Function llama_sample_token_mirostat

llama.cpp:8526–8573

}

llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu) {
    GGML_ASSERT(ctx);

    auto N = float(llama_n_vocab(llama_get_model(ctx)));
    int64_t t_start_sample_us;
    t_start_sample_us = ggml_time_us();

    llama_sample_softmax(nullptr, candidates);

    // Estimate s_hat using the most probable m tokens
    float s_hat = 0.0;
    float sum_ti_bi = 0.0;
    float sum_ti_sq = 0.0;
    for (size_t i = 0; i < size_t(m - 1) && i < candidates->size - 1; ++i) {
        float t_i = logf(float(i + 2) / float(i + 1));
        float b_i = logf(candidates->data[i].p / candidates->data[i + 1].p);
        sum_ti_bi += t_i * b_i;
        sum_ti_sq += t_i * t_i;
    }
    s_hat = sum_ti_bi / sum_ti_sq;

    // Compute k from the estimated s_hat and target surprise value
    float epsilon_hat = s_hat - 1;
    float k = powf((epsilon_hat * powf(2, *mu)) / (1 - powf(N, -epsilon_hat)), 1 / s_hat);

    // Sample the next word X using top-k sampling
    llama_sample_top_k(nullptr, candidates, int(k), 1);
    if (ctx) {
        ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
    }
    llama_token X = llama_sample_token(ctx, candidates);
    t_start_sample_us = ggml_time_us();

    // Compute error as the difference between observed surprise and target surprise value
    size_t X_idx = std::distance(candidates->data, std::find_if(candidates->data, candidates->data + candidates->size, [&](const llama_token_data & candidate) {
        return candidate.id == X;
    }));
    float observed_surprise = -log2f(candidates->data[X_idx].p);
    float e = observed_surprise - tau;

    // Update mu using the learning rate and error
    *mu = *mu - eta * e;

    if (ctx) {
        ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
    }
    return X;
}

llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu) {
    int64_t t_start_sample_us;