Method opt_epoch

src/llama-context.cpp:2868–2914

}

void llama_context::opt_epoch(
        ggml_opt_dataset_t        dataset,
        ggml_opt_result_t         result_train,
        ggml_opt_result_t         result_eval,
        int64_t                   idata_split,
        ggml_opt_epoch_callback   callback_train,
        ggml_opt_epoch_callback   callback_eval) {
    const uint32_t n_ctx    = this->n_ctx();
    const uint32_t n_batch  = std::min(cparams.n_batch,  n_ctx);
    const uint32_t n_ubatch = std::min(cparams.n_ubatch, n_batch);
    const  int64_t ndata    = ggml_opt_dataset_ndata(dataset);

    GGML_ASSERT(idata_split >= 0);
    GGML_ASSERT(idata_split <= ndata);

    const uint32_t ubatch_per_ctx = n_ctx / n_ubatch;

    struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
    std::vector<llama_token>        tokens(n_ctx);
    std::vector<llama_token> labels_sparse(n_ctx);

    int64_t idata = 0;

    int64_t t_loop_start = ggml_time_us();
    int64_t ndata_in_loop = idata_split*ubatch_per_ctx;
    for (; idata < idata_split; ++idata) {
        constexpr bool train = true;
        const int64_t idata_in_loop = idata*ubatch_per_ctx;

        ggml_opt_dataset_get_batch_host(dataset, tokens.data(), n_ctx*sizeof(llama_token), labels_sparse.data(), idata);
        opt_epoch_iter(dataset, result_train, tokens, labels_sparse, batch,
            callback_train, train, idata_in_loop, ndata_in_loop, t_loop_start);
    }

    t_loop_start = ggml_time_us();
    ndata_in_loop = (ndata - idata_split)*ubatch_per_ctx;
    for (; idata < ndata; ++idata) {
        constexpr bool train = false;
        const int64_t idata_in_loop = (idata - idata_split)*ubatch_per_ctx;

        ggml_opt_dataset_get_batch_host(dataset, tokens.data(), n_ctx*sizeof(llama_token), labels_sparse.data(), idata);
        opt_epoch_iter(dataset, result_eval, tokens, labels_sparse, batch,
            callback_eval, train, idata_in_loop, ndata_in_loop, t_loop_start);
    }

    llama_batch_free(batch);
}

//
// interface implementation