Function llama_sampler_temp_ext_apply

smallthinker/src/llama-sampling.cpp:1014–1083

}

static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
    const auto * ctx = (llama_sampler_temp_ext *) smpl->ctx;
    if (ctx->delta > 0) {
        const float min_temp = std::max(0.0f, ctx->temp - ctx->delta);
        const float max_temp = ctx->temp + ctx->delta;

        float exponent_val = ctx->exponent;

        // no need to do anything if there is only one (or zero) candidates
        if (cur_p->size <= 1) {
            return;
        }

        // Calculate maximum possible entropy
        float max_entropy = -logf(1.0f / cur_p->size);

        llama_sampler_softmax_impl(cur_p);

        // Calculate entropy of the softmax probabilities
        float entropy = 0.0f;
        for (size_t i = 0; i < cur_p->size; ++i) {
            float prob = cur_p->data[i].p;
            if (prob > 0.0f) { // Ensure no log(0)
                entropy -= prob * logf(prob);
            }
        }

        // Normalize the entropy (max_entropy cannot be 0 here because we checked cur_p->size != 1 above)
        float normalized_entropy = entropy / max_entropy;

        // Map the normalized entropy to the desired temperature range using the power function
        float dyn_temp = min_temp + (max_temp - min_temp) * powf(normalized_entropy, exponent_val);

    #ifdef DEBUG
        LLAMA_LOG_INFO("Your text maxtemp value is: %f\n", max_temp);
        LLAMA_LOG_INFO("Entropy: %f\n", entropy);
        LLAMA_LOG_INFO("Max Possible Entropy: %f\n", max_entropy);
        LLAMA_LOG_INFO("Normalized Entropy: %f\n", normalized_entropy);
        LLAMA_LOG_INFO("Exponent: %f\n", exponent_val);
        LLAMA_LOG_INFO("Dynamic Temperature (dyn_temp): %f\n", dyn_temp);
    #endif

        // Apply the dynamically calculated temperature scaling
        llama_sampler_temp_impl(cur_p, dyn_temp);

        // Re-compute softmax probabilities after scaling logits with dynamic temperature
        const double max_l_double = cur_p->data[0].logit;

        double cum_sum_double = 0.0;
        for (size_t i = 0; i < cur_p->size; ++i) {
            double p = exp(cur_p->data[i].logit - max_l_double);
            cur_p->data[i].p = p; // Store the scaled probability
            cum_sum_double += p;
        }

        for (size_t i = 0; i < cur_p->size; ++i) {
            cur_p->data[i].p /= cum_sum_double; // Re-normalize the probabilities
        }