update requirements and warning messages

optimum/exporters/openvino/__main__.py

@@ -283,7 +283,9 @@ class StoreAttr(object):
 
     synonyms_for_task = TasksManager.synonyms_for_task(task)
     synonyms_for_task.add(task)
-    if stateful and not ensure_export_task_support_stateful(task):
+
+    task_support_stateful = ensure_export_task_support_stateful(task)
+    if stateful and not task_support_stateful:
         stateful = False
 
     preprocessors = maybe_load_preprocessors(

optimum/exporters/openvino/convert.py

@@ -309,7 +309,7 @@ def export_pytorch(
             `int4_sym_g64` - INT4 symmetric weights w/ group size 64, "int4_asym_g64" - as previous but asymmetric w/ zero-point.
         compression_ratio (`Optional[float]`, defaults to `None`):
             Compression ratio between primary and backup precision (only relevant to INT4).
-        stateful (`Optional[bool]`, defaults to `True`):
+        stateful (`Optional[bool]`, defaults to `False`):
             Produce stateful model where all kv-cache inputs and outputs are hidden in the model and are not exposed as model inputs and outputs
 
     Returns:
@@ -413,7 +413,8 @@ def ts_patched_forward(*args, **kwargs):
                 # cannot raise because stateful is enabled by default and it would break backward compatibility for models that couldn't convert to OV directly
                 # TODO: Implement stateful for ONNX path as well, not doing it right now because of lack of validation
                 logger.warn(
-                    "[ WARNING ] Making stateful models is not supported when exporting to ONNX as an intermediate step. A stateless model will be exported instead. It may result in sub-optimal inference performance."
+                    "[ WARNING ] Making stateful models is not supported when exporting to ONNX as an intermediate step. "
+                    "A stateless model will be exported instead. It may result in sub-optimal inference performance."
                     "Provide a model that can be converted to OpenVINO without fallback to ONNX conversion path."
                 )
             return export_pytorch_via_onnx(

optimum/exporters/openvino/stateful.py

@@ -192,7 +192,8 @@ def ensure_stateful_is_available(warn=True):
     if is_openvino_version("<", "2023.3"):
         if warn:
             log.warn(
-                f"Could not create or use stateful model when using old version of openvino=={_openvino_version}. Install openvino>=2023.3.0."
+                f"Could not create or use stateful model when using old version of openvino=={_openvino_version}. It may result in sub-optimal inference performance."
+                "Install openvino>=2023.3.0."
             )
         return False
     return True

optimum/intel/openvino/modeling_decoder.py

@@ -141,9 +141,16 @@ def __init__(
         self.update_pkv_precision()
         if self.is_dynamic:
             self.model = self._reshape(self.model, -1, -1)
+        is_stateful_supported = ensure_stateful_is_available(warn=False)
+
+        if self.use_cache and not self.stateful:
+            logger.warn(
+                "Provided model does not contain state. It may lead to sub-optimal performance."
+                "Please reexport model with updated OpenVINO version >= 2023.3.0 calling the `from_pretrained` method with original model "
+                "and `export=True` parameter"
+            )
 
         if self.stateful:
-            is_stateful_supported = ensure_stateful_is_available()
             if stateful is None:
                 stateful = is_stateful_supported
             if model_has_sinks and not is_stateful_supported:

setup.py

@@ -43,7 +43,7 @@
         "onnxruntime<1.15.0",
         "transformers>=4.34.0",
     ],
-    "openvino": ["openvino>=2023.2", "onnx", "onnxruntime", "transformers>=4.34.0"],
+    "openvino": ["openvino>=2023.2", "onnx", "onnxruntime", "transformers>=4.36.0", "optimum>=1.16.1"],
     "nncf": ["nncf>=2.7.0"],
     "ipex": ["transformers<4.32.0", "intel-extension-for-pytorch", "onnx"],
     "diffusers": ["diffusers"],