Struct GgufTensorIndex 

pub struct GgufTensorIndex {
    pub tensor_data_start: u64,
    pub hyperparams: GgufHyperparams,
    pub max_tensor_bytes: usize,
    pub max_layer_tensor_bytes: usize,
    /* private fields */
}

Lookup table from tensor-name hash → GgufTensorInfo, built by walking the GGUF tensor-info section that immediately follows the KV metadata section.

Fields

tensor_data_start: u64

Absolute byte offset in the mmap where tensor payload data begins.

hyperparams: GgufHyperparams

max_tensor_bytes: usize

Largest tensor payload in the file (informational).

max_layer_tensor_bytes: usize

Largest layer matmul tensor (attn/ffn weights) — sizes reusable GPU staging.

Implementations

impl GgufTensorIndex

pub fn from_gguf(mmap: &[u8]) -> Self

pub fn from_components( named_tensors: &[(&[u8], GgufTensorInfo)], hyperparams: GgufHyperparams, tensor_data_start: u64, ) -> Self

Build a synthetic index from an explicit (name, info) list — used to boot from a P64 weight container so the entire GGUF-based hot path (get_layer_tensors / fetch_tensor_bytes / resident upload) works unchanged. The caller passes absolute blob offsets in each GgufTensorInfo.byte_offset and tensor_data_start = 0, pointing the byte source at the P64 bytes. Format-agnostic: the hot path never learns it is reading P64.

pub fn get_layer_tensors(&self, layer_idx: u32) -> LayerTensors

Retrieve attention + FFN tensor metadata for one transformer block.

pub fn output_norm_info(&self) -> Option<&GgufTensorInfo>

pub fn output_weight_info(&self) -> Option<&GgufTensorInfo>

pub fn output_weights_tied(&self) -> bool

True when output.weight is absent and logits use tied token_embd.weight.

pub fn logits_projection_info(&self) -> Option<&GgufTensorInfo>

Output projection weights: output.weight when present, else tied token_embd.weight.

pub fn weight_tie_probe(&self) -> (bool, u64, u64, [u64; 4], [u64; 4])

Diagnostic: byte offset + dims for tied-weight validation (MC3g).

pub fn token_embd_info(&self) -> Option<&GgufTensorInfo>

Cached token_embd.weight tensor metadata.

pub fn emb_dim(&self) -> usize

Return the embedding dimension (n_embd) from token_embd.weight, or 0 if unknown.

pub fn vocab_dim(&self) -> usize

Vocabulary size from token_embd.weight shape [n_embd, n_vocab].

pub fn dequantize_token_embedding_into( &self, mmap: &[u8], token_id: u32, out: &mut [f32], ) -> usize

Dequantize one token embedding into caller-supplied out (zero heap in hot path).

pub fn get_token_embedding(&self, mmap: &[u8], token_id: u32) -> Vec<f32>

Slice and dequantize a token embedding (test / legacy path; allocates Vec).

Trait Implementations

impl Clone for GgufTensorIndex

fn clone(&self) -> GgufTensorIndex

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.