root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-10 10:41:47 +00:00
Code Issues Actions 15 Packages Projects Releases Wiki Activity

tests : replace macros with functions

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2024-10-17 17:51:54 +03:00
parent e31c8790ff
commit 33a69ec742
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
1 changed files with 107 additions and 97 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

204
tests/test-sampling.cpp
View File

@ -18,155 +18,165 @@ static void dump(const llama_token_data_array * cur_p) {
#define DUMP(__cur_p) do { printf("%s:%d (%s)\n", __FILE__, __LINE__, __func__); dump((__cur_p)); printf("-\n"); } while(0)
#define APPLY(__cnstr, __cur_p) do { \
auto * cnstr = (__cnstr); \
llama_sampler_apply(cnstr, (__cur_p)); \
llama_sampler_free(cnstr); \
} while(0)
struct sampler_tester {
sampler_tester(size_t n_vocab) {
cur.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
const float logit = logf(token_id);
cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
}
#define CUR_P_FROM_PROBS() \
const size_t n_vocab = probs.size(); \
std::vector<llama_token_data> cur; \
cur.reserve(n_vocab); \
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { \
const float logit = logf(probs[token_id]); \
cur.emplace_back(llama_token_data{token_id, logit, 0.0f}); \
} \
llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false }
static void test_top_k(const std::vector<float> & probs, const std::vector<float> & expected_probs, int k) {
CUR_P_FROM_PROBS();
DUMP(&cur_p);
APPLY(llama_sampler_init_top_k(k), &cur_p);
APPLY(llama_sampler_init_dist (0), &cur_p);
DUMP(&cur_p);
GGML_ASSERT(cur_p.size == expected_probs.size());
for (size_t i = 0; i < cur_p.size; i++) {
GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-5);
cur_p = llama_token_data_array { cur.data(), cur.size(), -1, false };
}
sampler_tester(const std::vector<float> & probs, const std::vector<float> & probs_expected) : probs_expected(probs_expected) {
cur.reserve(probs.size());
for (llama_token token_id = 0; token_id < (llama_token)probs.size(); token_id++) {
const float logit = logf(probs[token_id]);
cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
}
cur_p = llama_token_data_array { cur.data(), cur.size(), -1, false };
}
void apply(llama_sampler * sampler) {
llama_sampler_apply(sampler, &cur_p);
llama_sampler_free(sampler);
}
void check() {
GGML_ASSERT(cur_p.size == probs_expected.size());
for (size_t i = 0; i < cur_p.size; i++) {
GGML_ASSERT(fabs(cur_p.data[i].p - probs_expected[i]) < 1e-5);
}
}
llama_token_data_array cur_p;
private:
const std::vector<float> probs_expected;
std::vector<llama_token_data> cur;
};
static void test_temp(const std::vector<float> & probs, const std::vector<float> & probs_expected, float temp) {
sampler_tester tester(probs, probs_expected);
DUMP(&tester.cur_p);
tester.apply(llama_sampler_init_temp(temp));
tester.apply(llama_sampler_init_dist(0));
DUMP(&tester.cur_p);
tester.check();
}
static void test_top_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
CUR_P_FROM_PROBS();
static void test_top_k(const std::vector<float> & probs, const std::vector<float> & probs_expected, int k) {
sampler_tester tester(probs, probs_expected);
DUMP(&cur_p);
APPLY(llama_sampler_init_top_p(p, 1), &cur_p);
APPLY(llama_sampler_init_dist (0), &cur_p);
DUMP(&cur_p);
DUMP(&cur_p);
DUMP(&tester.cur_p);
tester.apply(llama_sampler_init_top_k(k));
tester.apply(llama_sampler_init_dist (0));
DUMP(&tester.cur_p);
GGML_ASSERT(cur_p.size == expected_probs.size());
for (size_t i = 0; i < cur_p.size; i++) {
GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-3);
}
tester.check();
}
static void test_tfs(const std::vector<float> & probs, const std::vector<float> & expected_probs, float z) {
CUR_P_FROM_PROBS();
static void test_top_p(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p) {
sampler_tester tester(probs, probs_expected);
DUMP(&cur_p);
APPLY(llama_sampler_init_tail_free(z, 1), &cur_p);
DUMP(&cur_p);
DUMP(&tester.cur_p);
tester.apply(llama_sampler_init_top_p(p, 1));
tester.apply(llama_sampler_init_dist (0));
DUMP(&tester.cur_p);
GGML_ASSERT(cur_p.size == expected_probs.size());
for (size_t i = 0; i < cur_p.size; i++) {
GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-3);
}
tester.check();
}
static void test_min_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
CUR_P_FROM_PROBS();
static void test_tfs(const std::vector<float> & probs, const std::vector<float> & probs_expected, float z) {
sampler_tester tester(probs, probs_expected);
DUMP(&cur_p);
APPLY(llama_sampler_init_min_p(p, 1), &cur_p);
APPLY(llama_sampler_init_dist (0), &cur_p);
DUMP(&cur_p);
DUMP(&tester.cur_p);
tester.apply(llama_sampler_init_tail_free(z, 1));
DUMP(&tester.cur_p);
GGML_ASSERT(cur_p.size == expected_probs.size());
for (size_t i = 0; i < cur_p.size; i++) {
GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-3);
}
tester.check();
}
static void test_xtc(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p, float t) {
CUR_P_FROM_PROBS();
static void test_min_p(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p) {
sampler_tester tester(probs, probs_expected);
DUMP(&cur_p);
APPLY(llama_sampler_init_xtc(p, t, 0, 0), &cur_p);
DUMP(&cur_p);
DUMP(&tester.cur_p);
tester.apply(llama_sampler_init_min_p(p, 1));
tester.apply(llama_sampler_init_dist (0));
DUMP(&tester.cur_p);
GGML_ASSERT(cur_p.size == expected_probs.size());
for (size_t i = 0; i < cur_p.size; i++) {
GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-5);
}
tester.check();
}
static void test_typical(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
CUR_P_FROM_PROBS();
static void test_xtc(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p, float t) {
sampler_tester tester(probs, probs_expected);
DUMP(&cur_p);
APPLY(llama_sampler_init_typical(p, 1), &cur_p);
DUMP(&cur_p);
DUMP(&tester.cur_p);
tester.apply(llama_sampler_init_xtc(p, t, 0, 0));
DUMP(&tester.cur_p);
GGML_ASSERT(cur_p.size == expected_probs.size());
for (size_t i = 0; i < cur_p.size; i++) {
GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-3);
}
tester.check();
}
static void test_typical(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p) {
sampler_tester tester(probs, probs_expected);
DUMP(&tester.cur_p);
tester.apply(llama_sampler_init_typical(p, 1));
DUMP(&tester.cur_p);
tester.check();
}
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
const std::vector<float> & probs_expected, float repeat_penalty, float alpha_frequency, float alpha_presence
) {
GGML_ASSERT(probs.size() == expected_probs.size());
GGML_ASSERT(probs.size() == probs_expected.size());
CUR_P_FROM_PROBS();
sampler_tester tester(probs, probs_expected);
const size_t n_vocab = probs.size();
auto * sampler = llama_sampler_init_penalties(n_vocab, LLAMA_TOKEN_NULL, LLAMA_TOKEN_NULL, last_tokens.size(), repeat_penalty, alpha_frequency, alpha_presence, false, false);
for (size_t i = 0; i < last_tokens.size(); i++) {
llama_sampler_accept(sampler, last_tokens[i]);
}
DUMP(&cur_p);
APPLY(sampler, &cur_p);
APPLY(llama_sampler_init_dist(0), &cur_p);
DUMP(&cur_p);
DUMP(&tester.cur_p);
tester.apply(sampler);
tester.apply(llama_sampler_init_dist(0));
DUMP(&tester.cur_p);
GGML_ASSERT(cur_p.size == expected_probs.size());
for (size_t i = 0; i < cur_p.size; i++) {
GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-3);
}
tester.check();
}
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> cur;
cur.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
const float logit = logf(token_id);
cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
}
llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
sampler_tester tester(n_vocab);
llama_token min_token_id = 0;
const llama_token max_token_id = n_vocab-1;
for (auto s : samplers_sequence) {
switch (s){
case 'k': APPLY(llama_sampler_init_top_k(top_k), &cur_p); break;
case 'k': tester.apply(llama_sampler_init_top_k(top_k)); break;
case 'f': GGML_ABORT("tail_free test not implemented");
case 'y': GGML_ABORT("typical test not implemented");
case 'p': APPLY(llama_sampler_init_top_p(top_p, 1), &cur_p); break;
case 'm': APPLY(llama_sampler_init_min_p(min_p, 1), &cur_p); break;
case 'p': tester.apply(llama_sampler_init_top_p(top_p, 1)); break;
case 'm': tester.apply(llama_sampler_init_min_p(min_p, 1)); break;
case 't': GGML_ABORT("temperature test not implemented");
default : GGML_ABORT("Unknown sampler");
}
APPLY(llama_sampler_init_dist(0), &cur_p);
tester.apply(llama_sampler_init_dist(0));
auto & cur_p = tester.cur_p;
const int size = cur_p.size;