vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Add fixes to newer changes

Browse source
This commit is contained in:
Michael Klimenko 2024-01-28 18:34:27 +01:00
parent 024e566389
commit e75d00797e
1 changed files with 9 additions and 9 deletions
Download patch file Download diff file Expand all files Collapse all files

18
tests/test-sampling.cpp
View file

@ -154,14 +154,14 @@ static void test_sampler_queue(
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
const float logit = logf(token_id);
const float logit = logf(float(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_token min_token_id = 0;
const llama_token max_token_id = n_vocab-1;
const llama_token max_token_id = llama_token(n_vocab-1);
for (auto s : samplers_sequence) {
switch (s){
@ -176,7 +176,7 @@ static void test_sampler_queue(
llama_sample_softmax(nullptr, &candidates_p); // make sure tokens are sorted for tests
const int size = candidates_p.size;
const int size = int(candidates_p.size);
if (s == 'k') {
const int expected_size = std::min(size, top_k);
@ -186,10 +186,10 @@ static void test_sampler_queue(
GGML_ASSERT(candidates_p.data[0].id == max_token_id);
GGML_ASSERT(candidates_p.data[expected_size-1].id == min_token_id);
} else if (s == 'p') {
const int softmax_divisor = n_vocab * (n_vocab-1) / 2 - min_token_id * (min_token_id-1) / 2;
const int softmax_numerator_target = ceilf(top_p * softmax_divisor);
const int softmax_divisor = int(n_vocab * (n_vocab-1) / 2 - min_token_id * (min_token_id-1) / 2);
const int softmax_numerator_target = int(ceilf(top_p * softmax_divisor));
min_token_id = n_vocab;
min_token_id = llama_token(n_vocab);
int expected_size = 0;
int cumsum = 0;
do { // do-while because always at least one token is sampled
@ -209,11 +209,11 @@ static void test_sampler_queue(
GGML_ASSERT(candidates_p.data[0].id == max_token_id);
GGML_ASSERT(candidates_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 = int(ceilf((1.0f-min_p) * n_vocab));
expected_size = std::max(expected_size, 1);
expected_size = std::min(expected_size, size);
min_token_id = floorf(min_p * n_vocab);
min_token_id = llama_token(floorf(min_p * n_vocab));
min_token_id = std::max(min_token_id, 1);
min_token_id = std::max(min_token_id, (llama_token)(n_vocab - size));
min_token_id = std::min(min_token_id, (llama_token)(n_vocab - 1));
@ -270,7 +270,7 @@ int main(void) {
test_sampler_queue(10000, "p", 10000, 1.0f, 1.0f);
test_sampler_queue(10000, "p", 10000, 0.0f, 1.0f);
test_sampler_queue(10000, "m", 10000, 1.0f, 1.0f);
test_sampler_queue(10000, "m", 10000, 1.0f, 1e-12);
test_sampler_queue(10000, "m", 10000, 1.0f, 1e-12f);
test_sampler_queue(10000, "k", 100, 1.0000f, 1.0f);
test_sampler_queue(10000, "p", 10000, 0.0002f, 1.0f);