/**
 * attention.ts – Causal multi-head self-attention kernels.
 *
 * Implements tiled 16×16 causal attention suitable for WebGPU.
 * No Flash-Attention dependency — straightforward O(L²) with causal mask.
 *
 * Buffer layout:
 *   qkv_in  : [B, L, 3*D_model]   fused Q,K,V after wQKV projection
 *   out_buf : [B, L, D_model]
 *   scores  : [B, H, L, L]        intermediate (written then read by kernel)
 *
 * Dispatch attention_forward:  (ceil(L/16), H, B)
 * Dispatch softmax_forward:    (L, H, B)            — one workgroup per row
 * Dispatch attention_backward: (ceil(L/16), H, B)
 */
export declare const SOFTMAX_WGSL: string;
export declare const ATTENTION_FORWARD_WGSL: string;
export declare const ATTENTION_BACKWARD_WGSL: string;
//# sourceMappingURL=attention.d.ts.map