Introduce OVQuantizationConfig for nncf.quantize() parameters

optimum/intel/__init__.py

@@ -124,6 +124,7 @@
             "OVModelForVision2Seq",
             "OVModelForSequenceClassification",
             "OVModelForTokenClassification",
+            "OVQuantizationConfig",
             "OVWeightQuantizationConfig",
             "OVConfig",
         ]
@@ -243,6 +244,7 @@
             OVModelForSpeechSeq2Seq,
             OVModelForTokenClassification,
             OVModelForVision2Seq,
+            OVQuantizationConfig,
             OVWeightQuantizationConfig,
         )
 

optimum/intel/openvino/__init__.py

@@ -43,7 +43,7 @@
         from .trainer import OVTrainer
 
 
-from .configuration import OVConfig, OVWeightQuantizationConfig
+from .configuration import OVConfig, OVQuantizationConfig, OVWeightQuantizationConfig
 from .modeling import (
     OVModelForAudioClassification,
     OVModelForAudioFrameClassification,

optimum/intel/openvino/modeling_decoder.py

@@ -640,7 +640,7 @@ def _from_pretrained(
                 # from optimum.gptq.utils import get_seqlen
 
                 # seqlen = get_seqlen(causal_model)
-                nsamples = quantization_config.num_samples if quantization_config.num_samples else 128
+                nsamples = quantization_config.subset_size if quantization_config.subset_size else 128
                 dataset = get_dataset(quantization_config.dataset, tokenizer, seqlen=32, nsamples=nsamples)
                 dataset = prepare_dataset(dataset)
                 quantization_config = copy.deepcopy(quantization_config)

optimum/intel/openvino/modeling_diffusion.py

@@ -321,7 +321,7 @@ def _from_pretrained(
             if not isinstance(sd_model, supported_pipelines):
                 raise NotImplementedError(f"Quantization in hybrid mode is not supported for {cls.__name__}")
 
-            nsamples = quantization_config.num_samples if quantization_config.num_samples else 200
+            nsamples = quantization_config.subset_size if quantization_config.subset_size else 200
             unet_inputs = sd_model._prepare_unet_inputs(quantization_config.dataset, nsamples)
 
             from .quantization import _hybrid_quantization

optimum/intel/openvino/trainer.py

@@ -89,7 +89,7 @@
 
 from ..utils.constant import _TASK_ALIASES
 from ..utils.import_utils import is_transformers_version
-from .configuration import DEFAULT_QUANTIZATION_CONFIG, OVConfig
+from .configuration import OVConfig
 from .quantization import OVDataLoader
 from .training_args import OVTrainingArguments
 from .utils import (
@@ -136,6 +136,25 @@
 NNCF_LOG_FILE_NAME = "nncf_output.log"
 
 
+DEFAULT_QUANTIZATION_CONFIG = {
+    "algorithm": "quantization",
+    "preset": "mixed",
+    "overflow_fix": "disable",
+    "initializer": {
+        "range": {"num_init_samples": 300, "type": "mean_min_max"},
+        "batchnorm_adaptation": {"num_bn_adaptation_samples": 0},
+    },
+    "scope_overrides": {"activations": {"{re}.*matmul_0": {"mode": "symmetric"}}},
+    "ignored_scopes": [
+        "{re}.*Embedding.*",
+        "{re}.*add___.*",
+        "{re}.*layer_norm_.*",
+        "{re}.*matmul_1",
+        "{re}.*__truediv__.*",
+    ],
+}
+
+
 def _onnx_export_nncf_model(model: NNCFNetwork, config: OnnxConfig, output: Union[str, io.BytesIO], opset: int = None):
     # TODO: remove it when fix controller.strip(copy=True) behavior
     signature = inspect.signature(model.forward)
@@ -228,6 +247,16 @@ def __init__(
         if self.ov_config is not None:
             if self.ov_config.compression is None:
                 self.ov_config.compression = DEFAULT_QUANTIZATION_CONFIG
+            if (
+                isinstance(self.ov_config.compression, dict)
+                and "algorithm" in self.ov_config.compression
+                and self.ov_config.compression["algorithm"] == "quantization"
+            ):
+                self.ov_config.compression["export_to_onnx_standard_ops"] = self.ov_config.save_onnx_model
+            elif isinstance(self.ov_config.compression, list):
+                for i, algo_config in enumerate(self.ov_config.compression):
+                    if algo_config["algorithm"] == "quantization":
+                        self.ov_config.compression[i]["export_to_onnx_standard_ops"] = self.ov_config.save_onnx_model
 
             if self.args.do_train:
                 self._set_task()