extern/whisper.cpp
1
0
Fork 1
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-08-09 06:45:12 +02:00
Code Issues Packages Projects Releases Wiki Activity

talk-llama : sync llama.cpp

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2025-05-13 13:20:19 +03:00
parent a14c89aefa
commit f890560575
25 changed files with 2847 additions and 1125 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
examples/talk-llama/llama-sampling.cpp
View File

@ -1750,23 +1750,35 @@ static const char * llama_sampler_top_n_sigma_name(const struct llama_sampler *
static void llama_sampler_top_n_sigma_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
const auto * ctx = (llama_sampler_top_n_sigma *) smpl->ctx;
if (ctx->n <= 0.0f || cur_p->size <= 1) {
return;
}
// find max logit and calculate mean
float max = cur_p->data[0].logit;
float logits_sum = 0;
size_t valid_count = 0;
for (size_t i = 0; i < cur_p->size; ++i) {
if (cur_p->data[i].logit > max) {
max = cur_p->data[i].logit;
// Only count non-negative infinity values
if (cur_p->data[i].logit != -INFINITY) {
if (cur_p->data[i].logit > max) {
max = cur_p->data[i].logit;
}
logits_sum += cur_p->data[i].logit;
valid_count++;
}
logits_sum += cur_p->data[i].logit;
}
float mean = logits_sum/cur_p->size;
float mean = valid_count > 0 ? logits_sum/valid_count : 0;
// calculate standard deviation
float acc = 0;
for (size_t i = 0; i < cur_p->size; ++i) {
acc += pow(cur_p->data[i].logit - mean, 2);
// Skip -infinity in std calculation
if (cur_p->data[i].logit != -INFINITY) {
acc += pow(cur_p->data[i].logit - mean, 2);
}
}
float std = sqrt(acc/cur_p->size);
float std = valid_count > 0 ? sqrt(acc/valid_count) : 0;
//apply mask
for (size_t i = 0; i < cur_p->size; ++i) {