ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-05-01 11:51:53 +00:00
Code Issues Packages Projects Releases Wiki Activity

sampling: refactor sorting

Browse Source
This commit is contained in:
Kawrakow
2026-01-19 14:04:16 +00:00
parent fa58c20c42
commit efd36d2863
1 changed files with 74 additions and 105 deletions
Download Patch File Download Diff File Expand all files Collapse all files

129
src/llama-sampling.cpp
View File

@@ -31,57 +31,22 @@ void llama_set_rng_seed_impl(struct llama_sampling * smpl, uint32_t seed) {
smpl->rng.seed(seed);
}
void llama_sample_softmax_impl(struct llama_sampling * smpl, llama_token_data_array * candidates) {
GGML_ASSERT(candidates->size > 0);
const int64_t t_start_sample_us = ggml_time_us();
// Sort the logits in descending order
if (!candidates->sorted) {
std::sort(candidates->data, candidates->data + candidates->size, [](const llama_token_data & a, const llama_token_data & b) {
return a.logit > b.logit;
});
candidates->sorted = true;
static void llama_sort(llama_token_data_array * candidates, int32_t k) {
if (candidates->sorted || candidates->size < 2) {
return;
}
float max_l = candidates->data[0].logit;
float cum_sum = 0.0f;
for (size_t i = 0; i < candidates->size; ++i) {
float p = expf(candidates->data[i].logit - max_l);
candidates->data[i].p = p;
cum_sum += p;
}
for (size_t i = 0; i < candidates->size; ++i) {
candidates->data[i].p /= cum_sum;
}
if (smpl) {
smpl->t_sample_us += ggml_time_us() - t_start_sample_us;
}
}
void llama_sample_top_k_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, int32_t k, size_t min_keep) {
// TODO: move bucket sort to separate function so that top_p/tail_free/typical/softmax first is equally fast
// if (k >= (int32_t)candidates->size) {
// return;
// }
const int64_t t_start_sample_us = ggml_time_us();
if (k <= 0) {
if (k < 0) {
k = candidates->size;
}
k = std::max(k, (int) min_keep);
k = std::min(k, (int) candidates->size);
// Sort scores in descending order
if (!candidates->sorted) {
auto comp = [](const llama_token_data & a, const llama_token_data & b) {
return a.logit > b.logit;
};
if (k <= 128) {
if (k <= 1024) { //128) {
if (k == int(candidates->size)) {
std::sort(candidates->data, candidates->data + candidates->size, comp);
} else {
std::partial_sort(candidates->data, candidates->data + k, candidates->data + candidates->size, comp);
}
} else {
constexpr int nbuckets = 128;
constexpr float bucket_low = -10.0f;
@@ -134,6 +99,44 @@ void llama_sample_top_k_impl(struct llama_sampling * smpl, llama_token_data_arra
}
candidates->sorted = true;
}
void llama_sample_softmax_impl(struct llama_sampling * smpl, llama_token_data_array * candidates) {
GGML_ASSERT(candidates->size > 0);
const int64_t t_start_sample_us = ggml_time_us();
// Sort the logits in descending order if necessary
llama_sort(candidates, -1);
float max_l = candidates->data[0].logit;
float cum_sum = 0.0f;
for (size_t i = 0; i < candidates->size; ++i) {
float p = expf(candidates->data[i].logit - max_l);
candidates->data[i].p = p;
cum_sum += p;
}
for (size_t i = 0; i < candidates->size; ++i) {
candidates->data[i].p /= cum_sum;
}
if (smpl) {
smpl->t_sample_us += ggml_time_us() - t_start_sample_us;
}
}
void llama_sample_top_k_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, int32_t k, size_t min_keep) {
const int64_t t_start_sample_us = ggml_time_us();
if (k <= 0) {
k = candidates->size;
}
k = std::max(k, (int) min_keep);
k = std::min(k, (int) candidates->size);
llama_sort(candidates, k);
candidates->size = k;
if (smpl) {
@@ -208,13 +211,8 @@ void llama_sample_min_p_impl(struct llama_sampling * smpl, llama_token_data_arra
// if the candidates are sorted or the unsorted implementation failed, use this implementation
if (!min_p_applied) {
// Sort the logits in descending order
if (!candidates->sorted) {
std::sort(candidates->data, candidates->data + candidates->size, [](const llama_token_data & a, const llama_token_data & b) {
return a.logit > b.logit;
});
candidates->sorted = true;
}
// Sort the logits in descending order if needed
llama_sort(candidates, -1);
const float min_logit = candidates->data[0].logit + logf(p); // min logit for p_i >= p * p_max
size_t i = 1; // first token always matches
@@ -1178,35 +1176,6 @@ void llama_prep_adaptive_p_impl(struct llama_sampling * smpl,
adapt_p_ctx->cum_orig_prob = cum_prob;
if (smpl) smpl->t_sample_us += ggml_time_us() - t_start;
//if (!candidates->sorted) {
// std::sort(candidates->data, candidates->data + candidates->size,
// [](const llama_token_data & a, const llama_token_data & b) {
// return a.logit > b.logit;
// });
// candidates->sorted = true;
//}
//const float max_logit = candidates->data[0].logit;
//// decide how many tokens to track based on logit delta
//// i.e. do not track unlikely tokens
//auto iter = std::lower_bound(
// candidates->data,
// candidates->data + candidates->size,
// max_logit - kDelta, // delta
// [](const llama_token_data & data, const float delta) {
// return data.logit > delta;
// });
//const size_t n_track = std::distance(candidates->data, iter);
//// store orig_prob_map and cum_orig_prob to estimate original probability later
//float cum_prob = 0.0f;
//adapt_p_ctx->orig_prob_map.clear();
//for (size_t i = 0; i < n_track; ++i) {
// const float prob = expf(candidates->data[i].logit - max_logit);
// cum_prob += prob;
// adapt_p_ctx->orig_prob_map[candidates->data[i].id] = prob;
//}
//adapt_p_ctx->cum_orig_prob = cum_prob;
}
struct llama_sampler_adaptive_p * llama_init_adaptive_p_impl(