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

call a standalone function to untokenize output

Browse source
This commit is contained in:
wizard 2023-03-13 10:00:34 +08:00
parent 307dba3dd2
commit 1b87fe1e90
3 changed files with 72 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

7
main.cpp
View file

@ -947,9 +947,10 @@ int main(int argc, char ** argv) {
// display text // display text
if (!input_noecho) { if (!input_noecho) {
for (auto id : embd) { untokenize(sp, embd);
printf("%s", vocab.id_to_token[id].c_str()); // for (auto id : embd) {
} // printf("%s", vocab.id_to_token[id].c_str());
// }
// reset color to default if we there is no pending user input // reset color to default if we there is no pending user input
if (params.use_color && embd_inp.size() <= input_consumed) { if (params.use_color && embd_inp.size() <= input_consumed) {
printf(ANSI_COLOR_RESET); printf(ANSI_COLOR_RESET);

63
utils.cpp
View file

@ -5,6 +5,7 @@
#include <fstream> #include <fstream>
#include <regex> #include <regex>
#include <sentencepiece_processor.h> #include <sentencepiece_processor.h>
#include <sstream>
#include <iostream> #include <iostream>
#include <iterator> #include <iterator>
#include <string> #include <string>
@ -540,3 +541,65 @@ size_t ggml_quantize_q4_1(float * src, void * dst, int n, int k, int qk, int64_t
return (n/k)*row_size; return (n/k)*row_size;
} }
void untokenize(sentencepiece::SentencePieceProcessor & sp, std::vector<gpt_vocab::id> & embd)
{
// Convert the IDs in embd to tokens using SentencePiece
// std::vector<gpt_vocab::id> pieces;
// for (const auto& id : embd) {
// //std::string s = sp.DecodeIds(id);
// //s = std::regex_replace(s, std::regex("▁"), " ");
// // if (s.find("<0x") == 0 && s[s.length() - 1] == '>')
// // {
// // s = sp.IdToPiece(id);
// // }
// //printf("%s", s.c_str());
// pieces.push_back(id);
// // if(s.length() > 1)
// // tokens.push_back(" ");
// }
// // Insert spaces between tokens
// // std::string text;
// // for (const auto& token : tokens) {
// // // Add a space before the token if it is not the first token and it doesn't start with a special character
// // if (!text.empty() && !(token[0] == '\0x25' && token[1] == '\0x81') && token[0] != ' ') {
// // text += ' ';
// // }
// // text += sp.DecodePieces(tokens);
// // }
// //sp.DecodeIds(embd);
// std::string text =
// sp.DecodeIds(pieces);
// printf("%s", text.c_str());
for (auto id : embd) {
std::string s = sp.IdToPiece(id); //vocab.id_to_token[id];
if (s.find("<0x") == 0 && s[s.length() - 1] == '>')
{
// Extract the hexadecimal value from the token
std::string hex_value = s.substr(s.find("0x"));
// Convert the hexadecimal value to binary and print it
int decimal_value;
std::stringstream(hex_value) >> std::hex >> decimal_value;
std::bitset<8> binary_value(decimal_value);
char* bytes = reinterpret_cast<char*>(&decimal_value);
printf("%s", bytes);
}
else if(s.find("▁") == 0)
{
s = std::regex_replace(s, std::regex("▁"), " ");
//s.replace(0, 2, 1, ' ');
printf("%s", s.c_str());
}
else
{
printf("%s", s.c_str());
}
}
}

5
utils.h
View file

@ -7,6 +7,8 @@
#include <vector> #include <vector>
#include <random> #include <random>
#include <thread> #include <thread>
#include <sentencepiece_processor.h>
#include <sstream>
// //
// CLI argument parsing // CLI argument parsing
@ -102,3 +104,6 @@ void sample_top_k(std::vector<std::pair<double, gpt_vocab::id>> & logits_id, int
size_t ggml_quantize_q4_0(float * src, void * dst, int n, int k, int qk, int64_t * hist); size_t ggml_quantize_q4_0(float * src, void * dst, int n, int k, int qk, int64_t * hist);
size_t ggml_quantize_q4_1(float * src, void * dst, int n, int k, int qk, int64_t * hist); size_t ggml_quantize_q4_1(float * src, void * dst, int n, int k, int qk, int64_t * hist);
void untokenize(sentencepiece::SentencePieceProcessor & sp, std::vector<gpt_vocab::id> & embd);