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

examples : replace llama_kv_cache_seq_* with llama_past_seq_*

Browse source
This commit is contained in:
Francis Couture-Harpin 2024-06-10 14:44:42 -04:00
parent 372482dffe
commit 43d8d4bf9e
23 changed files with 125 additions and 112 deletions
Download patch file Download diff file Expand all files Collapse all files

13
examples/lookahead/lookahead.cpp
View file

@ -96,7 +96,7 @@ int main(int argc, char ** argv) {
llama_decode(ctx, llama_batch_get_one(&inp.back(), 1, n_input - 1, 0));
for (int s = 1; s < W + G + 1; ++s) {
llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
llama_past_seq_cp(ctx, 0, s, -1, -1);
}
const auto t_enc_end = ggml_time_us();
@ -438,17 +438,18 @@ 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(ctx, -1, n_past, -1);
// FIXME: recurrent and hybrid models
llama_past_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(ctx, seq_id_best);
llama_kv_cache_seq_cp (ctx, seq_id_best, 0, -1, -1);
llama_kv_cache_seq_rm (ctx, seq_id_best, -1, -1);
llama_past_seq_keep(ctx, seq_id_best);
llama_past_seq_cp (ctx, seq_id_best, 0, -1, -1);
llama_past_seq_rm (ctx, seq_id_best, -1, -1);
for (int s = 1; s < W + G + 1; ++s) {
llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
llama_past_seq_cp(ctx, 0, s, -1, -1);
}
}
}