FluxML/Flux.jl v0.16.15 on GitHub

Flux v0.16.15

Merged pull requests:

GPT-2 example: add a CUDA.jl version (#2765) (@CarloLucibello)
Mooncake: move the rule for mse here from Mooncake's Flux extension (#2772) (@AstitvaAggarwal)
Fix the gradient of logitcrossentropy for a 0-dimensional label (#2788) (@parth-patel010)
Fix Enzyme zero gradient for a loss returning a NamedTuple with aux arrays (#2790) (@CarloLucibello)
Add a precompile workload (#2792) (@CarloLucibello)
Count a change of exactly min_dist as no improvement in early_stopping and plateau (#2793) (@parth-patel010)
Make g of WeightNorm with dims = : an array, so that it is trained (#2797) (@parth-patel010)
One GC pacer per model for train! and trainstep!, with a full GC after a step that compiles (#2798) (@CarloLucibello)

Closed issues:

No precompile workload: first forward + gradient + update spends ~24 s in inference (#2781)
Gradient of logitcrossentropy fails for a single sample with a 0-dimensional label (#2787)
early_stopping never triggers when the loss stays constant (#2791)
train!'s paced GC may never free the first step's GPU arrays (#2795)
WeightNorm with dims = : makes g a scalar that is never trained (#2796)