| 82 | } |
| 83 | |
| 84 | static void test_typical(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) { |
| 85 | size_t n_vocab = probs.size(); |
| 86 | std::vector<llama_token_data> candidates; |
| 87 | candidates.reserve(n_vocab); |
| 88 | for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { |
| 89 | float logit = log(probs[token_id]); |
| 90 | candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); |
| 91 | } |
| 92 | |
| 93 | llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; |
| 94 | DUMP(&candidates_p); |
| 95 | llama_sample_typical(nullptr, &candidates_p, p, 1); |
| 96 | DUMP(&candidates_p); |
| 97 | |
| 98 | GGML_ASSERT(candidates_p.size == expected_probs.size()); |
| 99 | for (size_t i = 0; i < candidates_p.size; i++) { |
| 100 | GGML_ASSERT(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3); |
| 101 | } |
| 102 | } |
| 103 | |
| 104 | static void test_repetition_penalties( |
| 105 | const std::vector<float> & probs, const std::vector<llama_token> & last_tokens, |