Struct KdaDeviceState 

pub struct KdaDeviceState {
    pub conv: DevicePtr,
    pub recurrent: DevicePtr,
    /* private fields */
}

2026-09-25: One sequence’s KDA conv and recurrent state on the device. launch_k3_kda_decode_token_on_device updates it in place and copies only the output back; download copies it to a host KdaState.

launch_k3_kda_decode_token instead uploads and downloads the state every token; at KdaConfig::production the recurrent state alone is 96 x 128 x 128 FP32, 6.3 MB per layer.

Fields

conv: DevicePtr

recurrent: DevicePtr

Implementations

impl KdaDeviceState

pub fn alloc_and_upload(gpu: &dyn GpuBackend, host: &KdaState) -> Result<Self>

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

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