diff --git a/llama.cpp b/llama.cpp index 0f9d5346d..fa3b7c03c 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2024,9 +2024,18 @@ void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * } // Normalize the second derivatives - float second_derivatives_sum = std::accumulate(second_derivatives.begin(), second_derivatives.end(), 0.0f); - for (float & value : second_derivatives) { - value /= second_derivatives_sum; + { + const float second_derivatives_sum = std::accumulate(second_derivatives.begin(), second_derivatives.end(), 0.0f); + + if (second_derivatives_sum > 1e-6f) { + for (float & value : second_derivatives) { + value /= second_derivatives_sum; + } + } else { + for (float & value : second_derivatives) { + value = 1.0f / second_derivatives.size(); + } + } } float cum_sum = 0.0f; diff --git a/tests/test-sampling.cpp b/tests/test-sampling.cpp index 64f9455d7..4437c3948 100644 --- a/tests/test-sampling.cpp +++ b/tests/test-sampling.cpp @@ -200,4 +200,6 @@ int main(void) { test_frequency_presence_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.499977f, 0.499977f, 0.000023f, 0.000023f, 0.000000f}, 5.0f, 5.0f); printf("OK\n"); + + return 0; }