Evaluating sparse matrices in the REPL has a huge memory footprint

[ma-sadeghi]

Here's how to reproduce:

1. Create a large sparse matrix
2. Evaluate the matrix in the REPL
3. Monitor GPU usage

In my case, evaluating a ~ 12GB matrix in the REPL caused a ~ 28GB spike (12 + 28) in GPU memory usage. (Of course, the memory is released once the evaluation is done...)

[maleadt]

We just convert the sparse GPU array to a sparse CPU array and use SparseArray.jl's output methods; there's nothing special about evaluation in the REPL. So if there's a surprising memory use, that should be reproducible by just calling these conversion methods.

Anyway, I can't reproduce:

julia> using CUDA, SparseArrays

julia> x = cu(sprand(10000, 10000, 0.01));

julia> CUDA.memory_status()
Effective GPU memory usage: 0.94% (459.625 MiB/47.504 GiB)
Memory pool usage: 7.660 MiB (32.000 MiB reserved)

julia> x
10000×10000 CuSparseMatrixCSC{Float32, Int32} with 998972 stored entries:
...

julia> CUDA.memory_status()
Effective GPU memory usage: 0.94% (459.625 MiB/47.504 GiB)
Memory pool usage: 7.660 MiB (32.000 MiB reserved)

Please, when filing bugs, always include an actual reproducer, as suggested by the bug filing template. That template also asks about crucial information, like the CUDA.jl version, Julia version, etc.

[maleadt]

Going to close this for lack of response, but feel free to open a new issue with more details if you're still running into this on the latest version.