Struct IncrementalModelLoader 

pub struct IncrementalModelLoader<S: WeightSource> { /* private fields */ }

Wraps a WeightSource and provides layer-by-layer streaming iteration.

On construction, all tensor names are scanned to extract unique "layers.N." prefixes. The load_all_streaming method then iterates over these prefixes in sorted order, loading one layer at a time and invoking a user-supplied callback. Miscellaneous tensors (not matching the layers.N. pattern) are presented last, grouped under "_misc.".

Implementations

impl<S: WeightSource> IncrementalModelLoader<S>

pub fn new(source: S) -> Self

Construct a new loader from a weight source.

Tensor names are scanned once to build the list of unique layer prefixes.

pub fn load_layer( &mut self, prefix: &str, ) -> ModelResult<HashMap<String, Vec<f32>>>

Load all tensors whose names start with prefix.

The special prefix "_misc." loads all tensors that do not match the "layers.N." pattern.

Returns a HashMap<tensor_name, Vec<f32>> for the group.

pub fn load_all_streaming<F>(&mut self, callback: F) -> ModelResult<()>

where F: FnMut(&str, HashMap<String, Vec<f32>>) -> ModelResult<()>,

Stream through all layers in order, invoking callback once per layer prefix.

The callback receives:

• prefix: the layer prefix string (e.g. "layers.0." or "_misc.")
• tensors: a HashMap<tensor_name, Vec<f32>> for that layer

If callback returns an Err, iteration stops immediately and the error is propagated.

pub fn layer_prefixes(&self) -> &[String]

Return the list of unique layer prefixes discovered in the weight source.

The list is sorted lexicographically; the "_misc." bucket (if present) always appears last.

pub fn source(&self) -> &S

Return a shared reference to the underlying weight source.

pub fn into_source(self) -> S

Consume the loader and return ownership of the underlying weight source.