[ITensors] [BUG] Incompatibility between ITensors.jl and CUDA.jl v5.7.1

[zipeilee]

I am experiencing compatibility issues between ITensors.jl and CUDA.jl when using the latest CUDA.jl version (v5.7.1). Here is the behavior I observed:

When I update CUDA.jl to v5.7.1（(the latest version ), ITensors.jl is downgraded to v0.7.13, and ITensorMPS.jl is downgraded to v0.3.6.

If I try to upgrade ITensorMPS.jl to v0.3.16 or ITensors.jl to v0.9.0, then CUDA.jl is downgraded to v5.5.2 instead.

This prevents me from using the latest features of both ITensors.jl and CUDA.jl simultaneously. Could you clarify whether there are plans to support CUDA.jl v5.7.1 in the latest versions of ITensors.jl and ITensorMPS.jl?

My environment:

OS: Manjaro Linux

Julia version: 1.11.4

CUDA.jl version: v5.7.1 (before downgrade)

ITensors.jl version: v0.9.0 (before downgrade)

ITensorMPS.jl version: v0.3.16 (before downgrade)

Let me know if you need additional information. Thanks!

[mtfishman]

CUDA.jl is based on GPUArrays.jl. The latest versions of CUDA.jl are based on GPUArrays.jl v11, which did a big refactor to be based on KernelAbstractions.jl. Unfortunately that change broke some of our tests, so we weren't able to support it (#1599).

I can look into if there are simple workarounds for the issues we hit, I was hoping they would get fixed in GPUArrays.jl but I don't think they have yet.

[zipeilee]

Thanks for the explanation! I see that the transition to GPUArrays.jl v11 caused issues with some tests, preventing ITensors.jl from supporting the latest CUDA.jl versions.

In this case, do you think there are any workarounds I could try on my end to get ITensors.jl working with CUDA.jl v5.7.1? Or would it be better to stay on an older CUDA.jl version for now?

Additionally, if providing logs or testing potential fixes would be helpful, I’d be happy to assist. Let me know what you think!

[mtfishman]

It turns out the original issues I saw are now fixed, so I've upgraded the versions here: #1641. I'll close since this should be addressed in the next NDTensors.jl release.