Trait DynQuantizedStorage 

pub trait DynQuantizedStorage:
    Send
    + Sync
    + Debug {
    // Required methods
    fn dtype(&self) -> GgmlDType;
    fn block_size(&self) -> usize;
    fn storage_size_in_bytes(&self) -> usize;
    fn quantize(&mut self, src: &dyn DynBackendStorage) -> Result<(), Error>;
    fn quantize_imatrix(
        &mut self,
        src: &dyn DynBackendStorage,
        imatrix_weights: &[f32],
        n_per_row: usize,
    ) -> Result<(), Error>;
    fn quantize_onto(
        &mut self,
        src: &dyn DynBackendStorage,
    ) -> Result<(), Error>;
    fn quantize_imatrix_onto(
        &mut self,
        src: &dyn DynBackendStorage,
        imatrix_weights: &[f32],
        n_per_row: usize,
    ) -> Result<(), Error>;
    fn dequantize(
        &self,
        elem_count: usize,
    ) -> Result<Box<dyn DynBackendStorage>, Error>;
    fn data(&self) -> Result<Cow<'_, [u8]>, Error>;
    fn fwd(
        &self,
        self_shape: &Shape,
        input: &dyn DynBackendStorage,
        layout: &Layout,
    ) -> Result<(Box<dyn DynBackendStorage>, Shape), Error>;
    fn as_any(&self) -> &(dyn Any + 'static);
    fn device_arc_dyn(&self) -> Arc<dyn DynBackendDevice> ;

    // Provided methods
    fn dequantize_f16(
        &self,
        _elem_count: usize,
    ) -> Result<Box<dyn DynBackendStorage>, Error> { ... }
    fn device_ptr(&self) -> Result<*const u8, Error> { ... }
    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> { ... }
}

Object-safe per-backend quantized storage. Each backend (CPU, CUDA, Metal, …) supplies a concrete type that implements this trait; fuel-core holds them as Box<dyn DynQuantizedStorage> and dispatches without naming backends.

The _src arguments to quantize* are typed as &dyn DynBackendStorage so the implementor can downcast to its own concrete storage; the _onto variants take a HostBuffer-shaped CPU source.

Required Methods

fn dtype(&self) -> GgmlDType

fn block_size(&self) -> usize

fn storage_size_in_bytes(&self) -> usize

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

Quantize an in-device source storage onto self.

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

Quantize with importance-matrix weighting.

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

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<(), Error>

Quantize with importance matrix from a CPU source.

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

Dequantize to backend-native f32 storage.

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

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), Error>

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 + 'static)

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.

Provided Methods

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).

Dyn Compatibility

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors