNCFGraph, fx_graph: torch.fx.Graph):
             idx += 1
             if node.node_type in ["dequantize_per_tensor", "dequantize_per_channel"]:
                 source_node = get_graph_node_by_name(fx_graph, node.node_name)
-                dtypes_map[new_node_name] = IntDtype.INT8 if source_node.args[-1] == torch.int8 else IntDtype.UINT8
+                dtypes_map[new_node_name] = (
+                    TensorDataType.int8 if source_node.args[-1] == torch.int8 else TensorDataType.uint8
+                )
         norm_nncf_graph.add_nncf_node(
             node_name=attrs[node.NODE_NAME_ATTR],
             node_type=attrs[node.NODE_TYPE_ATTR],