Expand description
Weight-loading abstraction.
Native targets typically mmap a .safetensors file and read tensors
via byte-offset slices into the mapping. WASM has no mmap; weights
arrive as Vec<u8> from fetch() or Response.arrayBuffer(). Both
paths produce the same shape: a name → byte-slice lookup.
WeightLoader is the contract. Concrete implementations live in
in-tree model loaders (mmap-based) and here (BytesWeightLoader — works on
every target including WASM).
Structs§
- Bytes
Weight Loader - Owned in-memory weight loader. The simplest, most portable variant — works on every target including WASM.
Traits§
- Weight
Loader - A name-keyed view of weight tensor bytes.