Function bark_eval_fine_encoder_internal

bark.cpp:1963–2008

}

static bool bark_eval_fine_encoder_internal(
    struct bark_context* bctx,
    bark_sequence& input,
    std::vector<float>& logits,
    int nn,
    int n_threads) {
    auto& model = bctx->text_model.fine_model;
    auto& allocr = bctx->allocr;
    auto& hparams = model.hparams;
    auto& params = bctx->params;

    const int n_vocab = hparams.n_out_vocab;
    const int block_size = hparams.block_size;

    const int n_fine_codebooks = params.n_fine_codebooks;

    const int64_t t_predict_us_start = ggml_time_us();

    // reset the allocator to free all the memory allocated during the previous inference
    ggml_allocr_reset(allocr);

    struct ggml_cgraph* gf = bark_build_fine_gpt_graph(
        &model, allocr, input, nn, n_fine_codebooks, n_threads);

    // allocate tensors
    ggml_allocr_alloc_graph(allocr, gf);

    // run the computation
    if (ggml_backend_is_cpu(model.backend)) {
        ggml_backend_cpu_set_n_threads(model.backend, n_threads);
    }
#ifdef GGML_USE_METAL
    if (ggml_backend_is_metal(model.backend)) {
        ggml_backend_metal_set_n_cb(model.backend, n_threads);
    }
#endif
    ggml_backend_graph_compute(model.backend, gf);

    struct ggml_tensor* inpL = gf->nodes[gf->n_nodes - 1];

    ggml_backend_tensor_get(inpL, logits.data(), 0, sizeof(float) * n_vocab * block_size);

    model.t_predict_us += ggml_time_us() - t_predict_us_start;

    return true;
}

static bool bark_eval_fine_encoder(struct bark_context* bctx, int n_threads) {
    // input shape: [N, n_codes]