Struct RopeMultiBufferPack 

pub struct RopeMultiBufferPack {
    pub params_buf: MlxBuffer,
    pub rope_params_buf: MlxBuffer,
    pub sections_buf: MlxBuffer,
}

Pre-built triple of small parameter buffers for a rope_multi dispatch.

Held in the per-thread [ROPE_PACK_CACHE] so callers that issue repeated dispatches with stable shape (the qwen35 / qwen36 decode hot path: identical head_dim, rope_dim, n_heads, seq_len=1, freq_base, mode, sections every step) skip the per-call allocation triplet (~208 µs/token measured on M5 Max in cfa-20260426-adr015-wave2a-p3aprime). Decode-out-of-scope cases (variable seq_len) populate one entry per seq_len value seen, then reuse on re-encounter.

Fields

params_buf: MlxBuffer

rope_params_buf: MlxBuffer

sections_buf: MlxBuffer