Function test_top_p

tests/test-sampling.cpp:43–62

}

static void test_top_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
    size_t n_vocab = probs.size();
    std::vector<llama_token_data> candidates;
    candidates.reserve(n_vocab);
    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
        float logit = log(probs[token_id]);
        candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
    }

    llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
    llama_sample_softmax(nullptr, &candidates_p);
    DUMP(&candidates_p);
    llama_sample_top_p(nullptr, &candidates_p, p, 1);
    DUMP(&candidates_p);

    GGML_ASSERT(candidates_p.size == expected_probs.size());
    for (size_t i = 0; i < candidates_p.size; i++) {
        GGML_ASSERT(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3);
    }
}

static void test_tfs(const std::vector<float> & probs, const std::vector<float> & expected_probs, float z) {
    size_t n_vocab = probs.size();