tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-28 20:25:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama : add llama_sampling API + move grammar in libllama

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2024-08-05 10:08:25 +03:00
parent b69a480af4
commit f648ca2cee
48 changed files with 2481 additions and 2590 deletions
Download Patch File Download Diff File Expand all files Collapse all files

59
tests/test-sampling.cpp
View File

@ -1,5 +1,6 @@
#include "ggml.h"
#include "llama.h"
#include "llama-sampling.h"
#ifdef NDEBUG
#undef NDEBUG
@ -20,6 +21,7 @@ static void dump(const llama_token_data_array * candidates) {
static void test_top_k(const std::vector<float> & probs, const std::vector<float> & expected_probs, int k) {
const size_t n_vocab = probs.size();
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
@ -28,9 +30,9 @@ static void test_top_k(const std::vector<float> & probs, const std::vector<float
}
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
llama_sample_softmax(nullptr, &candidates_p);
llama_sampling_softmax_impl(&candidates_p);
DUMP(&candidates_p);
llama_sample_top_k(nullptr, &candidates_p, k, 1);
llama_sampling_top_k_impl(&candidates_p, k, 1);
DUMP(&candidates_p);
GGML_ASSERT(candidates_p.size == expected_probs.size());
@ -41,6 +43,7 @@ static void test_top_k(const std::vector<float> & probs, const std::vector<float
static void test_top_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
const size_t n_vocab = probs.size();
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
@ -49,9 +52,9 @@ static void test_top_p(const std::vector<float> & probs, const std::vector<float
}
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
llama_sample_softmax(nullptr, &candidates_p);
llama_sampling_softmax_impl(&candidates_p);
DUMP(&candidates_p);
llama_sample_top_p(nullptr, &candidates_p, p, 1);
llama_sampling_top_p_impl(&candidates_p, p, 1);
DUMP(&candidates_p);
GGML_ASSERT(candidates_p.size == expected_probs.size());
@ -62,6 +65,7 @@ static void test_top_p(const std::vector<float> & probs, const std::vector<float
static void test_tfs(const std::vector<float> & probs, const std::vector<float> & expected_probs, float z) {
const size_t n_vocab = probs.size();
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
@ -71,7 +75,7 @@ static void test_tfs(const std::vector<float> & probs, const std::vector<float>
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
DUMP(&candidates_p);
llama_sample_tail_free(nullptr, &candidates_p, z, 1);
llama_sampling_tail_free_impl(&candidates_p, z, 1);
DUMP(&candidates_p);
GGML_ASSERT(candidates_p.size == expected_probs.size());
@ -82,6 +86,7 @@ static void test_tfs(const std::vector<float> & probs, const std::vector<float>
static void test_min_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
const size_t n_vocab = probs.size();
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
@ -91,9 +96,9 @@ static void test_min_p(const std::vector<float> & probs, const std::vector<float
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
DUMP(&candidates_p);
llama_sample_min_p(nullptr, &candidates_p, p, 1);
llama_sampling_min_p_impl(&candidates_p, p, 1);
DUMP(&candidates_p);
llama_sample_softmax(nullptr, &candidates_p);
llama_sampling_softmax_impl(&candidates_p);
GGML_ASSERT(candidates_p.size == expected_probs.size());
for (size_t i = 0; i < candidates_p.size; i++) {
@ -103,6 +108,7 @@ static void test_min_p(const std::vector<float> & probs, const std::vector<float
static void test_typical(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
const size_t n_vocab = probs.size();
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
@ -112,7 +118,7 @@ static void test_typical(const std::vector<float> & probs, const std::vector<flo
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
DUMP(&candidates_p);
llama_sample_typical(nullptr, &candidates_p, p, 1);
llama_sampling_typical_impl(&candidates_p, p, 1);
DUMP(&candidates_p);
GGML_ASSERT(candidates_p.size == expected_probs.size());
@ -121,13 +127,14 @@ static void test_typical(const std::vector<float> & probs, const std::vector<flo
}
}
static void test_repetition_penalties(
