Skip to content

Commit

Permalink
server : fix handling of characters that span multiple tokens when st…
Browse files Browse the repository at this point in the history
…reaming (ggerganov#4446)
  • Loading branch information
shibe2 authored and teleprint-me committed Dec 21, 2023
1 parent ec3f263 commit 84c0489
Showing 1 changed file with 19 additions and 20 deletions.
39 changes: 19 additions & 20 deletions examples/server/server.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -376,7 +376,6 @@ struct llama_client_slot

int32_t num_prompt_tokens = 0;
int32_t num_prompt_tokens_processed = 0;
int32_t multibyte_pending = 0;

json prompt;
std::string generated_text;
Expand Down Expand Up @@ -425,7 +424,6 @@ struct llama_client_slot
stopped_word = false;
stopped_limit = false;
stopping_word = "";
multibyte_pending = 0;
n_past = 0;
sent_count = 0;
sent_token_probs_index = 0;
Expand Down Expand Up @@ -992,35 +990,36 @@ struct llama_server_context
slot.generated_text += token_str;
slot.has_next_token = true;

if (slot.multibyte_pending > 0)
// check if there is incomplete UTF-8 character at the end
bool incomplete = false;
for (unsigned i = 1; i < 5 && i <= slot.generated_text.size(); ++i)
{
slot.multibyte_pending -= token_str.size();
}
else if (token_str.size() == 1)
{
const char c = token_str[0];
// 2-byte characters: 110xxxxx 10xxxxxx
unsigned char c = slot.generated_text[slot.generated_text.size() - i];
if ((c & 0xC0) == 0x80)
{
// continuation byte: 10xxxxxx
continue;
}
if ((c & 0xE0) == 0xC0)
{
slot.multibyte_pending = 1;
// 3-byte characters: 1110xxxx 10xxxxxx 10xxxxxx
// 2-byte character: 110xxxxx ...
incomplete = i < 2;
}
else if ((c & 0xF0) == 0xE0)
{
slot.multibyte_pending = 2;
// 4-byte characters: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
// 3-byte character: 1110xxxx ...
incomplete = i < 3;
}
else if ((c & 0xF8) == 0xF0)
{
slot.multibyte_pending = 3;
}
else
{
slot.multibyte_pending = 0;
// 4-byte character: 11110xxx ...
incomplete = i < 4;
}
// else 1-byte character or invalid byte
break;
}

if (slot.multibyte_pending == 0)
if (!incomplete)
{
size_t pos = std::min(slot.sent_count, slot.generated_text.size());
const std::string str_test = slot.generated_text.substr(pos);
Expand Down Expand Up @@ -1055,7 +1054,7 @@ struct llama_server_context
}
}

if (slot.multibyte_pending > 0 && !slot.has_next_token)
if (incomplete)
{
slot.has_next_token = true;
}
Expand Down

0 comments on commit 84c0489

Please sign in to comment.