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

buffer incomplete multi-byte characters

Browse source
This commit is contained in:
anon 2023-05-31 10:40:42 -03:00
parent 27911d6d68
commit dd30219332
1 changed files with 35 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

37
examples/server/server.cpp
View file

@ -842,16 +842,49 @@ int main(int argc, char **argv)
"application/json"); "application/json");
} else { } else {
const auto chunked_content_provider = [&](size_t, DataSink &sink) { const auto chunked_content_provider = [&](size_t, DataSink &sink) {
size_t sent_count = 0;
int32_t multibyte_pending = 0;
while (llama.has_next_token) { while (llama.has_next_token) {
std::string token_text = llama.doCompletion(); std::string token_text = llama.doCompletion();
if (multibyte_pending > 0) {
multibyte_pending -= token_text.size();
} else if (token_text.size() == 1) {
const char c = token_text[0];
// 2-byte characters: 110xxxxx 10xxxxxx
if ((c & 0xE0) == 0xC0) {
multibyte_pending = 1;
// 3-byte characters: 1110xxxx 10xxxxxx 10xxxxxx
} else if ((c & 0xF0) == 0xE0) {
multibyte_pending = 2;
// 4-byte characters: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
} else if ((c & 0xF8) == 0xF0) {
multibyte_pending = 3;
} else {
multibyte_pending = 0;
}
}
if (multibyte_pending > 0) {
if (!llama.has_next_token) {
llama.has_next_token = true;
llama.n_remain++;
}
continue;
}
const size_t pos = std::min(sent_count, llama.generated_text.size());
std::string to_send = llama.generated_text.substr(pos);
sent_count += to_send.size();
json data; json data;
if (llama.has_next_token) { if (llama.has_next_token) {
data = {{"content", token_text}, {"stop", false}}; data = {{"content", to_send}, {"stop", false}};
} else { } else {
// Generation is done, send extra information. // Generation is done, send extra information.
data = { data = {
{"content", token_text}, {"content", to_send},
{"stop", true}, {"stop", true},
{"model", llama.params.model_alias}, {"model", llama.params.model_alias},
{"tokens_predicted", llama.num_tokens_predicted}, {"tokens_predicted", llama.num_tokens_predicted},