Function ds4_session_eval_layer_slice

ds4.c:26314–26660

    free(gpu_routed_mid_all);
    free(gpu_shared);
    free(gpu_shared_mid);
    free(gpu_shared_up);
    free(gpu_shared_gate);
    free(gpu_ffn_norm);
    free(gpu_ffn_cur);
    free(gpu_after_attn_hc);
    free(gpu_attn_out);
    free(gpu_kv);
    free(gpu_q);
    free(gpu_attn_norm);
    free(gpu_attn_cur);
    free(routed_q8_midscale);
    free(routed_q8_midq);
    free(routed_q8_xscale);
    free(routed_q8_xq);
    free(routed_midq);
    free(routed_xq);
    free(routed_mid_all);
    free(shared_xscale);
    free(shared_xq);
    free(cpu_after_ffn_hc);
    free(cpu_ffn_out);
    free(cpu_routed);
    free(cpu_shared);
    free(cpu_shared_mid);
    free(cpu_shared_up);
    free(cpu_shared_gate);
    free(cpu_ffn_norm);
    free(cpu_ffn_cur);
    free(cpu_after_attn_hc);
    free(cpu_attn_out);
    free(cpu_heads);
    free(cpu_kv);
    free(cpu_qr_norm);
    free(cpu_q);
    free(cpu_attn_norm);
    free(cpu_attn_cur);
}

static int metal_graph_decode_test(
        const ds4_model   *model,
        const ds4_weights *weights,
        const token_vec   *prompt,
        bool               quality) {
    if (prompt->len <= 0) {
        fprintf(stderr, "ds4: Metal graph test needs a non-empty prompt\n");
        return 1;
    }

    const int token = prompt->v[0];
    const ds4_layer_weights *layer = &weights->layer[0];
    const uint64_t hc_dim = (uint64_t)DS4_N_HC * DS4_N_EMBD;
    const uint64_t q_rank = layer->attn_q_a->dim[1];
    const uint64_t q_dim = (uint64_t)DS4_N_HEAD * DS4_N_HEAD_DIM;
    const uint64_t expert_in_dim = layer->ffn_gate_exps->dim[0];
    const uint64_t down_in_dim = layer->ffn_down_exps->dim[0];
    const uint64_t vocab_dim = weights->output->dim[1];
    const bool routed_q8_0 =