Replaced artifacts with JLL artifacts

[stemann]

Using extended platform selection based on platform augmentation tags, i.e. https://pkgdocs.julialang.org/v1.7/artifacts/#Extending-Platform-Selection - adapted for use with JLL Artifacts.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 88.67%. Comparing base (f79d863) to head (0cf858d).
Report is 38 commits behind head on main.

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #19      +/-   ##
==========================================
+ Coverage   87.98%   88.67%   +0.68%     
==========================================
  Files           3        3              
  Lines         383      371      -12     
==========================================
- Hits          337      329       -8     
+ Misses         46       42       -4     

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

[jw3126]

Thanks a lot @stemann . I have never looked at augment_platform! I will read the docs, before I can give any meaningful feedback. Please feel free to ping me, should I forget to come back to this PR in the next week.

[jw3126]

CUDA tests fail for me locally on this branch. I also discovered that my way of determining whether to run CUDA.jl tests does not work with recent CUDA versions and fixed that #20

[stemann]

Yes - it’s not even remotely ready.

CUDA binaries will not be available until JuliaPackaging/Yggdrasil#4386 is merged.

Current ONNXRuntime JLL provides “eager” artifacts. The CUDA EP PR will also make artifacts lazy.

The platform selection stuff is also only documented for Pkg/Julia v1.7, so that might also break v1.6 - and seems to be only/mostly relevant for “eager” artifacts.

[stemann]

I stumbled a bit on the platform selection thing. I think what needs to be done is to have the ONNXRuntime JLL re-built using platform augmentation...

[jw3126]

Very cool to hear from you again! If there is a clean way to make the jll artifacts work, that would be awesome.

[stemann]

Yeah - sorry - life happened :-) I'll try to find the time to finish this... (too...)

[jw3126]

One general comment. The current approach we take to artifacts has many downsides, but one upside. It is quite simple, no need to compile anything we simply reuse binaries which onnxruntime and CUDA.jl provide. Also no usage of Artifacts.jl internals etc.

If we go with your approach, it would be important to me, that the thing stays easy to maintain.
Which means only use documented Artifacts.jl features, share artifacts with other packages (or onnxruntime) where possible and compile scripts that are simple to adapt to new versions.

[stemann]

Of course - it will of course get a little harder, as there is one more package (the JLL) to update and release, but it shouldn't get much harder.

I can ping you on PRs for updates of the JLL recipe, so you can easily see the path for upgrades.

[jw3126]

Of course - it will of course get a little harder, as there is one more package (the JLL) to update and release, but it shouldn't get much harder.

I can ping you on PRs for updates of the JLL recipe, so you can easily see the path for upgrades.

Yeah I need to do an update myself.

[stemann]

I have just pushed JuliaPackaging/Yggdrasil#7561 which uses the same platform selection/augmentation block as CUDNN, i.e. the CUDA Runtime version is controlled via the CUDA_Runtime_jll version and/or local preferences.

It "also" builds the CUDA/CUDNN and TensorRT artifacts from source.

[stemann]

... still awaiting JuliaPackaging/Yggdrasil#7623 (needed for JuliaPackaging/Yggdrasil#7561)