Function llama_sample_xtc_apply

smallthinker/src/llama-sampling.cpp:1131–1159

}

static void llama_sample_xtc_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
    auto * ctx = (llama_sampler_xtc *) smpl->ctx;

    if (ctx->probability <= 0.0f
        || ctx->threshold > 0.5f
        || cur_p->size < 2) {
        return;
    }

    std::uniform_real_distribution<float> distribution(0.0f, 1.0f);
    float chance = distribution(ctx->rng);
    if (chance > ctx->probability) return;

    // in case it's not sorted/recalculated yet
    llama_sampler_softmax_impl(cur_p);

    int pos_last = 0;

    for (size_t i = 0; i < cur_p->size; ++i) {
        if (cur_p->data[i].p >= ctx->threshold) {
            pos_last = i;
        } else break;
    }

    if (cur_p->size - pos_last >= ctx->min_keep && pos_last > 0) {
        cur_p->data += pos_last;
        cur_p->size -= pos_last;
    }
}

static struct llama_sampler * llama_sampler_xtc_clone(const struct llama_sampler * smpl) {
    const auto * ctx = (const llama_sampler_xtc *) smpl->ctx;