Struct DirectCheckpoint 

pub struct DirectCheckpoint { /* private fields */ }

A validated Hugging Face checkpoint prepared for direct runtime loading.

The object owns immutable mappings for every SafeTensors shard. Its GGUF buffer contains metadata and tensor descriptors only; tensor bytes are read from the source mappings on demand.

Implementations

impl DirectCheckpoint

pub fn open(source: &Path, buffer_size: usize) -> Result<Self>

Open a checkpoint directory or SafeTensors file and prepare canonical llama.cpp metadata and names.

pub fn metadata_gguf(&self) -> &[u8]

Metadata-only GGUF consumed by the native model constructor.

pub fn tensor_count(&self) -> usize

Number of canonical tensors exposed by this checkpoint.

pub fn imatrix_layout(&self) -> Result<Vec<ImatrixTensorLayout>>

Canonical tensor names and importance-vector widths expected by llama.cpp.

pub fn read_tensor_f32(&self, name: &str, destination: &mut [f32]) -> Result<()>

Decode one canonical tensor into the caller-provided F32 destination.