pub(crate) fn load_constant_tensors(dir: &Path) -> Result<Vec<ConstantTensor>>Expand description
Read <tensor-name>.bin for every tensor the caller supplied.
The naming follows --calibration-data, which is the convention this
project already has for feeding a directory of external data in. Only the
file stem is used, so weight.bin supplies the tensor a WGSL kernel binds
as weight.
Nothing is invented here. A tensor with no file stays a graph input, and the backend says what that costs; filling one with zeros to win an acceleration would produce a model that runs, is attributed to the accelerator, and computes the wrong thing.