Python: QNN : profiling_file_path does not dump profiling events to CSV

[DavidLuong98]

I am working on profiling an onnx model using onnxruntime-qnn version 1.20.0 with QNN execution provider. However, the qnn profile events are never dumped to the csv.

Below is a sample code that I am working with:

import onnxruntime as ort

profile_csv_path = "some_path.csv"
provider_option = [
    {
        'backend_path': 'QnnHtp.dll',
        'htp_performance_mode': 'sustained_high_performance',
        'htp_graph_finalization_optimization_mode': '3',
        'soc_model': '60',
        'htp_arch': '73',
        'profiling_level': 'detailed',
        'profiling_file_path': profile_csv_path,
    }
]

sess_options = ort.SessionOptions()
sess_options.enable_profiling = True

model_path = "some_onnxmodel.onnx"

session = ort.InferenceSession(
        model_path,
        sess_options=sess_options,
        providers=['QNNExecutionProvider'],
        provider_options=provider_option)
        
batch_inputs = { # some batch input }
session.run(None, batch_inputs)
session.end_profiling()

I would have expected the dump of profiling_file_path to be written to profile_csv_path. However, the file remains empty.
An onnxruntime_profile__2025-02-04_19-10-30.json is still output from sess_options.enable_profiling = True which indicates that the model inference worked fine.

I manually created the csv since it seemed like the source code does not make the csv file when not present on disk. This still did not work.

It's quite hard to debug because and tell what's going on. It's difficult to indicate whether this is a developer api misuage, a bug within the source code, or some issue with the model. Should there be some feedback log when the qnn events are dumped to the csv ?

Lastly, my code could be off. Please let me know if there's a bug. Thank you.

[DavidLuong98]

couple issues :

1. I had installed onnxruntime-qnn and onnxruntime installed.
2. I am on an ARM64 device and installed AMD64 python by accident 🤣