From dd7eff57d8491792010b1002b8de6a4b54912e5c Mon Sep 17 00:00:00 2001 From: Ivan Stepanov Date: Sat, 29 Apr 2023 08:34:41 +0300 Subject: llama : new sampling algorithms (#1126) * Sample interface, new samplers. New samplers: - locally typical sampling - tail free sampling - frequency and presence penalty - mirostat Ignore EOS fix: -inf should be used. * mirostat * Added --logit-bias and --no-penalize-nl, removed std::span * Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k) Use C++11, clarify llama API documentation, rename Mirostat parameters to --mirostat_lr and --mirostat_ent, add temperature sampling for Mirostat, simplify Mirostat sampling API parameters (removed N and *k) * Save and load example adjust * Tests * Windows build fix * Windows test fix --- tests/test-sampling.cpp | 199 ++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 199 insertions(+) create mode 100644 tests/test-sampling.cpp (limited to 'tests/test-sampling.cpp') diff --git a/tests/test-sampling.cpp b/tests/test-sampling.cpp new file mode 100644 index 0000000..7eee4f6 --- /dev/null +++ b/tests/test-sampling.cpp @@ -0,0 +1,199 @@ +#include "llama.h" +#include "ggml.h" +#include +#include +#include +#include +#include +#include +#include + + +void dump(const llama_token_data_array * candidates) { + for (size_t i = 0; i < candidates->size; i++) { + printf("%d: %f (%f)\n", candidates->data[i].id, candidates->data[i].p, candidates->data[i].logit); + } +} + +#define DUMP(__candidates) do { printf("%s:%d (%s)\n", __FILE__, __LINE__, __func__); dump((__candidates)); printf("-\n"); } while(0) + + +void test_top_k(const std::vector & probs, + const std::vector & expected_probs, + int k) { + size_t n_vocab = probs.size(); + std::vector candidates; + candidates.reserve(n_vocab); + for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { + float logit = log(probs[token_id]); + candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); + } + + llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; + llama_sample_softmax(nullptr, &candidates_p); + DUMP(&candidates_p); + llama_sample_top_k(nullptr, &candidates_p, k); + DUMP(&candidates_p); + + assert(candidates_p.size == expected_probs.size()); + for (size_t i = 0; i < candidates_p.size; i++) { + assert(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-5); + } +} + + +void test_top_p(const std::vector & probs, + const std::vector & expected_probs, + float p) { + + size_t n_vocab = probs.size(); + std::vector candidates; + candidates.reserve(n_vocab); + for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { + float logit = log(probs[token_id]); + candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); + } + + llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; + llama_sample_softmax(nullptr, &candidates_p); + DUMP(&candidates_p); + llama_sample_top_p(nullptr, &candidates_p, p); + DUMP(&candidates_p); + + assert(candidates_p.size == expected_probs.size()); + for (size_t i = 0; i < candidates_p.size; i++) { + assert(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3); + } +} + + +void test_tfs(const std::vector & probs, + const std::vector & expected_probs, + float z) { + size_t n_vocab = probs.size(); + std::vector candidates; + candidates.reserve(n_vocab); + for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { + float logit = log(probs[token_id]); + candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); + } + + llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; + DUMP(&candidates_p); + llama_sample_tail_free(nullptr, &candidates_p, z); + DUMP(&candidates_p); + + assert(candidates_p.size == expected_probs.size()); + for (size_t i = 0; i < candidates_p.size; i++) { + assert(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3); + } +} + + +void test_typical(const std::vector & probs, + const std::vector & expected_probs, + float p) { + size_t n_vocab = probs.size(); + std::vector