On downside of this, I realized, is that this solution only works for data layouts that end with the field index, and only works on Julia 1.11.

An ideal solution would be to use an immutable version ofRef. We’ve had issues with Ref in the past, which is why we switched to using tuples.

We tried to recreate this issue with ordinary cuarrays, but we weren’t able to reproduce the error. Which suggests that there is something else different. Perhaps it’s worth taking a second look at this, since it could lead to a fix on our end, or a fix in GPUCompiler/julia inference.

I did manage to make a reproducer in JuliaGPU/CUDA.jl#2623. I think the proper path forward is to further simplify the reproducer in that issue and resolve the compiler limitation (whether that be in Base or in CUDA.jl), so that this issue is properly fixed. I'm going to close this PR, since it's really only a monkey patch that is not really widely applicable.