nicolas
/
ik_llama_opt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Add adaptive sampling clone and free functions to manage memory (#1851)

This commit is contained in:
Samuel Oliveira Alves 2026-05-21 02:11:17 -03:00 committed by GitHub
parent aefb8bdd99
commit 7b73f45541
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 35 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
common/sampling.cpp
View File

@ -12,6 +12,9 @@
#include <nlohmann/json.hpp>
using json = nlohmann::ordered_json;
struct llama_sampler_adaptive_p * llama_clone_adaptive_p(const struct llama_sampler_adaptive_p * adapt_p_ctx);
void llama_free_adaptive_p(struct llama_sampler_adaptive_p * adapt_p_ctx);
struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params) {
const llama_vocab * vocab = llama_model_get_vocab(model);
@ -160,9 +163,11 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
}
case llama_sampler_type::ADAPTIVE_P:
{
GGML_ASSERT(vocab);
auto n_vocab = llama_vocab_n_tokens(vocab);
result->adapt_p_ctx = llama_init_adaptive_p(n_vocab, params.adaptive_target, params.adaptive_decay, params.adaptive_updt_w_cur, result->rng());
if (params.adaptive_target >= 0.0f) {
GGML_ASSERT(vocab);
auto n_vocab = llama_vocab_n_tokens(vocab);
result->adapt_p_ctx = llama_init_adaptive_p(n_vocab, params.adaptive_target, params.adaptive_decay, params.adaptive_updt_w_cur, result->rng());
}
break;
}
default:
@ -185,6 +190,8 @@ void common_sampler_free(struct common_sampler * ctx) {
}
if (ctx->smpl)
llama_sampler_dry_free(ctx->smpl);
if (ctx->adapt_p_ctx)
llama_free_adaptive_p(ctx->adapt_p_ctx);
if (ctx->rbudget)
common_reasoning_budget_free(ctx->rbudget);
delete ctx;
@ -255,6 +262,14 @@ void common_sampler_clone(common_sampler * src, common_sampler * dst) {
dst->smpl = llama_sampler_dry_clone(src->smpl);
}
if (dst->adapt_p_ctx) {
llama_free_adaptive_p(dst->adapt_p_ctx);
dst->adapt_p_ctx = nullptr;
}
if (src->adapt_p_ctx) {
dst->adapt_p_ctx = llama_clone_adaptive_p(src->adapt_p_ctx);
}
if (dst->rbudget) {
common_reasoning_budget_free(dst->rbudget);
dst->rbudget = nullptr;
@ -454,7 +469,7 @@ static void sampler_queue(
llama_sample_temp(ctx_main, &cur_p, temp);
}
break;
case llama_sampler_type::ADAPTIVE_P: use_adaptive_p = true; break;
case llama_sampler_type::ADAPTIVE_P: use_adaptive_p = ctx_sampling->adapt_p_ctx != nullptr; break;
default : break;
}

4
src/llama-sampling.h
View File

@ -90,6 +90,10 @@ struct llama_sampler_adaptive_p * llama_init_adaptive_p_impl(int n_vocab,
const bool updt_w_cur,
const uint32_t seed);
struct llama_sampler_adaptive_p * llama_clone_adaptive_p(const struct llama_sampler_adaptive_p * adapt_p_ctx);
void llama_free_adaptive_p(struct llama_sampler_adaptive_p * adapt_p_ctx);
void llama_prep_adaptive_p_impl(
struct llama_sampling * smpl,
llama_token_data_array * candidates,

12
src/llama.cpp
View File

@ -10533,6 +10533,18 @@ struct llama_sampler_adaptive_p * llama_init_adaptive_p(int n_vocab, const float
return llama_init_adaptive_p_impl(n_vocab, target, decay, updt_w_cur, seed);
}
struct llama_sampler_adaptive_p * llama_clone_adaptive_p(const struct llama_sampler_adaptive_p * adapt_p_ctx) {
if (adapt_p_ctx == nullptr) {
return nullptr;
}
return new llama_sampler_adaptive_p(*adapt_p_ctx);
}
void llama_free_adaptive_p(struct llama_sampler_adaptive_p * adapt_p_ctx) {
delete adapt_p_ctx;
}
void llama_review_adaptive_p(struct llama_sampler_adaptive_p * adapt_p_ctx, const size_t n_unsent, const bool rewind_status) {
llama_review_adaptive_p_impl(adapt_p_ctx, n_unsent, rewind_status);
}