Struct CpuQStorage 

pub struct CpuQStorage(pub Box<dyn QuantizedType>);

Tuple Fields

0: Box<dyn QuantizedType>

Trait Implementations

impl Debug for CpuQStorage

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl DynQuantizedStorage for CpuQStorage

fn dtype(&self) -> GgmlDType

fn block_size(&self) -> usize

fn storage_size_in_bytes(&self) -> usize

fn quantize(&mut self, src: &dyn DynBackendStorage) -> Result<()>

Quantize an in-device source storage onto self.

fn quantize_imatrix( &mut self, src: &dyn DynBackendStorage, imatrix_weights: &[f32], n_per_row: usize, ) -> Result<()>

Quantize with importance-matrix weighting.

fn quantize_onto(&mut self, src: &dyn DynBackendStorage) -> Result<()>

Quantize a CPU source onto self (cross-device).

fn quantize_imatrix_onto( &mut self, src: &dyn DynBackendStorage, imatrix_weights: &[f32], n_per_row: usize, ) -> Result<()>

Quantize with importance matrix from a CPU source.

fn dequantize(&self, elem_count: usize) -> Result<Box<dyn DynBackendStorage>>

Dequantize to backend-native f32 storage.

fn data(&self) -> Result<Cow<'_, [u8]>>

Raw bytes of the quantized data (host-readable copy).

fn fwd( &self, self_shape: &Shape, input: &dyn DynBackendStorage, layout: &Layout, ) -> Result<(Box<dyn DynBackendStorage>, Shape)>

QMatMul forward against an in-device input storage. Returns a fresh device storage and its output shape. self_shape is the weight shape (the QTensor’s own shape).

fn as_any(&self) -> &dyn Any

Identity returns the host device tag this storage is on.

fn device_arc_dyn(&self) -> Arc<dyn DynBackendDevice>

Owning device handle. Lets fuel-core recover a Device from a QTensor without keeping a parallel device field on QTensor.

fn dequantize_f16( &self, _elem_count: usize, ) -> Result<Box<dyn DynBackendStorage>, Error>

Dequantize to f16 (CUDA fast path; default impl rejects).

fn device_ptr(&self) -> Result<*const u8, Error>

Device pointer for callers that need raw addressing (CUDA only today).

fn indexed_moe_forward( &self, _self_shape: &Shape, _input: &dyn DynBackendStorage, _input_layout: &Layout, _ids: &dyn DynBackendStorage, _ids_layout: &Layout, ) -> Result<(Box<dyn DynBackendStorage>, Shape), Error>

Indexed MoE forward (CUDA-only today; default rejects).