Flux.jl

08ed6809 - [tests] Expand Reactant test coverage: layers, losses, optimisers, precision (#2750)

[tests] Expand Reactant test coverage: layers, losses, optimisers, precision (#2750) * [tests] Expand Reactant test coverage Add tests on the Reactant device for: - gradient parity with Zygote for layers not in TEST_MODELS: conv options, pooling, upsampling, normalisation (also in test mode), recurrent layers over sequences and cells, Attention, activations, containers, Dropout and AlphaDropout in test mode; - gradient parity for every loss, w.r.t. the model and input only; - trainstep! trajectory parity with eager for the optimiser rules; - f16/bf16 training and Float64 kept on the device. Compile the Reactant gradient helpers at HIGHEST dot_general/convolution precision, so that CUDA's TF32-like default doesn't eat into the tolerances. Known upstream bugs are marked broken: FluxML/Optimisers.jl#234, #235 (CUDA only), #236 and #237. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * [tests] Merge and trim the Reactant tests; Optimisers 0.4.10 fixes Optimisers 0.4.10 fixes RAdam, AMSGrad/AdaGrad setup on GPU, Rprop and AccumGrad on Reactant, so their trajectory tests are no longer broken. Merge the ten new files into four (layers, conv, losses, optimisers) and chain several layers per gradient case, roughly halving the number of compilations: each file and case pays Julia and XLA compilation, which dominates their time on CI. Compile the trajectory steps at full precision: on CUDA, TF32 rounding flips signs that Rprop amplifies into a different trajectory. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Link Reactant.jl#3504 from the onehot Embedding test Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>