CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-13 10:57:15 +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

28
examples/passkey/passkey.cpp
View File

@@ -130,11 +130,11 @@ int main(int argc, char ** argv) {
const int ib = i/n_batch - 1;
const int bd = n_batch_grp*(n_grp - 1);
llama_past_seq_add (ctx, 0, n_past - n_batch, n_past, ib*bd);
llama_past_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
llama_kv_cache_update(ctx);
llama_kv_cache_seq_add (ctx, 0, n_past - n_batch, n_past, ib*bd);
llama_kv_cache_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
llama_kv_cache_update (ctx);
n_past = llama_past_seq_pos_max(ctx, 0) + 1;
n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
}
common_batch_clear(batch);
@@ -164,12 +164,12 @@ int main(int argc, char ** argv) {
LOG_INF("%s: shifting KV cache with %d\n", __func__, n_discard);
llama_past_seq_rm (ctx, 0, n_keep , n_keep + n_discard);
llama_past_seq_add (ctx, 0, n_keep + n_discard, n_ctx, -n_discard);
//llama_kv_cache_defrag(ctx);
llama_kv_cache_update(ctx);
llama_kv_cache_seq_rm (ctx, 0, n_keep , n_keep + n_discard);
llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard);
//llama_kv_cache_defrag (ctx);
llama_kv_cache_update (ctx);
n_past = llama_past_seq_pos_max(ctx, 0) + 1;
n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
common_batch_clear(batch);
@@ -195,12 +195,12 @@ int main(int argc, char ** argv) {
if (n_discard > 0) {
LOG_INF("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
llama_past_seq_rm (ctx, 0, n_keep , n_keep + n_discard);
llama_past_seq_add (ctx, 0, n_keep + n_discard, n_ctx, -n_discard);
//llama_kv_cache_defrag(ctx);
llama_kv_cache_update(ctx);
llama_kv_cache_seq_rm (ctx, 0, n_keep , n_keep + n_discard);
llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard);
//llama_kv_cache_defrag (ctx);
llama_kv_cache_update (ctx);
n_past = llama_past_seq_pos_max(ctx, 0) + 1;
n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
}
}