vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : update llama_kv_self API

Browse source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2025-01-14 16:47:34 +02:00
parent fd05ab87aa
commit 17b363afd3
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
30 changed files with 387 additions and 205 deletions
Download patch file Download diff file Expand all files Collapse all files

11
examples/parallel/parallel.cpp
View file

@ -134,7 +134,6 @@ int main(int argc, char ** argv) {
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
llama_kv_cache * kv = llama_get_kv_cache(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
@ -202,7 +201,7 @@ int main(int argc, char ** argv) {
// assign the system KV cache to all parallel sequences
for (int32_t i = 1; i <= n_clients; ++i) {
llama_kv_cache_seq_cp(kv, 0, i, -1, -1);
llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
}
LOG_INF("\n");
@ -234,9 +233,9 @@ int main(int argc, char ** argv) {
if (batch.n_tokens == 0) {
// all sequences have ended - clear the entire KV cache
for (int i = 1; i <= n_clients; ++i) {
llama_kv_cache_seq_rm(kv, i, -1, -1);
llama_kv_self_seq_rm(ctx, i, -1, -1);
// but keep the system prompt
llama_kv_cache_seq_cp(kv, 0, i, -1, -1);
llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
}
LOG_INF("%s: clearing the KV cache\n", __func__);
@ -372,8 +371,8 @@ int main(int argc, char ** argv) {
}
// delete only the generated part of the sequence, i.e. keep the system prompt in the cache
llama_kv_cache_seq_rm(kv, client.id + 1, -1, -1);
llama_kv_cache_seq_cp(kv, 0, client.id + 1, -1, -1);
llama_kv_self_seq_rm(ctx, client.id + 1, -1, -1);
llama_kv_self_seq_cp(ctx, 0, client.id + 1, -1, -1);
const auto t_main_end = ggml_time_us();