main : support special tokens as reverse/anti prompt (#5847)

* Support special tokens as reverse/anti prompt.

* Tokenize antiprompts only once.

* main : minor

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
DAN™ 2024-03-04 02:57:20 -05:00 committed by GitHub
parent 67be2ce101
commit 5a51cc1bb4
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194

View File

@ -511,6 +511,14 @@ int main(int argc, char ** argv) {
std::vector<llama_token> embd; std::vector<llama_token> embd;
std::vector<llama_token> embd_guidance; std::vector<llama_token> embd_guidance;
// tokenized antiprompts
std::vector<std::vector<llama_token>> antiprompt_ids;
antiprompt_ids.reserve(params.antiprompt.size());
for (const std::string & antiprompt : params.antiprompt) {
antiprompt_ids.emplace_back(::llama_tokenize(ctx, antiprompt, false, true));
}
struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams); struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
while ((n_remain != 0 && !is_antiprompt) || params.interactive) { while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
@ -769,6 +777,18 @@ int main(int argc, char ** argv) {
} }
} }
// check for reverse prompt using special tokens
llama_token last_token = llama_sampling_last(ctx_sampling);
for (std::vector<llama_token> ids : antiprompt_ids) {
if (ids.size() == 1 && last_token == ids[0]) {
if (params.interactive) {
is_interacting = true;
}
is_antiprompt = true;
break;
}
}
if (is_antiprompt) { if (is_antiprompt) {
LOG("found antiprompt: %s\n", last_output.c_str()); LOG("found antiprompt: %s\n", last_output.c_str());
} }