NNlib.jl icon indicating copy to clipboard operation
NNlib.jl copied to clipboard
verified publisher icon FluxML

add Metal extension for batched_mul

Open mcabbott opened this issue 1 year ago • 3 comments

Closes #581

PR Checklist

  • [x] Tests are added
  • [ ] Documentation, if applicable

mcabbott avatar Nov 15 '24 03:11 mcabbott

https://github.com/JuliaGPU/Metal.jl/issues/381

chengchingwen avatar Dec 01 '24 03:12 chengchingwen

Thanks I hadn't seen that.

Got a wrong answer in this test on CI (tiny arrays though) but didn't investigate further:

https://github.com/FluxML/NNlib.jl/pull/614/files#diff-df0d2a37225f09d22727651479dc1cd59f2b8358f4eb1e2be98c9b04e215be86R31-R34

mcabbott avatar Dec 01 '24 03:12 mcabbott

IIRC the bug might also happen on tiny arrays if it's within a sequence of calls. It's really hard to detect though.

chengchingwen avatar Dec 01 '24 04:12 chengchingwen