Implement tofile on tensors to reduce data write time by 40%

[justinchuby]

This PR introduces the tofile method on tensors (similarly named as the one on numpy arrays), which allows for faster write and lower memory usage on external data by bypassing tobytes().

Compatibility with existing TensorProtocols is maintained in the external data module by using tofile only when it is available in the class. The TorchTensor class in PyTorch exporter should be updated accordingly to leverage the new logic when saving.

Note that io time to disk is reduced by 40% below.

Note

TensorProtocol is not updated because we do isinstance() checks on external implementations (PyTorch). Adding the method in the protocol will cause isinstance check to fail on those implementations that have not added the tofile method.

Reference: https://github.com/microsoft/onnxscript/pull/2241/files/b2381658492510a9bcc8c0a8574db7368e33bceb

Before:

________________________________________________________
Executed in   48.08 secs    fish           external
   usr time   60.54 secs    0.00 millis   60.54 secs
   sys time   23.06 secs    1.22 millis   23.06 secs

After:

________________________________________________________
Executed in   45.69 secs    fish           external
   usr time   60.68 secs  244.00 micros   60.68 secs
   sys time   22.22 secs  518.00 micros   22.22 secs

Fix #207

[codecov]

Codecov Report

❌ Patch coverage is 81.96721% with 11 lines in your changes missing coverage. Please review.
✅ Project coverage is 76.92%. Comparing base (feb51e5) to head (e3df4c9).
⚠️ Report is 3 commits behind head on main.

Files with missing lines	Patch %	Lines
src/onnx_ir/_core.py	80.85%	7 Missing and 2 partials ⚠️
src/onnx_ir/external_data.py	66.66%	1 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #210      +/-   ##
==========================================
+ Coverage   76.83%   76.92%   +0.08%     
==========================================
  Files          40       40              
  Lines        4922     4992      +70     
  Branches      980      996      +16     
==========================================
+ Hits         3782     3840      +58     
- Misses        856      864       +8     
- Partials      284      288       +4     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[justinchuby]

cc @iksnagreb

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[justinchuby]

@titaiwangms @gramalingam this is ready for review, thanks.

[titaiwangms]

I am wondering whether tofile() makes sense to LazyTensor. hmm

[justinchuby]

Can you say more?

[titaiwangms]

I just thought it's not even real until it's evaluated. Intuitively, not very suitable with tofile(), which we want to write it to disk. But I guess in general expectation, we want all tensors have this method. It's understandable.

[justinchuby]

It is actually useful: even when the tensor is lazily evaluated, we still want to avoid tobytes() making a copy of the tensor data before writing to file. The screenshots on the PR description are showing lazy tensors.