tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-07-01 21:15:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama : cont

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2025-01-13 14:56:52 +02:00
parent f78b396ee7
commit e4550fbafc
19 changed files with 128 additions and 79 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
examples/lookahead/lookahead.cpp
View File

@ -60,6 +60,7 @@ 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);
@ -95,7 +96,7 @@ 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(ctx, 0, s, -1, -1);
llama_kv_cache_seq_cp(kv, 0, s, -1, -1);
}
const auto t_enc_end = ggml_time_us();
@ -437,17 +438,17 @@ 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(ctx, -1, n_past, -1);
llama_kv_cache_seq_rm(kv, -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_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);
for (int s = 1; s < W + G + 1; ++s) {
llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
llama_kv_cache_seq_cp(kv, 0, s, -1, -1);
}
}
}