Struct GgufWeights 

pub struct GgufWeights { /* private fields */ }

All F32-decoded tensors of one GGUF file, keyed by tensor name.

Implementations

impl GgufWeights

pub fn meta_u32(&self, key: &str) -> Option<u32>

U32 metadata value (accepts u32/u64/i32/i64 spellings).

pub fn load_2d(&self, name: &str) -> Result<Tensor<2>>

Load a 2D [rows, cols] F32 tensor.

pub fn load_weight_2d( &self, name: &str, expected_d_out: usize, ) -> Result<Tensor<2>>

Load a 2D weight with the load_weight transpose guard: PyTorch stores (d_out, d_in); if the file holds it transposed (the Q8_0 transpose trick some converters apply), flip it back.

pub fn load_1d(&self, name: &str) -> Result<Tensor<1>>

Load a 1D [n] F32 tensor. Like the retired candle bridge (burn_weight_1d), this flattens any stored shape: real checkpoints sometimes store row vectors as [1, n].

pub fn load_3d(&self, name: &str) -> Result<Tensor<3>>

Load a 3D [d0, d1, d2] F32 tensor.

pub fn load_flat(&self, name: &str) -> Result<Vec<f32>>

Raw F32 values of any tensor (for embedding tables, rope vectors, …).