Function model_warm_weights

ds4.c:2331–2356

Optional startup pass that touches tensor pages before timing generation. */

    m->fd = -1;
}

static void model_prefetch_cpu_mapping(const ds4_model *m) {
    if (!m || !m->map || m->size == 0) return;

    /*
     * CPU generation touches expert weights according to router decisions, so a
     * long decode can fault in model pages that the prompt never touched. On
     * current Darwin kernels we have seen those late file-backed faults trigger
     * an OS-level VM panic in map-count accounting. This hint does not copy or
     * pin the GGUF; it just asks the kernel to start bringing the read-only
     * mapping into the page cache before token generation reaches it.
     */
#if defined(POSIX_MADV_WILLNEED)
    const int rc = posix_madvise((void *)m->map, (size_t)m->size, POSIX_MADV_WILLNEED);
    if (rc != 0) {
        ds4_log(stderr,
                DS4_LOG_WARNING,
                "ds4: warning: POSIX_MADV_WILLNEED failed for CPU model mapping: %s\n",
                strerror(rc));
    }
#else
    (void)m;
#endif
}

/* Read the GGUF metadata table.  Values stay in the mmap; we store offsets so
 * later validation can decode only the keys it needs. */
static void parse_metadata(ds4_model *m, ds4_cursor *c) {
    /* n_kv comes from the header. Every entry consumes at least one byte in the