sdgoij/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-05-11 19:44:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

common : revert reasoning budget +inf logit bias (#22740)

Browse Source
This commit is contained in:
Aldehir Rojas
2026-05-08 10:46:43 -05:00
committed by GitHub
parent 5d6f18a638
commit f9cd456ea5
2 changed files with 7 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
common/reasoning-budget.cpp
View File

@@ -158,8 +158,6 @@ static void common_reasoning_budget_apply(struct llama_sampler * smpl, llama_tok
for (size_t i = 0; i < cur_p->size; i++) {
if (cur_p->data[i].id != forced) {
cur_p->data[i].logit = -INFINITY;
} else {
cur_p->data[i].logit = +INFINITY; // force the token
}
}
}

14
tests/test-reasoning-budget.cpp
View File

@@ -70,20 +70,20 @@ static void test_reasoning_budget(
llama_sampler_apply(sampler, &cur_p);
// Check if forcing is active (all logits except one should be -INFINITY)
size_t not_neg_inf = 0;
llama_token not_neg_inf_token = -1;
size_t finite_count = 0;
llama_token finite_token = -1;
for (size_t j = 0; j < cur.size(); j++) {
if (std::isfinite(cur[j].logit) || cur[j].logit > 0) { // +INFINITY
not_neg_inf++;
not_neg_inf_token = cur[j].id;
if (std::isfinite(cur[j].logit)) {
finite_count++;
finite_token = cur[j].id;
}
}
llama_sampler_accept(sampler, sequence[i]);
fprintf(stderr, " i=%zu: token=%d, not_neg_inf_count=%zu, not_neg_inf_token=%d\n", i, (int)sequence[i], not_neg_inf, (int)not_neg_inf_token);
fprintf(stderr, " i=%zu: token=%d, finite_count=%zu, finite_token=%d\n", i, (int)sequence[i], finite_count, (int)finite_token);
if (not_neg_inf == 1) {
if (finite_count == 1) {
if (actual_force_start == SIZE_MAX) {
actual_force_start = i;
}