/**
 * ssd.ts – Structured State Space Duality (SSD) kernels for Mamba-2.
 *
 * Implements a chunked SSD algorithm:
 *   A_bar_t = exp(-softplus(A_h) · softplus(dt_t + dt_bias_h))   [scalar per head]
 *   h_t     = A_bar_t · h_{t-1} + B_t · x_t                      [MIMO per head]
 *   y_t     = C_t · h_t
 *
 * The sequence is split into chunks of `chunk_len` time steps.
 * Within each chunk the recurrence is run sequentially; the carry-over
 * state `h` is passed forward between chunks via the state_carry buffer.
 *
 * Dispatch for ssd_chunk_forward:  (num_chunks, H, B)
 * Dispatch for ssd_chunk_backward: (num_chunks, H, B)
 *
 * Buffer layout (all f32, row-major):
 *   x           : [B, L, D_inner]     where D_inner = H * d_head
 *   B_proj      : [B, L, n_groups, N]
 *   C_proj      : [B, L, n_groups, N]
 *   dt          : [B, L, H]
 *   A_log       : [H]                 log(-A), positive scalar per head
 *   dt_bias     : [H]
 *   D_vec       : [H]                 skip connection per head
 *   out         : [B, L, D_inner]     scan output (written by kernel)
 *   state_carry : [num_chunks+1, B, H, N, d_head]  inter-chunk states
 */
export declare const SSD_FORWARD_WGSL: string;
export declare const SSD_BACKWARD_WGSL: string;
//# sourceMappingURL=ssd.d.ts.map