Some ONNX model takes over 50mins to load

[lorenzo-park]

What would be the cause?

I exported a model to ONNX format. However, its loading time is extremely slow, like over 40mins to load it. The model file size is about 1.16GB. I tried the other model file(This file is >2GB, so it contains weight files as well) but the loading time is less than 1min.

sess_opt = ort.SessionOptions()
        sess_opt.log_severity_level = 0
        self.onnx_model_session = ort.InferenceSession('export_output/model_original_sampler/model.onnx', sess_opt, providers=['CPUExecutionProvider'])

...
2025-02-10 18:09:47.622645747 [I:onnxruntime:, graph_transformer.cc:15 Apply] GraphTransformer MemcpyTransformer modified: 0 with status: OK
2025-02-10 18:09:47.713474820 [V:onnxruntime:, session_state.cc:1148 VerifyEachNodeIsAssignedToAnEp] Node placements
2025-02-10 18:09:47.713592805 [V:onnxruntime:, session_state.cc:1151 VerifyEachNodeIsAssignedToAnEp]  All nodes placed on [CPUExecutionProvider]. Number of nodes: 191006
2025-02-10 18:09:48.419443741 [V:onnxruntime:, session_state.cc:128 CreateGraphInfo] SaveMLValueNameIndexMapping
2025-02-10 18:09:48.553279738 [V:onnxruntime:, session_state.cc:174 CreateGraphInfo] Done saving OrtValue mappings.
2025-02-10 18:09:48.553347095 [I:onnxruntime:, allocation_planner.cc:2567 CreateGraphPartitioner] Use DeviceBasedPartition as default
2025-02-10 18:56:40.416055734 [I:onnxruntime:, session_state_utils.cc:276 SaveInitializedTensors] Saving initialized tensors.

As you see in the log, the loading takes almost 50mins from the moment Use DeviceBasedPartition as default is logged to the moment Saving initialized tensors. is logged.

I am trying to dig what is happening here, but no clues. Here is the model file in google drive https://drive.google.com/file/d/1C3lF5IkL3DDvAh7-Ktg6i8R32pzS3DiL/view?usp=sharing

Versions
onnx==1.17.0
onnxruntime==1.20.1
onnxruntime-gpu==1.20.1