flashinfer.cake_fused_qk_rope_append

cake_fused_qk_rope_append fuses, in one launch, the per-head RMSNorm of the query and key projections, the NeoX-style rotary embedding of both, and the append of the new key and value rows into a paged BF16 KV cache. It serves prefill and decode (head_dim 128; 8 query heads with 1 KV head, or 64 with 8) on SM90a, SM100a and SM103a GPUs, writes only into caller-owned buffers, and is CUDA Graph capturable. launch_plan() exposes the host-side stage selection.

cake_fused_qk_rmsnorm_rope_append_paged_kv_cache(...)	Fused Q/K RMSNorm + NeoX RoPE + BF16 Q output + paged K/V append in one launch.
launch_plan(num_rows, num_requests, ...[, ...])	Grid and host-derived scalars (mirrors the Cake kernel's launch_plan).