Paulo Valente

[WIP/Proposal] feat: add working first draft

3

This PR aims to add custom symbols as part of the symbol finder This is a step to improve upon Nx development and usage, but done in a more generalized...

The key which is based on mix.lock fails to point to the correct file (probably due to the working_directory change).

We need to add support for complex numbers for the function and for it's grad implementation - [ ] BinaryBackend - [ ] EXLA - [ ] Torchx - [...

We need to add support for complex numbers for the function and for it's grad implementation - [ ] BinaryBackend - [ ] EXLA - [ ] Torchx - [...

Backend-agnostic stateless PRNGs

2

As discussed in #331, we need an approach for having portable PRNGs that can be applied for any backend. > We definitely need the ability to seed RNGs, and I'm...

Currently, the Nx.BinaryBackend does not support wide matrices in SVD

Nx.LinAlg.qr currently only supports tall and square matrices

Allocate intermediate tensors in the same device as the input tensor

1

Some functions in Torchx.Backend create intermediate Iotas and similar tensors. These should be allocated in the same device as the function's input tensor.

Example of failing code: ```elixir iex(1)> t = Nx.tensor([1, 2, 3]) #Nx.Tensor< s64[3] Torchx.Backend(cpu) [1, 2, 3] > iex(2)> Nx.broadcast(t, {2, 3, 2}, axes: [1], names: [:x, :y, :z]) **...

feat: add persistent term to state retrieving benchmark

2

I noticed that there was a benchmark about state retrieval which did not include `persistent_term`s. I added them to the benchmark. However, there is no benchmark for writing speed or...