CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-08 10:07:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama : remove implicit recurrent state rollbacks

Browse Source
This commit is contained in:
Francis Couture-Harpin
2024-11-24 20:35:30 -05:00
parent 124c222f76
commit 8006f3b3c8
25 changed files with 411 additions and 1119 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
examples/parallel/parallel.cpp
View File

@@ -199,7 +199,7 @@ 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_past_seq_cp(ctx, 0, i, -1, -1);
llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
}
LOG_INF("\n");
@@ -231,9 +231,9 @@ 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_past_seq_rm(ctx, i, -1, -1);
llama_kv_cache_seq_rm(ctx, i, -1, -1);
// but keep the system prompt
llama_past_seq_cp(ctx, 0, i, -1, -1);
llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
}
LOG_INF("%s: clearing the KV cache\n", __func__);
@@ -370,8 +370,8 @@ int main(int argc, char ** argv) {
}
// delete only the generated part of the sequence, i.e. keep the system prompt in the cache
llama_past_seq_rm(ctx, client.id + 1, -1, -1);
llama_past_seq_cp(ctx, 0, client.id + 1, -1, -1);
llama_kv_cache_seq_rm(ctx, client.id + 1, -1, -1);
llama_kv_cache_seq_cp(ctx, 0, client.id + 1, -1, -1);
const auto t_main_end = ggml_time_us();