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

13
examples/lookahead/lookahead.cpp
View file

@ -60,7 +60,6 @@ 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);
@ -96,7 +95,7 @@ int main(int argc, char ** argv) {
llama_decode(ctx, llama_batch_get_one(&inp.back(), 1));
for (int s = 1; s < W + G + 1; ++s) {
llama_kv_cache_seq_cp(kv, 0, s, -1, -1);
llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
}
const auto t_enc_end = ggml_time_us();
@ -438,17 +437,17 @@ int main(int argc, char ** argv) {
// KV cache management
// if no verification token matched, we simply remove all cells from this batch -> no fragmentation
llama_kv_cache_seq_rm(kv, -1, n_past, -1);
llama_kv_self_seq_rm(ctx, -1, n_past, -1);
if (seq_id_best != 0) {
// if a verification token matched, we keep the best sequence and remove the rest
// this leads to some KV cache fragmentation
llama_kv_cache_seq_keep(kv, seq_id_best);
llama_kv_cache_seq_cp (kv, seq_id_best, 0, -1, -1);
llama_kv_cache_seq_rm (kv, seq_id_best, -1, -1);
llama_kv_self_seq_keep(ctx, seq_id_best);
llama_kv_self_seq_cp (ctx, seq_id_best, 0, -1, -1);
llama_kv_self_seq_rm (ctx, seq_id_best, -1, -1);
for (int s = 1; s < W + G + 1; ++s) {
llama_kv_cache_seq_cp(kv, 0, s, -1, -1);
llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
}
}
}