Function tensor_expert_bytes

ds4.c:5529–5549

Locate one expert's 2D matrix inside a 3D GGUF expert tensor. */

            fprintf(stderr, "ds4: n-gram embeddings must use original BF16, got type %u\n", t);
            exit(1);
        }
        if (w->ple_embd->ndim != 2 || w->ple_embd->dim[0] != DS4_N_PLE_HEAD_DIM ||
            w->ple_embd->dim[1] < g_ds4_qwen4_ple.n_rows) {
            fprintf(stderr, "ds4: per_layer_token_embd layout [%" PRIu64 ", %" PRIu64 "] does not "
                    "cover %u x %" PRIu64 " hash rows\n",
                    w->ple_embd->dim[0], w->ple_embd->dim[1], DS4_N_PLE_HEAD_DIM,
                    g_ds4_qwen4_ple.n_rows);
            exit(1);
        }
    }

    const bool have_output = weights_have_output_head(w);
    if (require_output && !have_output) ds4_die("required output head tensors are missing");
    if (weights_have_partial_output_head(w) && !have_output) ds4_die("partial output head in GGUF");
    if (have_output) {
        tensor_expect_layout(w->output_hc_norm, DS4_TENSOR_F32, 1, hc_dim, 0, 0);
        tensor_expect_qwen4_dense_layout(w->output_hc_down, 2, hc_dim, DS4_N_HC_LOWRANK, 0);
        tensor_expect_qwen4_dense_layout(w->output_hc_up, 2, DS4_N_HC_LOWRANK, hc_dim, 0);
        tensor_expect_qwen4_dense_layout(w->output, 2, DS4_N_EMBD, DS4_N_VOCAB, 0);
    }

    for (uint32_t il = layer_start; il <= layer_end; il++) {
        const ds4_layer_weights *l = &w->layer[il];
        if (!weights_qwen4_layer_has_required(l, il)) {