export part poc

optimum/exporters/openvino/model_configs.py

@@ -15,8 +15,6 @@
 from copy import deepcopy
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Union
 from optimum.exporters.onnx.base import ConfigBehavior
-import logging
-logger = logging.getLogger(__name__)
 
 from packaging import version
 from transformers import PretrainedConfig
@@ -836,7 +834,6 @@ def patch_model_for_export(
 class WhisperOpenVINOConfig(WhisperOnnxConfig):
     def __init__(self, config: PretrainedConfig, task: str = "feature-extraction", int_dtype: str = "int64", float_dtype: str = "fp32", use_past: bool = False, use_past_in_inputs: bool = False, behavior: ConfigBehavior = ConfigBehavior.MONOLITH, preprocessors: Optional[List[Any]] = None, legacy: bool = False, stateful: bool = False):
         self.stateful = stateful
-        logger.warn(f"config stateful: {self.stateful}")
         super().__init__(config, task, int_dtype, float_dtype, use_past, use_past_in_inputs, behavior, preprocessors, legacy)
 
 
@@ -847,12 +844,10 @@ def _create_dummy_input_generator_classes(self, **kwargs) -> List[DummyInputGene
         forces the other generators to use the same batch size, meaning they will all produce inputs of the same batch
         size. Override this method for custom behavior.
         """
-        logger.warn(f"config stateful: {self.stateful}")
         if self.stateful:
             if "encoder_sequence_length" not in kwargs:
                 sequence_len = kwargs.get("sequence_length", DEFAULT_DUMMY_SHAPES["sequence_length"])
                 kwargs["encoder_sequence_length"] = sequence_len + 2
-        logger.warn(kwargs)
         return super()._create_dummy_input_generator_classes(**kwargs)
 
     def with_behavior(

optimum/exporters/openvino/stateful.py

@@ -196,7 +196,6 @@ def ensure_export_task_support_stateful(task: str, is_encoder_decoder:bool = Fal
     )
 
     is_stateful = task.endswith("-with-past") and task.replace("-with-past", "") in _ENCODER_DECODER_TASKS_WITH_PAST
-    log.warn(f"is_stateful {is_stateful}")
     return is_stateful
 
 

optimum/intel/openvino/modeling_base_seq2seq.py

@@ -100,7 +100,7 @@ def _save_pretrained(self, save_directory: Union[str, Path]):
         """
         src_files = [self.encoder_model, self.decoder_model]
         dst_file_names = [OV_ENCODER_NAME, OV_DECODER_NAME]
-        if self.use_cache:
+        if self.decoder_with_past_model is not None:
             src_files.append(self.decoder_with_past_model)
             dst_file_names.append(OV_DECODER_WITH_PAST_NAME)