candidates; + candidates.reserve(n_vocab); + for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { + float logit = log(probs[token_id]); + candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); + } + + llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; + DUMP(&candidates_p); + llama_sample_typical(nullptr, &candidates_p, p); + DUMP(&candidates_p); + + assert(candidates_p.size == expected_probs.size()); + for (size_t i = 0; i < candidates_p.size; i++) { + assert(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3); + } +} + + +void test_repetition_penalty( + const std::vector & probs, + const std::vector & last_tokens, + const std::vector & expected_probs, + float penalty) { + assert(probs.size() == expected_probs.size()); + + size_t n_vocab = probs.size(); + std::vector candidates; + candidates.reserve(n_vocab); + for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { + float logit = log(probs[token_id]); + candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); + } + + llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; + llama_sample_softmax(nullptr, &candidates_p); + DUMP(&candidates_p); + llama_sample_repetition_penalty(nullptr, &candidates_p, (llama_token *)last_tokens.data(), last_tokens.size(), penalty); + llama_sample_softmax(nullptr, &candidates_p); + DUMP(&candidates_p); + + assert(candidates_p.size == expected_probs.size()); + for (size_t i = 0; i < candidates_p.size; i++) { + assert(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-6); + } +} + + +void test_frequency_presence_penalty( + const std::vector & probs, + const std::vector & last_tokens, + const std::vector & expected_probs, + float alpha_frequency, float alpha_presence) { + assert(probs.size() == expected_probs.size()); + + size_t n_vocab = probs.size(); + std::vector candidates; + candidates.reserve(n_vocab); + for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { + float logit = log(probs[token_id]); + candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); + } + + llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; + llama_sample_softmax(nullptr, &candidates_p); + // DUMP(&candidates_p); + llama_sample_frequency_and_presence_penalties(nullptr, &candidates_p, (llama_token *)last_tokens.data(), last_tokens.size(), alpha_frequency, alpha_presence); + llama_sample_softmax(nullptr, &candidates_p); + // DUMP(&candidates_p); + + assert(candidates_p.size == expected_probs.size()); + for (size_t i = 0; i < candidates_p.size; i++) { + assert(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3); + } +} + +int main(void) { + ggml_time_init(); + + test_top_k({0.1, 0.2, 0.3, 0.4}, {0.4}, 1); + test_top_k({0.1, 0.2, 0.3, 0.4}, {0.4, 0.3, 0.2}, 3); + + test_top_p({0.1, 0.2, 0.3, 0.4}, {0.4}, 0); + test_top_p({0.1, 0.2, 0.3, 0.4}, {0.4, 0.3}, 0.7); + test_top_p({0.1, 0.2, 0.3, 0.4}, {0.4, 0.3, 0.2, 0.1}, 1); + + test_tfs({0.1, 0.15, 0.2, 0.25, 0.3}, {0.3}, 0.25); + test_tfs({0.1, 0.15, 0.2, 0.25, 0.3}, {0.3, 0.25}, 0.75); + test_tfs({0.1, 0.15, 0.2, 0.25, 0.3}, {0.3, 0.25}, 0.99); + + test_typical({0.97, 0.01, 0.01, 0.01}, {0.97}, 0.5); + test_typical({0.4, 0.2, 0.2, 0.2}, {0.2, 0.2, 0.2}, 0.5); + + test_repetition_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0}, {0.25, 0.25, 0.25, 0.25, 0}, 50.0); + test_repetition_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0, 1, 2}, {0.5, 0.5, 0, 0, 0}, 50.0); + test_repetition_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0, 1, 2, 0, 0}, {0.5, 0.5, 0, 0, 0}, 50.0); + + test_frequency_presence_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0}, {0.249997, 0.249997, 0.249997, 0.249997, 0.000011}, 5.0, 5.0); + test_frequency_presence_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0, 1, 2}, {0.499966, 0.499966, 0.000023, 0.000023, 0.000023}, 5.0, 5.0); + test_frequency_presence_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0, 1, 2, 0, 0}, {0.499977, 0.499977, 0.000023, 0.000023, 0.000000}, 5.0, 5.0); + + printf("OK\n"); +} -- cgit v1.2.3