Anton Oresten

CSV.jl fails to precompile on Ubuntu server, v0.10.5 and up.

1

## Description I am encountering a precompilation error with CSV.jl on an Ubuntu server. This issue starts from version 0.10.5. ## Environment - Julia Version: 1.9.4 - CSV.jl Version: 0.10.5...

`bt` not defined in changeWeights!

1

Hi! I was trying to change the weights of a BallTreeDensity when I encountered this error. Here's the relevant source code: ```julia getIndexOf(bd::BallTreeDensity, i::Int) = getIndexOf(bd.bt, i) function changeWeights!(bd::BallTreeDensity, newWeights::Array{Float64,1})...

`batched_transpose` with multiple batch dimensions

3

### Motivation and description There exists a method for `batched_mul` that reshapes arrays to allow for an arbitrary number of batch dimensions: ```julia function batched_mul(x::AbstractArray{T1,N}, y::AbstractArray{T2,N}) where {T1,T2,N} batch_size =...

This PR adds a new `sample` method for sparse weights, as well as tests. It brings the time complexity from `O(n)` to `O(n_nonzero)`. This would be useful for e.g. [top-p...

Using `gemv` for `batched_vec`

1

`batched_vec` is currently implemented to use `batched_mul` (which calls batched gemm) only with some extra reshapes. Some basic benchmarks (on an RTX PRO 6000 Blackwell) suggest batched gemv is sometimes...

Add `batched_mul` support for oneAPI

2

Like #614, this PR adds partial support for a previously unsupported backend. It was just a matter of hooking up the bindings, using that PR as template. I think `NNlib.batched_mul`...

[CUDNN] Support BFloat16

9

This PR defines methods for making cuDNN work with `BFloat16s.BFloat16`. In the following example, I show how the new methods fixes the `BFloat16` backward pass of `Flux.logitcrossentropy`: Before Note: Core.BFloat16...

Mention built-in `(; a, b) = x` syntax

1

I was not aware of the NamedTuple syntax originally discussed in https://github.com/JuliaLang/julia/issues/28579. I think a quick mention of this would be good, since many users may just be needing this...

Numerical issues with `BFloat16`

4

I've been trying to do training in BFloat16, and I just figured out the primary cause for NaNs in the backward pass, namely that momentum parameters in the optimisers state...

Require LLVM 19 for `llvm_storage` on aarch64

3

Closes #113 `-(::BFloat16)` gives an LLVM error on 1.12 only: 1.11 uses LLVM 16, 1.12 uses 18, and 1.13 uses 20. This PR changes the requirement for LLVM storage on...