RomeoV

Here is a github gist which reproduces the error: https://gist.github.com/RomeoV/ca397a6b883c1cf567f2503d135084d8 The setup is generally inspired by [the VAE tutorial in the FastAI doc](https://fluxml.ai/FastAI.jl/dev/FastAI@dev/doc/docs/notebooks/vae.ipynb.html).

Trying out this code yields an error ```julia julia> A_gpu = Dagger.@spawn CUDA.Matrix(A) Error in eager scheduler: MethodError: no method matching get_parent(::CUDAExt.CuArrayDeviceProc) ``` The Readme in this repo points in...

Actually, it seems `DaggerGPU.CuArrayDeviceProc` isn't even defined anymore... However, I did find the following line in the tests: https://github.com/JuliaGPU/DaggerGPU.jl/blob/319a71e44587dc0983dc4cd5c64f18d672a3e44d/test/runtests.jl#L83C1-L86C12 ```julia c = Dagger.with_options(;scope=Dagger.scope(cuda_gpu=1)) do @test fetch(Dagger.@spawn isongpu(b)) Dagger.@spawn sum(b) end...

The error still persists if replace the failing call with a more pessimistic ```julia for i in 1:100 CUDA.@sync run_with_dagger() GC.gc(true) end ``` which further points at a memory leak.

Actually, it turns out the memory leak only happens when I launch with multiple processes (i.e. Distributed). If I don't everything runs fine. Probably the GC has trouble when the...

I was hoping that binding the spawn commands to the main worker would fix the problem, but it doesn't seem so. Specifically, calling with `julia --project=bench -p4` and then setting...

Try setting ```julia import FastAI.Flux.MLUtils._default_executor _default_executor() = ThreadedEx() ``` See also #142 .

Bump :) Has anyone gotten this to work?

Maybe it's worth revisiting this with package extensions? JSON is one of the few formats I've run into that don't support FileIO yet. EDIT: Seems like e.g. TOML also doesn't...

Seems like there's also #366.