voice_audio_multi/tests/higgs_audio_tts/test_sampler.cpp

#include "engine/models/higgs_audio_tts/sampler.h"

#include <cstdint>
#include <iostream>
#include <limits>
#include <stdexcept>
#include <string>
#include <vector>

namespace {

void test_greedy_filter_is_opt_in() {
    using engine::models::higgs_audio_tts::HiggsCodebookSampler;
    using engine::models::higgs_audio_tts::HiggsSamplingOptions;

    const float nan = std::numeric_limits<float>::quiet_NaN();
    const float inf = std::numeric_limits<float>::infinity();
    const std::vector<std::vector<float>> rows = {
        {nan, -inf, 1.0F, 2.0F},
        {0.0F, inf, 1.0F, 2.0F},
        {-3.0F, -1.0F, -1.0F, -2.0F},
    };
    const std::vector<int32_t> legacy_expected = {0, 1, 1};
    const std::vector<int32_t> filtered_expected = {3, 3, 1};
    for (const bool temperature_greedy : {false, true}) {
        for (const bool filter : {false, true}) {
            HiggsCodebookSampler sampler(1, 4);
            auto state = sampler.make_state();
            HiggsSamplingOptions options;
            options.temperature = temperature_greedy ? 0.0F : 1.0F;
            options.top_k = temperature_greedy ? 4 : 1;
            if (filter) options.filter_nan_logits = true;
            for (size_t i = 0; i < rows.size(); ++i) {
                const auto & codes = sampler.step(rows[i].data(), rows[i].size(), state, options);
                const auto expected = filter ? filtered_expected[i] : legacy_expected[i];
                if (codes.size() != 1 || codes.front() != expected) {
                    throw std::runtime_error("Higgs greedy filter changed default behavior or finite argmax");
                }
            }
        }
    }
}

void test_greedy_filter_reports_when_all_logits_are_non_finite() {
    using engine::models::higgs_audio_tts::HiggsCodebookSampler;
    using engine::models::higgs_audio_tts::HiggsSamplingOptions;

    const float nan = std::numeric_limits<float>::quiet_NaN();
    const float inf = std::numeric_limits<float>::infinity();
    const std::vector<float> logits = {nan, inf, -inf};
    for (const bool temperature_greedy : {false, true}) {
        HiggsCodebookSampler sampler(1, 3);
        auto state = sampler.make_state();
        HiggsSamplingOptions options;
        options.temperature = temperature_greedy ? 0.0F : 1.0F;
        options.top_k = temperature_greedy ? 3 : 1;
        options.filter_nan_logits = true;
        try {
            (void)sampler.step(logits.data(), logits.size(), state, options);
        } catch (const std::runtime_error & error) {
            if (std::string(error.what()).find("sampler kept no finite logits") != std::string::npos) {
                continue;
            }
            throw;
        }
        throw std::runtime_error("Higgs greedy filter accepted logits without a finite value");
    }
}

void test_top_k_ignores_non_finite_logits() {
    using engine::models::higgs_audio_tts::HiggsCodebookSampler;
    using engine::models::higgs_audio_tts::HiggsSamplingOptions;

    HiggsCodebookSampler sampler(1, 4);
    auto state = sampler.make_state();
    HiggsSamplingOptions options;
    options.temperature = 1.0F;
    options.top_k = 2;
    options.filter_nan_logits = true;
    options.has_seed = true;
    options.seed = 3;

    const float nan = std::numeric_limits<float>::quiet_NaN();
    const std::vector<float> logits = {0.0F, 1.0F, 2.0F, nan};
    const auto & codes = sampler.step(logits.data(), logits.size(), state, options);

    if (codes.size() != 1 || codes.front() < 1 || codes.front() > 2) {
        throw std::runtime_error("Higgs top-k sampling selected outside the finite top two");
    }
}

void test_top_k_reports_when_all_logits_are_non_finite() {
    using engine::models::higgs_audio_tts::HiggsCodebookSampler;
    using engine::models::higgs_audio_tts::HiggsSamplingOptions;

    HiggsCodebookSampler sampler(1, 3);
    auto state = sampler.make_state();
    HiggsSamplingOptions options;
    options.temperature = 1.0F;
    options.top_k = 2;
    options.filter_nan_logits = true;
    options.has_seed = true;
    const float nan = std::numeric_limits<float>::quiet_NaN();
    const float inf = std::numeric_limits<float>::infinity();
    const std::vector<float> logits = {nan, inf, -inf};

    try {
        (void)sampler.step(logits.data(), logits.size(), state, options);
    } catch (const std::runtime_error & error) {
        if (std::string(error.what()).find("sampler kept no finite logits") != std::string::npos) {
            return;
        }
        throw;
    }
    throw std::runtime_error("Higgs sampler accepted logits without a finite value");
}

} // namespace

int main() {
    test_greedy_filter_is_opt_in();
    test_greedy_filter_reports_when_all_logits_are_non_finite();
    test_top_k_ignores_non_finite_logits();
    test_top_k_reports_when_all_logits_are_non_finite();
    std::cout << "higgs_audio_tts_sampler_test passed\n";
    return 0;
}