Function payload_write_tensor_span

ds4.c:23451–23471

Accelerator tensors are copied through a fixed-size CPU buffer. We do not mmap the * cache file and we do not allocate a second graph-sized blob just to serialize * it; both would be poor fits for this very large model. */

                                                   false,
                                                   freq_base,
                                                   freq_scale,
                                                   ext_factor,
                                                   attn_factor,
                                                   DS4_ROPE_YARN_BETA_FAST,
                                                   DS4_ROPE_YARN_BETA_SLOW,
                                                   DS4_RMS_EPS) != 0;
    }
    if (!decode_q_norm_rope_fused) {
        if (ok) ok = ds4_gpu_head_rms_norm_tensor(metal_graph_q(g), 1, tp_heads, DS4_N_HEAD_DIM, DS4_RMS_EPS) != 0;
        if (ok) {
            metal_graph_debug_dump_tensor("Qnorm", metal_graph_q(g), q_dim, il, pos);
        }
        if (ok) ok = ds4_gpu_rope_tail_tensor(metal_graph_q(g), 1, tp_heads, DS4_N_HEAD_DIM,
                                                DS4_N_ROT, pos,
                                                compressed ? (uint32_t)DS4_ROPE_ORIG_CTX : 0,
                                                false, freq_base, freq_scale, ext_factor, attn_factor,
                                                DS4_ROPE_YARN_BETA_FAST, DS4_ROPE_YARN_BETA_SLOW) != 0;
    }
    DS4_METAL_PROFILE_DECODE_STAGE("q_path");
    if (ok) {
        metal_graph_debug_dump_tensor("Qcur", metal_graph_q(g), q_dim, il, pos);
    }
    if (!qkv_rms_fused) {
        if (ok) ok = metal_graph_matmul_dense_quant_tensor(metal_graph_kv_raw(g),