CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-21 12:16:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

common : more accurate sampling timing (#17382)

Browse Source 
* common : more accurate sampling timing

* eval-callback : minor fixes

* cont : add time_meas impl

* cont : fix log msg [no ci]

* cont : fix multiple definitions of time_meas

* llama-cli : exclude chat template init from time measurement

* cont : print percentage of unaccounted time

* cont : do not reset timings
This commit is contained in:
Georgi Gerganov
2025-11-20 13:40:10 +02:00
committed by GitHub
parent 5088b435d4
commit 196f5083ef
7 changed files with 102 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
src/llama-impl.cpp
View File

@@ -20,10 +20,10 @@ static llama_logger_state g_logger_state;
time_meas::time_meas(int64_t & t_acc, bool disable) : t_start_us(disable ? -1 : ggml_time_us()), t_acc(t_acc) {}
time_meas::~time_meas() {
if (t_start_us >= 0) {
t_acc += ggml_time_us() - t_start_us;
}
if (t_start_us >= 0) {
t_acc += ggml_time_us() - t_start_us;
}
}
void llama_log_set(ggml_log_callback log_callback, void * user_data) {
ggml_log_set(log_callback, user_data);

9
src/llama-sampling.cpp
View File

@@ -472,9 +472,6 @@ static void llama_sampler_chain_reset(struct llama_sampler * smpl) {
for (auto * smpl : chain->samplers) {
llama_sampler_reset(smpl);
}
chain->t_sample_us = 0;
chain->n_sample = 0;
}
static struct llama_sampler * llama_sampler_chain_clone(const struct llama_sampler * smpl) {
@@ -2670,8 +2667,7 @@ struct llama_perf_sampler_data llama_perf_sampler(const struct llama_sampler * c
void llama_perf_sampler_print(const struct llama_sampler * chain) {
const auto data = llama_perf_sampler(chain);
LLAMA_LOG_INFO("%s: sampling time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n",
__func__, data.t_sample_ms, data.n_sample, data.t_sample_ms / data.n_sample, 1e3 / data.t_sample_ms * data.n_sample);
LLAMA_LOG_INFO("%s: samplers time = %10.2f ms / %5d runs\n", __func__, data.t_sample_ms, data.n_sample);
}
void llama_perf_sampler_reset(struct llama_sampler * chain) {
@@ -2681,5 +2677,6 @@ void llama_perf_sampler_reset(struct llama_sampler * chain) {
auto * ctx = (struct llama_sampler_chain *) chain->ctx;
ctx->t_sample_us = ctx->n_sample = 0;
ctx->t_sample_us = 0;
ctx->n_sample = 0;
}