Add dynamic quantization config (#661)

* add warning

* add model arch

* remove calibration dataset argument

* format

* remove comments

* replace preset with sym for compatibility between configs

* format

* add dynamic quantization

* add dynamic config

* remove test deprecated config parameter

* add bits and sym to base config
* add config test

* format

* add kv cache precision

* format

* add test

* move compilation step

* set kv cache precision for seq2seq models

Author: echarlaix

optimum/intel/__init__.py

@@ -60,11 +60,23 @@
         raise OptionalDependencyNotAvailable()
 except OptionalDependencyNotAvailable:
     _import_structure["utils.dummy_openvino_and_nncf_objects"].extend(
-        ["OVQuantizer", "OVTrainingArguments", "OVQuantizationConfig", "OVWeightQuantizationConfig"]
+        [
+            "OVQuantizer",
+            "OVTrainingArguments",
+            "OVQuantizationConfig",
+            "OVWeightQuantizationConfig",
+            "OVDynamicQuantizationConfig",
+        ]
     )
 else:
     _import_structure["openvino"].extend(
-        ["OVQuantizer", "OVTrainingArguments", "OVQuantizationConfig", "OVWeightQuantizationConfig"]
+        [
+            "OVQuantizer",
+            "OVTrainingArguments",
+            "OVQuantizationConfig",
+            "OVWeightQuantizationConfig",
+            "OVDynamicQuantizationConfig",
+        ]
     )
 
 
@@ -192,13 +204,20 @@
             raise OptionalDependencyNotAvailable()
     except OptionalDependencyNotAvailable:
         from .utils.dummy_openvino_and_nncf_objects import (
+            OVDynamicQuantizationConfig,
             OVQuantizationConfig,
             OVQuantizer,
             OVTrainingArguments,
             OVWeightQuantizationConfig,
         )
     else:
-        from .openvino import OVQuantizationConfig, OVQuantizer, OVTrainingArguments, OVWeightQuantizationConfig
+        from .openvino import (
+            OVDynamicQuantizationConfig,
+            OVQuantizationConfig,
+            OVQuantizer,
+            OVTrainingArguments,
+            OVWeightQuantizationConfig,
+        )
 
     try:
         if not (is_openvino_available() and is_nncf_available() and is_accelerate_available()):

optimum/intel/openvino/__init__.py

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