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

solve process prompt bug

Browse source 
fix #6823 
for example 
```
What credit card company is on the banner in the background?\nAnswer the question using a single word or phrase.
```
notice the different tokenization result of '\n'
before fix, the verbose-prompt result is:
```
  2061 -> 'What'
  3884 -> ' credit'
  2657 -> ' card'
  1664 -> ' company'
   318 -> ' is'
   319 -> ' on'
   262 -> ' the'
 17625 -> ' banner'
   287 -> ' in'
   262 -> ' the'
  4469 -> ' background'
    30 -> '?'
    59 -> '\'
    77 -> 'n'
 33706 -> 'Answer'
   262 -> ' the'
  1808 -> ' question'
  1262 -> ' using'
   257 -> ' a'
  2060 -> ' single'
  1573 -> ' word'
   393 -> ' or'
  9546 -> ' phrase'
    13 -> '.'

```
after fix:
```
2061 -> 'What'
  3884 -> ' credit'
  2657 -> ' card'
  1664 -> ' company'
   318 -> ' is'
   319 -> ' on'
   262 -> ' the'
 17625 -> ' banner'
   287 -> ' in'
   262 -> ' the'
  4469 -> ' background'
    30 -> '?'
   198 -> '
'
 33706 -> 'Answer'
   262 -> ' the'
  1808 -> ' question'
  1262 -> ' using'
   257 -> ' a'
  2060 -> ' single'
  1573 -> ' word'
   393 -> ' or'
  9546 -> ' phrase'
    13 -> '.
```
This commit is contained in:
Leo Zhang 2024-05-13 19:32:50 +08:00 • committed by GitHub
parent 1c570d8bee
commit 3176c2f561
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 11 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

10
examples/llava/llava-cli.cpp
View file

@ -272,6 +272,15 @@ static void llama_log_callback_logTee(ggml_log_level level, const char * text, v
LOG_TEE("%s", text);
}
static std::string replaceWithN(std::string& input) {
size_t pos = 0;
while ((pos = input.find("\\n", pos)) != std::string::npos) {
input.replace(pos, 2, "\n");
pos += 1;
}
return input;
}
int main(int argc, char ** argv) {
ggml_time_init();
@ -310,6 +319,7 @@ int main(int argc, char ** argv) {
}
// process the prompt
params.prompt = replaceWithN(params.prompt);
process_prompt(ctx_llava, image_embed, &params, params.prompt);
llama_print_timings(ctx_llava->ctx_llama);