Struct KdaSeqState 

pub struct KdaSeqState {
    pub conv: DevicePtr,
    pub recurrent: DevicePtr,
}

The per-sequence state a KDA layer carries. Both buffers are read-modify-write.

🪤 The recurrent element is FP32 by reference semantics — HF stores it via last_recurrent_state.to(torch.float32) and vLLM’s kda_state_dtype returns (conv_dtype, torch.float32) regardless of mamba_cache_dtype. --ssm-h-dtype f16 is not available to KDA without deviating from the reference.

🪤 The conv buffer is Atlas’s conv_kernel-wide convention, one slot wider than HF’s conv_kernel - 1: Atlas shifts left before convolving, so slot 0 is shifted out and never participates. HF[0..k-1] == Atlas[1..k] pre-shift. Any code moving state between the two conventions must apply that offset.

Fields

conv: DevicePtr

[conv_dim, conv_kernel] FP32.

recurrent: DevicePtr

[heads, head_dim, head_dim] FP32, K-major.

Trait Implementations

impl Clone for KdaSeqState

fn clone(&self) -> KdaSeqState

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for KdaSeqState

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

Formats the value using the given formatter.

impl Copy for KdaSeqState