ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-26 17:20:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

Correctly accumulate sampling time for adaptive_p

Browse Source
This commit is contained in:
Kawrakow
2026-01-19 10:17:07 +00:00
parent 61eccfcf0d
commit a96e5449cc
5 changed files with 20 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
common/sampling.cpp
View File

@@ -471,7 +471,7 @@ static llama_token llama_sampling_sample_impl(
id = llama_sample_token_mirostat_v2(ctx_main, &cur_p, mirostat_tau, mirostat_eta, &ctx_sampling->mirostat_mu);
} else if (adaptive_target >= 0.0f && ctx_sampling->adapt_p_ctx!=nullptr) {
// adaptive p sampling
llama_prep_adaptive_p(&cur_p, ctx_sampling->adapt_p_ctx);
llama_prep_adaptive_p(ctx_main, &cur_p, ctx_sampling->adapt_p_ctx);
sampler_queue(ctx_main, params, ctx_sampling, cur_p, std::max(1, params.min_keep));
id = llama_sample_token_adaptive_p(ctx_main, &cur_p, ctx_sampling->adapt_p_ctx);
} else {

2
include/llama.h
View File

@@ -1389,7 +1389,7 @@ LLAMA_API struct llama_grammar* llama_sampler_init_grammar_lazy_patterns(
const float decay,
const uint32_t seed);
void llama_prep_adaptive_p(
void llama_prep_adaptive_p(struct llama_context * ctx,
llama_token_data_array * candidates,
struct llama_sampler_adaptive_p * adapt_p_ctx);

18
src/llama-sampling.cpp
View File

@@ -1061,15 +1061,15 @@ llama_token llama_sample_token_adaptive_p_impl(
const size_t idx = std::distance(ctx->cum_probs.begin(), iter);
llama_token id = candidates->data[idx].id;
smpl->t_sample_us += ggml_time_us() - t_start_sample_us;
smpl->n_sample++;
if (auto it = ctx->orig_prob_map.find(id); it != ctx->orig_prob_map.end()) {
float update_prob = it->second / ctx->cum_orig_prob;
ctx->weighted_sum = ctx->decay * ctx->weighted_sum + update_prob;
ctx->total_weight = ctx->decay * ctx->total_weight + 1.0f;
}
smpl->t_sample_us += ggml_time_us() - t_start_sample_us;
smpl->n_sample++;
//float update_prob = candidates->data[idx].p; // not ideal
//if (ctx->orig_prob_map.contains(id)) {
// // selected token id is among tracked ids
@@ -1083,13 +1083,16 @@ llama_token llama_sample_token_adaptive_p_impl(
return id;
}
void llama_sample_adaptive_p_impl(llama_token_data_array * candidates, struct llama_sampler_adaptive_p * adapt_p_ctx) {
void llama_sample_adaptive_p_impl(struct llama_sampling * ctx, llama_token_data_array * candidates,
struct llama_sampler_adaptive_p * adapt_p_ctx) {
if (adapt_p_ctx->target < 0.0f) {
// sampler is disabled
llama_sample_softmax_impl(nullptr, candidates);
return;
}
auto t_start = ggml_time_us();
// incomplete softmax because final division can be fused
float max_l = candidates->data[0].logit;
if (!candidates->sorted) {
@@ -1130,12 +1133,15 @@ void llama_sample_adaptive_p_impl(llama_token_data_array * candidates, struct ll
}
candidates->sorted = false;
adapt_p_ctx->max_xform_logit = max_logit;
ctx->t_sample_us += ggml_time_us() - t_start;
}
void llama_prep_adaptive_p_impl(
void llama_prep_adaptive_p_impl(struct llama_sampling * smpl,
llama_token_data_array * candidates,
struct llama_sampler_adaptive_p * adapt_p_ctx) {
constexpr float kDelta = 16.6f;
auto t_start = ggml_time_us();
if (!candidates->sorted) {
float max_logit = candidates->data[0].logit;
for (int j = 1; j < int(candidates->size); ++j) {
@@ -1152,6 +1158,7 @@ void llama_prep_adaptive_p_impl(
}
}
adapt_p_ctx->cum_orig_prob = cum_prob;
if (smpl) smpl->t_sample_us += ggml_time_us() - t_start;
return;
}
@@ -1169,6 +1176,7 @@ void llama_prep_adaptive_p_impl(
adapt_p_ctx->orig_prob_map[candidates->data[j].id] = prob;
}
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,

2
src/llama-sampling.h
View File

@@ -89,10 +89,12 @@ struct llama_sampler_adaptive_p * llama_init_adaptive_p_impl(
const uint32_t seed);
void llama_prep_adaptive_p_impl(
struct llama_sampling * smpl,
llama_token_data_array * candidates,
struct llama_sampler_adaptive_p * adapt_p_ctx);
void llama_sample_adaptive_p_impl(
struct llama_sampling * smpl,
llama_token_data_array * candidates,
struct llama_sampler_adaptive_p * adapt_p_ctx);

6
src/llama.cpp
View File

@@ -7691,11 +7691,11 @@ void llama_sample_adaptive_p(
[[maybe_unused]] struct llama_context * ctx,
llama_token_data_array * candidates,
struct llama_sampler_adaptive_p * adapt_p_ctx) {
llama_sample_adaptive_p_impl(candidates, adapt_p_ctx);
llama_sample_adaptive_p_impl(&ctx->sampling, candidates, adapt_p_ctx);
}
void llama_prep_adaptive_p(llama_token_data_array * candidates, struct llama_sampler_adaptive_p * adapt_p_ctx) {
llama_prep_adaptive_p_impl(candidates, adapt_p_ctx);
void llama_prep_adaptive_p(struct llama_context * ctx, llama_token_data_array * candidates, struct llama_sampler_adaptive_p * adapt_p_ctx) {
llama_prep_adaptive_p_impl(&ctx->sampling, candidates, adapt_p_ctx);
}