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

Add heuristic algo for speculative

Browse source
This commit is contained in:
Lengyue 2023-09-04 04:51:30 -04:00
parent 35195689cd
commit 98230ef656
No known key found for this signature in database
GPG key ID: 67CD446B0133EF95
1 changed files with 14 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

15
examples/speculative/speculative.cpp
View file

@ -84,7 +84,7 @@ int main(int argc, char ** argv) {
//GGML_ASSERT(n_vocab == llama_n_vocab(ctx_dft));
// how many tokens to draft each time
const int n_draft = params.n_draft;
int n_draft = params.n_draft;
int n_predict = 0;
int n_drafted = 0;
@ -116,6 +116,8 @@ int main(int argc, char ** argv) {
// sample from the drafted tokens if any
int i_dft = 0;
bool all_accepted = false;
while (true) {
const llama_token id = llama_sample_token(ctx_tgt, NULL, NULL, params, last_tokens, candidates, i_dft);
@ -141,6 +143,9 @@ int main(int argc, char ** argv) {
++n_past_dft;
++i_dft;
if (i_dft == (int) drafted.size()) {
all_accepted = true;
}
continue;
}
@ -154,6 +159,14 @@ int main(int argc, char ** argv) {
break;
}
if (drafted.size() > 0 && all_accepted) {
n_draft += 2;
LOG("all drafted tokens accepted, n_draft = %d\n", n_draft);
} else {
n_draft -= 1;
LOG("drafted token rejected, n_draft = %d\n", n_draft);
}
if (n_predict > params.n_predict || has_eos) {
break;
}