Function model_map_span_vec_include_layer_decode

ds4.c:4275–4286

* Decode-time spans for one layer. The static set excludes routed expert * tensors because the streaming expert cache serves them — but a boosted * layer (per-expert bytes != the global slab class) can never be cached, and * both the selected-addr prefill fallback and the decode no-cache fallback * read its experts through ds4_gpu_wrap_model_range / wrap_model_exact_range * over the model map

    if (n < 0 || (size_t)n >= sizeof(name)) ds4_die("tensor name is too long");
    return required_tensor(m, name);
}

static ds4_tensor *tensor_by_mtp_stage_suffix(
        const ds4_model *m,
        uint32_t         stage,
        const char      *suffix) {
    char name[160];
    int n = snprintf(name, sizeof(name), "mtp.%u.%s", stage, suffix);
    if (n < 0 || (size_t)n >= sizeof(name)) ds4_die("tensor name is too long");
    return model_find_tensor(m, name);
}

static void tensor_expect_layout(
        const ds4_tensor *t,
        uint32_t          type,