Avik Pal
The curl installation did not work, so I installed via `cargo`. But running juliaup produces: ``` name = "Juliaup" operating_system = "Linux 22.04 (jammy) [64-bit]" crate_version = "1.11.22" explanation =...
New location https://github.com/SciML/FastAlmostBandedMatrices.jl
Currently we throw an error for Banded QR for wide matrices. See discussion in https://github.com/SciML/BoundaryValueDiffEq.jl/pull/122
Currently, it is mostly a copy of `Krylov.GMRES` but with added features to support a fast routine for block diagonal matrices. Some ways to simplify the code: 1. Use givens...
* ~`sigmoid_fast` fails on GPU as pointed out in the NNlib PR. Need to work around that~ * ~Specialize `WrappedFunction` on `Broadcast.BroadcastFunction`~ --> can't do this mutation will violate purity...
Fixes https://github.com/LuxDL/Lux.jl/issues/494 - [x] FluxMPI Parity - [x] Optimisers - [x] CUDA MPI Support - [x] AMDGPU MPI Support - [x] Tutorial (ImageNet Example is completely functional again) - [...
- [x] ~https://github.com/LuxDL/DocumenterVitepress.jl/issues/48~ - [x] ~https://github.com/LuxDL/DocumenterVitepress.jl/issues/69~ - [ ] https://github.com/LuxDL/DocumenterVitepress.jl/issues/119 - [ ] ~Buildkite deployment is not correctly detected inside DocumenterVitepress so we force the directory cleanup with `clean_md_output`~ --...
FastBroadcast is really good on CPUs (on GPUs it doesn't really matter a lot). We should migrate to use that in the case where it can give speedups like `__apply_activation`....
Currently all our tests use Zygote. We should update `test_gradient_correctness_fdm` to compare across all AD packages and FiniteDifferences. - [x] Zygote - [x] Tracker - [x] ReverseDiff - [x] ForwardDiff...
For an immediate solution, see https://github.com/avik-pal/FluxMPI.jl/. With Package Extensions, this doesn't need to be in a separate package. Instead we can have backends which get activated via extensions for distributed...