Struct MlaDeviceKv 

pub struct MlaDeviceKv {
    pub k: DevicePtr,
    pub v: DevicePtr,
    pub seq_len: usize,
    pub cap: usize,
    /* private fields */
}

Device-resident MLA KV. Append is one-row D2D/H2D; SDPA reads the buffer. Host MlaKv remains the CPU oracle. Do not re-upload [0..T] each token.

Fields

k: DevicePtr

v: DevicePtr

seq_len: usize

cap: usize

Implementations

impl MlaDeviceKv

pub fn k_row(&self) -> usize

pub fn v_row(&self) -> usize

pub fn alloc( gpu: &dyn GpuBackend, cap: usize, k_row: usize, v_row: usize, ) -> Result<Self>

pub fn alloc_and_upload( gpu: &dyn GpuBackend, host: &MlaKv, cap: usize, k_row: usize, v_row: usize, ) -> Result<Self>

pub fn validate_cfg(&self, cfg: &MlaConfig) -> Result<()>

pub fn download(&self, gpu: &dyn GpuBackend, host: &mut MlaKv) -> Result<()>

pub fn free(self, gpu: &dyn GpuBackend) -> Result<()>