[tmva][sofie] Fix reading of ONNX external weight data

[guitargeek]

The external-data reading in the ONNX parser had two problems:

• The data file name was resolved once per parser instance: it defaulted to .onnx.data on the first Parse call and was never reset, and the ifstream stayed open across calls. Reusing one parser instance for several models with external data silently read every subsequent model's weights at the stored offsets of the first model's data file, producing garbage weights (or short reads leaving malloc'ed buffers uninitialized).

• The "location" key of a tensor's external_data, which per the ONNX spec names the data file relative to the model directory, was ignored entirely, so files whose data location does not follow the .onnx.data convention could not be read at all.

Resolve the data file per tensor: an explicitly set file (SetExternalDataFile) takes precedence, then the tensor's stored location relative to the model directory, then the conventional .onnx.data. Track which file is open, and reset the whole external-data state after each Parse call.

This surfaced with models exported by the torch.export-based ONNX exporter of PyTorch 2.x, which stores initializers externally.

Add a unit test covering the resolution of the external data file (stored location relative to the model directory, per-Parse-call default name, SetExternalDataFile precedence) and the reset of the state between Parse calls. The test hand-writes the minimal ONNX protobuf wire format, so it does not depend on the onnx Python package.

🤖 Done with the help of AI

[github-actions]

Test Results

    22 files      22 suites   3d 11h 17m 10s ⏱️
 3 870 tests  3 868 ✅ 0 💤 2 ❌
75 967 runs  75 959 ✅ 6 💤 2 ❌

For more details on these failures, see this check.

Results for commit d45bce4.