forked from microsoft/onnxruntime
-
Notifications
You must be signed in to change notification settings - Fork 52
Open
Description
Describe the issue
If a GPT-like ONNX model contains SkipLayerNormalization
nodes, an error occurs during loading with the OpenVINOExecutionProvider
; however, the model works well with other ExecutionProviders, such as CPU, CUDA, and DML. The following is the error message:
line 561, in _create_inference_session
sess.initialize_session(providers, provider_options, disabled_optimizers)
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
onnxruntime.capi.onnxruntime_pybind11_state.RuntimeException: [ONNXRuntimeError] : 6 : RUNTIME_EXCEPTION : Exception during initialization: /onnxruntime/onnxruntime/core/providers/openvino/ov_interface.cc:79 std::shared_ptr<ov::Model> onnxruntime::openvino_ep::OVCore::ReadModel(std::string&&, const std::string&) [OpenVINO-EP] [OpenVINO-EP] Exception while Reading network: Check 'onnx_node.get_outputs_size() <= outputs_size' failed at src/frontends/onnx/frontend/src/core/graph.cpp:392:
FrontEnd API failed with GeneralFailure:
Expected output number of SkipLayerNormalization node is 4 while the implementation provides 1 outputs
To reproduce
import onnxruntime
onnx_model_A = "Whisper_Encoder.onnx"
provider_options = [
{
'device_type': 'CPU',
'precision': 'ACCURACY',
'num_of_threads': 1,
'num_streams': 1,
'enable_opencl_throttling': True,
'enable_qdq_optimizer': False,
'disable_dynamic_shapes': False
}
]
session_opts = onnxruntime.SessionOptions()
session_opts.log_severity_level = 0
ort_session_A = onnxruntime.InferenceSession(onnx_model_A, sess_options=session_opts, providers=['OpenVINOExecutionProvider'], provider_options=provider_options)
Urgency
No response
Platform
Linux
OS Version
Ubuntu 24.04
ONNX Runtime Installation
Released Package
ONNX Runtime Version or Commit ID
1.22.1
ONNX Runtime API
Python
Architecture
X64
Execution Provider
OpenVINO
Execution Provider Library Version
1.22.0
Model File
Is this a quantized model?
No
Metadata
Metadata
Assignees
Labels
No labels