root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 19:04:35 +00:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

examples : do not assume BOS when shifting context (#5622)

Browse Source
This commit is contained in:
Jared Van Bortel 2024-02-21 10:33:54 -05:00 committed by GitHub
parent 5022cf242d
commit 89febfed93
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 14 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
examples/main/main.cpp
View File

@ -334,6 +334,8 @@ int main(int argc, char ** argv) {
// number of tokens to keep when resetting context // number of tokens to keep when resetting context
if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct || params.chatml) { if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct || params.chatml) {
params.n_keep = (int)embd_inp.size(); params.n_keep = (int)embd_inp.size();
} else {
params.n_keep += add_bos; // always keep the BOS token
} }
// prefix & suffix for instruct mode // prefix & suffix for instruct mode
@ -383,7 +385,7 @@ int main(int argc, char ** argv) {
} }
} }
if (params.n_keep > 0) { if (params.n_keep > add_bos) {
LOG_TEE("%s: static prompt based on n_keep: '", __func__); LOG_TEE("%s: static prompt based on n_keep: '", __func__);
for (int i = 0; i < params.n_keep; i++) { for (int i = 0; i < params.n_keep; i++) {
LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str()); LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
@ -540,14 +542,14 @@ int main(int argc, char ** argv) {
break; break;
} }
const int n_left = n_past - params.n_keep - 1; const int n_left = n_past - params.n_keep;
const int n_discard = n_left/2; const int n_discard = n_left/2;
LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n", LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
n_past, n_left, n_ctx, params.n_keep, n_discard); n_past, n_left, n_ctx, params.n_keep, n_discard);
llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1 , params.n_keep + n_discard + 1); llama_kv_cache_seq_rm (ctx, 0, params.n_keep , params.n_keep + n_discard);
llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard); llama_kv_cache_seq_shift(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
n_past -= n_discard; n_past -= n_discard;

13
examples/server/server.cpp
View File

@ -1487,14 +1487,15 @@ struct llama_server_context
if (slot.is_processing() && system_tokens.size() + slot.cache_tokens.size() >= (size_t) slot.n_ctx) if (slot.is_processing() && system_tokens.size() + slot.cache_tokens.size() >= (size_t) slot.n_ctx)
{ {
// Shift context // Shift context
const int n_left = system_tokens.size() + slot.n_past - slot.params.n_keep - 1; const int n_keep = slot.params.n_keep + add_bos_token;
const int n_left = system_tokens.size() + slot.n_past - n_keep;
const int n_discard = n_left / 2; const int n_discard = n_left / 2;
LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, slot.params.n_keep, n_left, n_discard); LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, n_keep, n_left, n_discard);
llama_kv_cache_seq_rm (ctx, slot.id, slot.params.n_keep + 1 , slot.params.n_keep + n_discard + 1); llama_kv_cache_seq_rm (ctx, slot.id, n_keep , n_keep + n_discard);
llama_kv_cache_seq_shift(ctx, slot.id, slot.params.n_keep + 1 + n_discard, system_tokens.size() + slot.n_past, -n_discard); llama_kv_cache_seq_shift(ctx, slot.id, n_keep + n_discard, system_tokens.size() + slot.n_past, -n_discard);
for (size_t i = slot.params.n_keep + 1 + n_discard; i < slot.cache_tokens.size(); i++) for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++)
{ {
slot.cache_tokens[i - n_discard] = slot.cache_tokens[i]; slot.cache_tokens[i - n_discard] = slot.cache_tokens[i];
} }
@ -1507,7 +1508,7 @@ struct llama_server_context
LOG_VERBOSE("context shift", { LOG_VERBOSE("context shift", {
{ "n_ctx", n_ctx }, { "n_ctx", n_ctx },
{ "n_keep", params.n_keep }, { "n_keep", n_keep },
{ "n_left", n_left }, { "n_left", n_left },
}); });
} }