Carlo Lucibello
Carlo Lucibello
Thanks for the effort, that almost did it: ```julia # NOT WORKING julia> graphplot(A, node_weights = 1:n, line = (3, 0.5, :blues), marker_z = 1:n, names = 1:n ) #...
Is NNAttentionLib.matmul simple enough that can be ported without carrying over another library dependance?
So you say that the attention layer in pytorch is hardly used in practice? https://pytorch.org/docs/stable/generated/torch.nn.MultiheadAttention.html
> I'm not sure what you mean. I'm pretty sure everyone in this thread wants to add the layer to Flux if that's what you're getting at. Sorry I misread...
@JuliaRegistrator register() subdir=lib/NNlibCUDA
@DilumAluthge there is also failing `import NNlibCUDA` with CUDA 2.6, while CUDA 3.0.0 not being released yet. I just wanted to start the registration process to make testing against CUDA#master...
> We wouldn't have to deal with registering it oddly don't see any oddity here, this subpackage thing is actually quite cool and already well supported by julia's toolings >...
@ToucheSir Dhairya and I disagree on a lot of stuff, which is of course unfortunate for two people working on the same project. There is nothing about "personal" and "enmity"...
> I think setting up ci isn't something new. We also shouldn't have nnlib depending on so many heavy deps like cuda. That makes it very heavy for the rest...
So I´d like to give this package-in-a-subdir a try, there is no problem in pulling out at any moment anyway. Maybe the people from KernelAbstractions @vchuravy @jpsamaroo can tell us...