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

speculative : be more rewarding upon guessing max drafted tokens

Browse source
This commit is contained in:
Georgi Gerganov 2023-09-05 08:54:08 +03:00
parent 4e6e951399
commit d9559b78f3
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
1 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

4
examples/speculative/speculative.cpp
View file

@ -189,8 +189,8 @@ int main(int argc, char ** argv) {
LOG("all_accepted = %d\n", all_accepted);
if (all_accepted && n_draft == n_dradt_cur) {
LOG(" - max drafted tokens accepted - n_draft += 2\n");
n_draft += 2;
LOG(" - max drafted tokens accepted - n_draft += 8\n");
n_draft = std::min(30, n_draft + 8);
} else if (all_accepted) {
LOG(" - partially drafted tokens accepted - no change\n");
} else {