How to enable parallel with NdArray backend?

[Thor-x86]

Currently, I'm running modern-lstm. However, I can only utilize a single core with ndarray-blas-openblas feature and using CUDA/ROCm are not possible for me because AMD Vega 10 does not support current ROCm version anymore.

Meanwhile, I noticed that there is an option to enable Rayon. I tried to add burn-common crate separately, but it seems like the rayon feature is not working. Any idea how to make it works? Thank you!

[laggui]

using CUDA/ROCm are not possible for me because AMD Vega 10 does not support current ROCm version anymore

Have you tried wgpu? I think Vulkan should also be supported.

But to answer the original question, burn-ndarray has "simd" and "multi-threads" features which are enabled by default. If you are not using default features, you need to enable multi-threads for rayon.

[Thor-x86]

Wow that was a quick response 😁

I tried Vulkan, but it became 4x slower. Also amdgpu_top reported only 10%-66% GPU usage (mostly 38%). Also tried WGPU with WgpuDevice::Cpu but it reported no adapter found.

For burn-ndarray, I enabled simd and still compiling. I will update you soon

[Thor-x86]

Update: It works, even though not near 100% core utilization. Thank you!

[laggui]

Ahh it might be that leveraging the GPU does not bring much gains for this task, which might explain the wgpu / vulkan perf.

For CPU you could also try the tch backend.