Nikola Janjušević

Float16 CUDA `conv` broken on 5D tensors

8

Float16 CUDA `conv` seems to be broken for 5D tensors, but not 3D or 4D tensors. FluxML/Flux.jl#2184 (using Julia 1.8.3 on a A100 GPU.) ```julia julia> conv(rand(Float16, 16, 16, 1,...

I'm wondering if you think it's possible to write similar code for cuda version for AD of sparse matrix times dense vector. If so, would it be somewhat straightforward from...

`pad_reflect` broken

1

seems the new release (v0.9.18) breaks padding. before: v0.9.17 ```julia julia> using NNlib julia> x = rand(1:9, 3,3,1,1) 3×3×1×1 Array{Int64, 4}: [:, :, 1, 1] = 7 3 2 1...

NCCL Complex wrapper

1

The NCCL backend in distributed utils does not support complex values [see issue](https://github.com/NVIDIA/nccl/issues/539). Can we add a conveinent wrapper in the NCCLEXT to support broadcast, reduce, etc., likely via using...

handle complex datatype

11

Addresses #56. Tests added for ComplexF64. Feedback appreciated.

Complex number wrapper

6

NCCL does not support complex numbers directly and does not plan to ([see issue](https://github.com/NVIDIA/nccl/issues/539)). Are we willing to add a wrapper to NCCL.jl to make using complex numbers more convienient?...