Revert the building container back to ubuntu 20.04

[LittleLittleCloud]

unblock dotnet/machinelearning#7182

Need to wait the PR #1351 to be merged in first

Why not reverting back to ubuntu 22.04

The glibcxx in ubuntu 22.04 is 3.4.30 and is newer than the one in centos-streaming9 (3.4.29)

[LittleLittleCloud]

@NiklasGustafsson We might still need to revert container back to ubuntu20? Let's still keep this PR open for a while until the ml.net repo side has verified that the libtorch 2.2.1 work with ubuntu 20 | centos 9 helix test machine

[NiklasGustafsson]

@NiklasGustafsson We might still need to revert container back to ubuntu20? Let's still keep this PR open for a while until the ml.net repo side has verified that the libtorch 2.2.1 work with ubuntu 20 | centos 9 helix test machine

@LittleLittleCloud -- still the case?

[LittleLittleCloud]

@NiklasGustafsson

Still the case. Just get some time to verify it on centos streaming 9.

Good news is libtorch.so can be loaded on centos 9, bad news is libLibTorchsharp requires a higher glibc version.

So we might still need to convert the building container back (to ubuntu 20.2 this time), or use a lower version glibcxx when compile libtorchsharp, or install a newer glibcxx in helix test machine

[NiklasGustafsson]

Before publishing to NuGet, we need the built packages tested with ML.NET, so that we're not releasing something that doesn't actually solve the problem.

[LittleLittleCloud]

Before publishing to NuGet, we need the built packages tested with ML.NET, so that we're not releasing something that doesn't actually solve the problem.

I have verified the cpu package with mlnet. Where can I verify the cuda package

[NiklasGustafsson]

Before publishing to NuGet, we need the built packages tested with ML.NET, so that we're not releasing something that doesn't actually solve the problem.

I have verified the cpu package with mlnet. Where can I verify the cuda package

You're the one who has a 4090, right? :-)

[LittleLittleCloud]

Before publishing to NuGet, we need the built packages tested with ML.NET, so that we're not releasing something that doesn't actually solve the problem.

I have verified the cpu package with mlnet. Where can I verify the cuda package

You're the one who has a 4090, right? :-)

Yes, I also have 1060 and 4050 as well, so a lot of scenarios with GPU I can test with

[michaelgsharp]

Just curious, how much does the compiler affect the cuda builds?

[NiklasGustafsson]

Just curious, how much does the compiler affect the cuda builds?

It's not the compiler as much as libc that's of concern. Since the CUDA libtorch is a separate distribution, it should be tested, and the build pipeline does not include any GPU machines.

[michaelgsharp]

@LittleLittleCloud I can show you how to enable the GPU for the big tests we have if you haven't done them before. We have at least 1 big test for each scenario in ML.NET that should fully exercise the gpu. Or at least for the stuff its touching anyways.

[LittleLittleCloud]

@michaelgsharp That would be great, I will reach out to you privately