happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

context : do not reserve scheduler for warmups (#18867)

This commit is contained in:
Georgi Gerganov 2026-01-15 19:35:57 +02:00 committed by GitHub
parent 13f1e4a9ca
commit be8e3d9515
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 5 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
src/llama-context.cpp
View File

@ -337,7 +337,7 @@ llama_context::llama_context(
cparams.pipeline_parallel = pipeline_parallel;
if (cparams.pipeline_parallel) {
LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(sched.get()));
LLAMA_LOG_INFO("%s: pipeline parallelism enabled\n", __func__);
}
sched_reserve();
@ -537,7 +537,8 @@ void llama_context::sched_reserve() {
const int64_t t_end_us = ggml_time_us();
LLAMA_LOG_INFO("%s: reserve took %.2f ms\n", __func__, (t_end_us - t_start_us)/1000.0);
LLAMA_LOG_INFO("%s: reserve took %.2f ms, sched copies = %d\n",
__func__, (t_end_us - t_start_us)/1000.0, ggml_backend_sched_get_n_copies(sched.get()));
}
void llama_context::synchronize() {
@ -1011,7 +1012,8 @@ void llama_context::set_warmup(bool value) {
cparams.warmup = value;
sched_need_reserve = true;
// warmups are usually with small batches, so no need to reserve
//sched_need_reserve = true;
}
bool llama_context::set_sampler(llama_seq_id seq_id, llama_sampler * sampler) {