FluxML/Flux.jl v0.16.11

on GitHub

Flux v0.16.11

Diff since v0.16.10

Merged pull requests:

• Bump julia-actions/setup-julia from 2 to 3 (#2673) (@dependabot[bot])
• Update ecosystem.md (#2674) (@msainsburydale)
• Update AGENTS.md extension table (#2678) (@CarloLucibello)
• Support auxiliary outputs in withgradient for Mooncake (#2682) (@CarloLucibello)
• Bump codecov/codecov-action from 6 to 7 (#2683) (@dependabot[bot])
• Update Buildkite pipeline for the new JuliaGPU cluster (#2685) (@maleadt)
• Bump actions/checkout from 6 to 7 (#2687) (@dependabot[bot])
• unbreak enzyme tests (#2688) (@CarloLucibello)
• Bump dorny/paths-filter from 4.0.1 to 4.0.2 (#2689) (@dependabot[bot])
• Update default for FLUX_TEST_ENZYME (#2690) (@wsmoses)
• docs: fix typos and trim trailing whitespace (#2691) (@jeis4wpi)
• unbreak some metal tests (#2692) (@CarloLucibello)
• Fix check_equal_leaves so gradient comparisons actually run (#2693) (@CarloLucibello)
• Add option to disable the caching allocator in train! (#2695) (@CarloLucibello)
• Add a ResNet-18 / Tiny-ImageNet-200 example (#2696) (@CarloLucibello)
• Adaptive GPU memory management in train! (#2697) (@CarloLucibello)
• Add bf16 precision helper (#2698) (@CarloLucibello)
• Overhaul NEWS.md and document changelog upkeep (#2699) (@CarloLucibello)
• Mixed-precision f16/bf16 for BatchNorm/InstanceNorm/GroupNorm (#2700) (@CarloLucibello)
• Delegate normalization layers to NNlib functional operators (#2701) (@CarloLucibello)
• Add Zygote-vs-Reactant training benchmark (perf/reactant) (#2704) (@CarloLucibello)
• cleanup repo (#2705) (@CarloLucibello)
• perf/reactant: benchmark Adam, fix BatchNorm loss-probe artifact (#2706) (@CarloLucibello)
• Compile train! on Reactant devices (#2707) (@CarloLucibello)
• add padding_idx to Embedding (#2708) (@CarloLucibello)
• Add public train_step!, build train! on top of it (#2709) (@CarloLucibello)
• Support auxiliary outputs in Enzyme withgradient (#2710) (@CarloLucibello)
• Support auxiliary loss outputs in train_step! and train! (#2711) (@CarloLucibello)
• Reactant trainstep!: single-forward aux, host-side outputs, cache warning (#2712) (@CarloLucibello)
• Reactant docs guide + trainstep! in the ResNet example & benchmark (#2713) (@CarloLucibello)
• Resolve AD backend once for train!/trainstep!; honour explicit adtype on Reactant (#2714) (@CarloLucibello)

Closed issues:

• Missing functionalities - Metal with Conv and ConvTranspose layers (#2278)
• example for using apple GPU with flux (#2396)
• Missing BFloat16 support (#2573)
• Suppot Mooncake AD (#2640)
• test Enzyme on GPU (#2643)
• add integration test for Flux in Enzyme (#2644)
• activate Enzyme tests on julia 1.12 (#2657)
• Upcoming modification in Mooncake's friendly tangents (#2676)
• Chain gradient issue (#2677)
• add a padding_idx to Embedding (#2684)