Bytes a variant is expected to use once loaded on a device: the measured memory when the manifest
has one, otherwise an estimate from the download size. 0 when neither is known.
The estimate reflects how ONNX Runtime holds weights: WebAssembly expands fp16 to fp32 (about 4× the
file), WebGPU keeps a copy in GPU buffers, the native CPU runtime maps them nearly as stored.
Bytes a variant is expected to use once loaded on a device: the measured
memorywhen the manifest has one, otherwise an estimate from the download size.0when neither is known.The estimate reflects how ONNX Runtime holds weights: WebAssembly expands fp16 to fp32 (about 4× the file), WebGPU keeps a copy in GPU buffers, the native CPU runtime maps them nearly as stored.