Function llama_sampler_softmax_impl

smallthinker/src/llama-sampling.cpp:203–226

}

static void llama_sampler_softmax_impl(llama_token_data_array * cur_p) {
    GGML_ASSERT(cur_p->size > 0);

    // Sort the logits in descending order
    if (!cur_p->sorted) {
        std::sort(cur_p->data, cur_p->data + cur_p->size, [](const llama_token_data & a, const llama_token_data & b) {
            return a.logit > b.logit;
        });
        cur_p->sorted = true;
    }

    float max_l = cur_p->data[0].logit;
    float cum_sum = 0.0f;

    for (size_t i = 0; i < cur_p->size; ++i) {
        float p = expf(cur_p->data[i].logit - max_l);
        cur_p->data[i].p = p;
        cum_sum += p;
    }

    for (size_t i = 0; i < cur_p->size; ++i) {
        cur_p->data[i].p /= cum_sum;
    }
}

static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k) {
    // TODO: move bucket sort to separate function so that top_p/typical/softmax first is equally fast