Function perplexity_v2

tools/perplexity/perplexity.cpp:296–442

}

static results_perplexity perplexity_v2(llama_context * ctx, const common_params & params) {
    // Download: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
    // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
    // Output: `perplexity: 13.5106 [114/114]`
    // BOS tokens will be added for each chunk before eval

    const llama_model * model = llama_get_model(ctx);
    const llama_vocab * vocab = llama_model_get_vocab(model);

    const bool add_bos = llama_vocab_get_add_bos(vocab);
    GGML_ASSERT(!llama_vocab_get_add_eos(vocab));

    LOG_INF("%s: tokenizing the input ..\n", __func__);

    std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, true);

    const int n_ctx = llama_n_ctx(ctx);

    if (int(tokens.size()) < 2*n_ctx) {
        LOG_ERR("%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*n_ctx,
                n_ctx);
        LOG_ERR("%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
        return {std::move(tokens), 0., {}, {}};
    }

    std::vector<float> logit_history;
    std::vector<float> prob_history;

    logit_history.resize(tokens.size());
    prob_history.resize(tokens.size());

    if (params.ppl_stride <= 0) {
        LOG_ERR("%s: stride is %d but must be greater than zero!\n",__func__,params.ppl_stride);
        return {tokens, -1, logit_history, prob_history};
    }

    const int calc_chunk = n_ctx;

    LOG_INF("%s: have %zu tokens. Calculation chunk = %d\n", __func__, tokens.size(), calc_chunk);

    if (int(tokens.size()) <= calc_chunk) {
        LOG_ERR("%s: there are only %zu tokens, this is not enough for a context size of %d and stride %d\n",__func__,
                tokens.size(), n_ctx, params.ppl_stride);
        return {tokens, -1, logit_history, prob_history};
    }

    const int n_chunk_max = (tokens.size() - calc_chunk + params.ppl_stride - 1)  / params.ppl_stride;

    const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
    const int n_batch = params.n_batch;

    const int n_vocab = llama_vocab_n_tokens(vocab);

    int count = 0;
    double nll = 0.0;

    const int n_seq = std::max(1, n_batch / n_ctx);
    LOG_INF("%s: computing over %d chunks, n_ctx=%d, batch_size=%d, n_seq=%d\n", __func__, n_chunk, n_ctx, n_batch, n_seq);