tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-28 04:15:21 +00:00
Code Issues Packages Projects Releases Wiki Activity

kv-cache : rework kv_idxs, support seq_cp

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2025-06-25 14:48:47 +03:00
parent 0bb1da5854
commit 6663128448
5 changed files with 229 additions and 98 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
examples/parallel/parallel.cpp
View File

@ -290,10 +290,8 @@ int main(int argc, char ** argv) {
for (int i = 1; i <= n_clients; ++i) {
llama_memory_seq_rm(mem, i, -1, -1);
if (is_sp_shared) {
// but keep the system prompt
llama_memory_seq_cp(mem, 0, i, -1, -1);
}
// but keep the system prompt
llama_memory_seq_cp(mem, 0, i, -1, -1);
}
LOG_INF("%s: clearing the KV cache\n", __func__);
@ -452,11 +450,8 @@ int main(int argc, char ** argv) {
}
// delete only the generated part of the sequence, i.e. keep the system prompt in the cache
llama_memory_seq_rm(mem, client.id + 1, -1, -1);
if (is_sp_shared) {
llama_memory_seq_cp(mem, 0, client.id + 1, -1, -1);
}
llama_memory_seq_rm(mem, client.id + 1, -1, -1);
llama_memory_seq_cp(mem, 0, client.id + 1, -1, -1);
const auto t_main_end = ggml_time_us();