Repository navigation
GPUArrays rand seeding fails when container is bigger than workgroup size #643
Description
See linked comment below for context, but in short I'm getting different results than CI on the rand doctest that uses GPUArrays.default_rng(MtlArray)
I was just getting different results locally and figured if it's unstable between devices maybe better to filter it out but maybe we should look into that instead (in a future pr). Lmk if I should revert the regex filter.
Originally posted by @christiangnrd in JuliaGPU/Metal.jl#641 (comment)
Activity
[-]Issues(?) with GPUArrays `rand` seeding[/-][+]GPUArrays `rand` seeding fails when container is bigger than workgroup size[/+]christiangnrd commentedon Oct 30, 2025
maleadt commentedon Oct 19, 2026
Fixed by #707, which replaced the stateful Tausworthe generator with a counter-based Philox RNG: values only depend on the seed, counter and element index, so they no longer change with the launch configuration or the back-end.
rand(GPUArrays.RNG{CuArray}(42), Float32, n)now matches the JLArray output bit-for-bit for everyn.