flashinfer.cake_fused_qk_rope_fp8_append.launch_plan

flashinfer.cake_fused_qk_rope_fp8_append.launch_plan(num_rows: int, num_requests: int, num_kv_heads: int) → dict[str, Any]

Grid of the fused kernel’s default build (mirrors the Cake kernel’s launch_plan).

One CTA per (packed row, kv head) plus one CTA per (request, kv head) for the last-page tail clear and split_k_flag. A stage that folds the clearing work into the row CTAs has its own grid rule: cake_jit.launch_grid() of its module record.