mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 28 Packages Projects Releases Wiki Activity

common : inhibit lazy grammar sampler while reasoning is active (#20970)

Browse Source 
* common : inhibit grammar while reasoning budget is active

* cont : update force_pos in accept

* cont : fix tests

* cont : tweak should apply logic

* cont : return early not using grammar sampler

* Add tests

* cont : prevent backend sampling when reasoning budget enabled

* cont : fix typo

---------

Co-authored-by: Piotr Wilkin <piotr.wilkin@syndatis.com>
This commit is contained in:
Aldehir Rojas
2026-03-27 12:30:40 -05:00
committed by GitHub
parent ff934e29bc
commit 59d840209a
8 changed files with 295 additions and 106 deletions
Download Patch File Download Diff File Expand all files Collapse all files
common/reasoning-budget.cpp
+12 -11
View File
@@ -115,9 +115,11 @@ static void common_reasoning_budget_accept(struct llama_sampler * smpl, llama_to
break;
}
case REASONING_BUDGET_FORCING:
// force_pos is advanced in apply(), not here.
// This ensures the first forced token isn't skipped when the sampler
// is initialized directly in FORCING state (e.g. COUNTING + budget=0)
ctx->force_pos++;
if (ctx->force_pos >= ctx->forced_tokens.size()) {
ctx->state = REASONING_BUDGET_DONE;
LOG_INF("reasoning-budget: forced sequence complete, done\n");
}
break;
case REASONING_BUDGET_DONE:
break;
@@ -144,14 +146,6 @@ static void common_reasoning_budget_apply(struct llama_sampler * smpl, llama_tok
cur_p->data[i].logit = -INFINITY;
}
}
// advance to next forced token (done here rather than in accept so that
// the first forced token isn't skipped when starting in FORCING state)
ctx->force_pos++;
if (ctx->force_pos >= ctx->forced_tokens.size()) {
ctx->state = REASONING_BUDGET_DONE;
LOG_INF("reasoning-budget: forced sequence complete, done\n");
}
}
static void common_reasoning_budget_reset(struct llama_sampler * smpl) {
@@ -261,3 +255,10 @@ struct llama_sampler * common_reasoning_budget_init(
common_reasoning_budget_state initial_state) {
return common_reasoning_budget_init_state(vocab, start_tokens, end_tokens, forced_tokens, budget, initial_state);
}
common_reasoning_budget_state common_reasoning_budget_get_state(const struct llama_sampler * smpl) {
if (!smpl) {
return REASONING_BUDGET_IDLE;
}
return ((const common_reasoning_budget_ctx *)smpl->ctx)->state;
}