static void test_penalties(
const std::vector<float> & probs, const std::vector<llama_token> & last_tokens,
const std::vector<float> & expected_probs, float repeat_penalty, float alpha_frequency, float alpha_presence
) {
GGML_ASSERT(probs.size() == expected_probs.size());
const size_t n_vocab = probs.size();
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
@ -135,11 +142,16 @@ static void test_repetition_penalties(
candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
}
llama_token_cnt token_count;
for (size_t i = 0; i < last_tokens.size(); i++) {
token_count[last_tokens[i]]++;
}
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
llama_sample_softmax(nullptr, &candidates_p);
llama_sampling_softmax_impl(&candidates_p);
DUMP(&candidates_p);
llama_sample_repetition_penalties(nullptr, &candidates_p, (const llama_token *) last_tokens.data(), last_tokens.size(), repeat_penalty, alpha_frequency, alpha_presence);
llama_sample_softmax(nullptr, &candidates_p);
llama_sampling_penalties_impl(&candidates_p, token_count, repeat_penalty, alpha_frequency, alpha_presence);
llama_sampling_softmax_impl(&candidates_p);
DUMP(&candidates_p);
GGML_ASSERT(candidates_p.size == expected_probs.size());
@ -148,8 +160,7 @@ static void test_repetition_penalties(
}
}
static void test_sampler_queue(
const size_t n_vocab, const std::string samplers_sequence, const int top_k, const float top_p, const float min_p
static void test_sampler_queue(const size_t n_vocab, const std::string & samplers_sequence, const int top_k, const float top_p, const float min_p
) {
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
@ -165,16 +176,16 @@ static void test_sampler_queue(
for (auto s : samplers_sequence) {
switch (s){
case 'k': llama_sample_top_k (nullptr, &candidates_p, top_k, 1); break;
case 'k': llama_sampling_top_k_impl(&candidates_p, top_k, 1); break;
case 'f': GGML_ABORT("tail_free test not implemented");
case 'y': GGML_ABORT("typical test not implemented");
case 'p': llama_sample_top_p (nullptr, &candidates_p, top_p, 1); break;
case 'm': llama_sample_min_p (nullptr, &candidates_p, min_p, 1); break;
case 'p': llama_sampling_top_p_impl(&candidates_p, top_p, 1); break;
case 'm': llama_sampling_min_p_impl(&candidates_p, min_p, 1); break;
case 't': GGML_ABORT("temperature test not implemented");
default : GGML_ABORT("Unknown sampler");
}
llama_sample_softmax(nullptr, &candidates_p); // make sure tokens are sorted for tests
llama_sampling_softmax_impl(&candidates_p); // make sure tokens are sorted for tests
const int size = candidates_p.size;
@ -259,13 +270,13 @@ int main(void) {
test_typical({0.97f, 0.01f, 0.01f, 0.01f}, {0.97f}, 0.5f);
test_typical({0.4f, 0.2f, 0.2f, 0.2f}, {0.2f, 0.2f, 0.2f}, 0.5f);
test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0.25f, 0.25f, 0.25f, 0.25f, 0}, 50.0f, 0.0f, 0.0f);
test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.5f, 0.5f, 0, 0, 0}, 50.0f, 0.0f, 0.0f);
test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.5f, 0.5f, 0, 0, 0}, 50.0f, 0.0f, 0.0f);
test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0.25f, 0.25f, 0.25f, 0.25f, 0}, 50.0f, 0.0f, 0.0f);
test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.5f, 0.5f, 0, 0, 0}, 50.0f, 0.0f, 0.0f);
test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.5f, 0.5f, 0, 0, 0}, 50.0f, 0.0f, 0.0f);
test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0.249997f, 0.249997f, 0.249997f, 0.249997f, 0.000011f}, 1.0f, 5.0f, 5.0f);
test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.499966f, 0.499966f, 0.000023f, 0.000023f, 0.000023f}, 1.0f, 5.0f, 5.0f);
test_repetition_penalties({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}, 1.0f, 5.0f, 5.0f);
test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0.249997f, 0.249997f, 0.249997f, 0.249997f, 0.000011f}, 1.0f, 5.0f, 5.0f);
test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.499966f, 0.499966f, 0.000023f, 0.000023f, 0.000023f}, 1.0f, 5.0f, 5.0f);
test_penalties({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}, 1.0f, 5.0f, 5.0f);
test_sampler_queue(10000, "k", 10000, 1.0f, 1.0f);
test_sampler_queue(10000, "k", 1, 1.0f, 1.0f);