API Reference
Luximm.jl exposes three groups of public symbols, all re-exported from the top-level Luximm module.
| Group | Purpose | Page |
|---|---|---|
| Models | The family-agnostic create_model / create_pretrained pair, variant tables, and per-variant config structs. | Models |
| Layers | Building blocks shared across families: weight-standardized conv, channel-axis LayerNorm, Global Response Norm, and timm-equivalent initializers. | Layers |
| Interop | PyTorch and HuggingFace plumbing: HDF5 parity fixtures, state_dict application, axis transforms, HuggingFace Hub cache, SafeTensors loading. | Interop |
The top-level Luximm module re-exports every public symbol from these three groups, so end users normally just call using Luximm. The submodule-qualified names (Luximm.Interop.read_parity, Luximm.Layers.std_conv) remain available for cases where the shorter top-level name would collide with something else in the caller's namespace.
Module map
Luximm
├── Luximm.Models # family constructors, loaders, variant tables
├── Luximm.Layers # std_conv, layernorm2d, grn_layer, init recipes
└── Luximm.Interop # parity, HF Hub, SafeTensorsInternal block constructors (resnet_basic_block, convnextv2_block, the stage scaffolding under ConvNeXtCommon, and the per-family resnet / bit_resnetv2 / convnext / convnextv2 constructors themselves) are deliberately unexported. They are implementation details of create_model and create_pretrained; rely on those two as the public construction surface.