Comments

Author: daniil-lyakhov

src/nncf/experimental/quantization/structs.py

@@ -11,6 +11,7 @@
 
 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
@@ -49,7 +50,7 @@ def __init__(
         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 RuntimeError(msg)
+            raise nncf.ParameterNotSupportedError(msg)
         self.dest_dtype = dest_dtype
 
     def __str__(self) -> str:

tests/common/experimental/test_structs.py

@@ -11,12 +11,28 @@
 
 import pytest
 
+import nncf
 from nncf.experimental.quantization.structs import ExtendedQuantizerConfig
 from nncf.tensor.definitions import TensorDataType
 
 
-def test_extended_q_config_non_supported_dest_dtype():
-    with pytest.raises(RuntimeError):
-        ExtendedQuantizerConfig(dest_dtype=TensorDataType.int4)
-    with pytest.raises(RuntimeError):
-        ExtendedQuantizerConfig(dest_dtype=None)
+@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/post_training/pipelines/image_classification_base.py

@@ -32,7 +32,6 @@
 from nncf.common.logging.track_progress import track
 from nncf.experimental.torch.fx import OpenVINOQuantizer
 from nncf.experimental.torch.fx import quantize_pt2e
-from nncf.torch import disable_patching
 from tests.post_training.pipelines.base import DEFAULT_VAL_THREADS
 from tests.post_training.pipelines.base import FX_BACKENDS
 from tests.post_training.pipelines.base import BackendType
@@ -130,7 +129,7 @@ def _validate(self) -> None:
         return []
 
     def _compress_torch_ao(self, quantizer):
-        with torch.no_grad(), disable_patching():
+        with torch.no_grad():
             prepared_model = prepare_pt2e(self.model, quantizer)
             subset_size = self.compression_params.get("subset_size", 300)
             for data in islice(self.calibration_dataset.get_inference_data(), subset_size):
@@ -167,7 +166,7 @@ def _compress_nncf_pt2e(self, quantizer):
         if self.compression_params.get("model_type", False):
             smooth_quant = self.compression_params["model_type"] == nncf.ModelType.TRANSFORMER
 
-        with disable_patching(), torch.no_grad():
+        with torch.no_grad():
             self.compressed_model = quantize_pt2e(
                 self.model,
                 quantizer,
@@ -186,7 +185,7 @@ def _compress(self):
 
             return
         if self.backend in [BackendType.FX_TORCH, BackendType.CUDA_FX_TORCH]:
-            with disable_patching(), torch.no_grad():
+            with torch.no_grad():
                 super()._compress()
                 return
 

tests/torch2/conftest.py

@@ -50,7 +50,6 @@ 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: