CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

sampling : optimize samplers by reusing bucket sort (#15665)

Browse Source 
* sampling : optimize sorting using bucket sort in more places

ggml-ci

* sampling : do not sort in dist sampler

ggml-ci

* sampling : avoid heap allocations for sort buffers

ggml-ci

* common : add option to sort sampling candidates by probability

ggml-ci

* sampling : revert the change for preserving sort buffers

* sampling : use std::copy instead of memcpy

* sampling : clarify purpose of partial sort helpers

ggml-ci

* cont : remove wrong comment [no ci]

* common : update comment

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

---------

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
This commit is contained in:
Georgi Gerganov
2025-08-31 20:41:02 +03:00
committed by GitHub
parent 0d161f021a
commit e92d53b29e
9 changed files with 227 additions and 171 deletions
Download Patch File Download Diff File Expand all files Collapse all files

60
tests/test-sampling.cpp
View File

@@ -197,10 +197,10 @@ static void test_sampler_queue(const size_t n_vocab, const std::string & sampler
sampler_tester tester(n_vocab);
llama_token min_token_id = 0;
const llama_token max_token_id = n_vocab-1;
const llama_token max_token_id = n_vocab - 1;
for (auto s : samplers_sequence) {
switch (s){
switch (s) {
case 'k': tester.apply(llama_sampler_init_top_k(top_k)); break;
case 'y': GGML_ABORT("typical test not implemented");
case 'p': tester.apply(llama_sampler_init_top_p(top_p, 1)); break;
@@ -243,10 +243,10 @@ static void test_sampler_queue(const size_t n_vocab, const std::string & sampler
}
GGML_ASSERT(size == expected_size);
GGML_ASSERT(cur_p.data[0].id == max_token_id);
GGML_ASSERT(cur_p.data[expected_size-1].id == min_token_id);
GGML_ASSERT(!cur_p.sorted || cur_p.data[0].id == max_token_id);
GGML_ASSERT(!cur_p.sorted || cur_p.data[expected_size-1].id == min_token_id);
} else if (s == 'm') {
int expected_size = ceilf((1.0f-min_p) * n_vocab);
int expected_size = ceilf((1.0f - min_p) * n_vocab);
expected_size = std::max(expected_size, 1);
expected_size = std::min(expected_size, size);
@@ -256,14 +256,14 @@ static void test_sampler_queue(const size_t n_vocab, const std::string & sampler
min_token_id = std::min(min_token_id, (llama_token)(n_vocab - 1));
GGML_ASSERT(size == expected_size);
GGML_ASSERT(cur_p.data[0].id == max_token_id);
GGML_ASSERT(cur_p.data[expected_size-1].id == min_token_id);
GGML_ASSERT(!cur_p.sorted || cur_p.data[0].id == max_token_id);
GGML_ASSERT(!cur_p.sorted || cur_p.data[expected_size-1].id == min_token_id);
} else {
GGML_ABORT("fatal error");
}
}
printf("Sampler queue %3s OK with n_vocab=%05zu top_k=%05d top_p=%f min_p=%f\n",
printf("Sampler queue %3s OK with n_vocab=%05zu top_k=%5d top_p=%f min_p=%f\n",
samplers_sequence.c_str(), n_vocab, top_k, top_p, min_p);
}
@@ -308,28 +308,28 @@ static void test_perf() {
int main(void) {
ggml_time_init();
test_temp({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1.0f);
test_temp({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f, 0.0f, 0.0f, 0.0f}, 0.0f);
test_temp({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f, 0.2f, 0.3f, 0.4f}, 1.0f);
test_temp({0.1f, 0.2f, 0.3f, 0.4f}, {0.0f, 0.0f, 0.0f, 1.0f}, 0.0f);
test_temp_ext({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1.0f, 0.0f, 1.0f);
test_temp_ext({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f, 0.0f, 0.0f, 0.0f}, 0.0f, 0.0f, 1.0f);
test_temp_ext({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f, 0.2f, 0.3f, 0.4f}, 1.0f, 0.0f, 1.0f);
test_temp_ext({0.1f, 0.2f, 0.3f, 0.4f}, {0.0f, 0.0f, 0.0f, 1.0f}, 0.0f, 0.0f, 1.0f);
test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f}, 1);
test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.44444f, 0.33333f, 0.22222f}, 3);
test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 4);
test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 0);
test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f, 0.2f, 0.3f, 0.4f}, 0);
test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f}, 0);
test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.571429f, 0.428571f}, 0.7f);
test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.44444f, 0.33333f, 0.22222f}, 0.8f);
test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1.0f);
test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f, 0.2f, 0.3f, 0.4f}, 1.0f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/1.0f, 0.3f/1.0f, 0.2f/1.0f, 0.1f/1.0f}, 0.00f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/1.0f, 0.3f/1.0f, 0.2f/1.0f, 0.1f/1.0f}, 0.24f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.9f, 0.3f/0.9f, 0.2f/0.9f}, 0.26f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.9f, 0.3f/0.9f, 0.2f/0.9f}, 0.49f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.7f, 0.3f/0.7f}, 0.51f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.7f, 0.3f/0.7f}, 0.74f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f/1.0f, 0.2f/1.0f, 0.3f/1.0f, 0.4f/1.0f}, 0.00f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f/1.0f, 0.2f/1.0f, 0.3f/1.0f, 0.4f/1.0f}, 0.24f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.2f/0.9f, 0.3f/0.9f, 0.4f/0.9f}, 0.26f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.2f/0.9f, 0.3f/0.9f, 0.4f/0.9f}, 0.49f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.3f/0.7f, 0.4f/0.7f}, 0.51f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.3f/0.7f, 0.4f/0.7f}, 0.74f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f}, 0.76f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f}, 1.00f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f}, 1.05f);
@@ -345,23 +345,23 @@ 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_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_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0, 0.25f, 0.25f, 0.25f, 0.25f}, 50.0f, 0.0f, 0.0f);
test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0, 0, 0, 0.5f, 0.5f}, 50.0f, 0.0f, 0.0f);
test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0, 0, 0, 0.5f, 0.5f}, 50.0f, 0.0f, 0.0f);
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_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0.000011f, 0.249997f, 0.249997f, 0.249997f, 0.249997f}, 1.0f, 5.0f, 5.0f);
test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.000023f, 0.000023f, 0.000023f, 0.499966f, 0.499966f}, 1.0f, 5.0f, 5.0f);
test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.000000f, 0.000023f, 0.000023f, 0.499977f, 0.499977f}, 1.0f, 5.0f, 5.0f);
test_dry({0.25f, 0.25f, 0.25f, 0.25f}, {0, 1}, {0.25f, 0.25f, 0.25f, 0.25f}, 1.0f, 1.1f, 2, 4, {});
test_dry({0.25f, 0.25f, 0.25f, 0.25f}, {0, 1, 2, 0, 1}, {0.296923f, 0.296923f, 0.296923f, 0.109232f}, 1.0f, 1.1f, 2, 5, {});
test_dry({0.25f, 0.25f, 0.25f, 0.25f}, {0, 1, 2, 0, 1}, {0.296923f, 0.296923f, 0.109232f, 0.296923f}, 1.0f, 1.1f, 2, 5, {});
test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 3, 4, 0, 1}, {0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, 1.0f, 1.1f, 2, 6, {{3}});
test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 1}, {0.241818f, 0.241818f, 0.241818f, 0.241818f, 0.032727f}, 2.0f, 1.1f, 2, 5, {});
test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 1}, {0.241818f, 0.241818f, 0.032727f, 0.241818f, 0.241818f}, 2.0f, 1.1f, 2, 5, {});
test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 3, 4, 0, 1}, {0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, 1.0f, 1.1f, 4, 7, {});
test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.571429f, 0.428571f, 0.0f, 0.0f}, 1.00f);
test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 0.00f); // top_n_sigma == 0 now represents a no-op rather than greedy decoding as of PR#13345
test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f, 0.2f, 0.3f, 0.4f}, 0.00f); // top_n_sigma == 0 now represents a no-op rather than greedy decoding as of PR#13345
test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 3.00f);
test_sampler_queue(10000, "k", 10000, 1.0f, 1.0f);
@@ -372,7 +372,7 @@ int main(void) {
test_sampler_queue(10000, "m", 10000, 1.0f, 1e-12);
test_sampler_queue(10000, "k", 100, 1.0000f, 1.0f);
test_sampler_queue(10000, "p", 10000, 0.0002f, 1.0f);
test_sampler_queue(10000, "p", 10000, 0.0003f, 1.0f);
test_sampler_queue(10000, "p", 10000, 0.8000f, 1.0f);
test_sampler_queue(10000, "m", 10000, 1.0000f, 9997.9f/9999.0f);
test_sampler_queue(10000, "m", 10000, 1.0000f, 0.1f);