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

Fix spm whitespaces (#2806)

Browse source 
* llama.cpp : fix spm whitespace escaping + clean up

* main.cpp : spm - add whitespace in front of prompt

* test-tokenizer-0.cpp : spm - add whitespace in front of prompt
This commit is contained in:
klosax 2023-08-26 13:45:53 +02:00 • committed by GitHub
parent bae5c5f679
commit 2ba83c8685
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
3 changed files with 27 additions and 41 deletions
Download patch file Download diff file Expand all files Collapse all files

17
examples/main/main.cpp
View file

@ -189,12 +189,19 @@ int main(int argc, char ** argv) {
}
}
const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
// Add BOS if SPM tokenizer
const bool add_bos = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
// tokenize the prompt
std::vector<llama_token> embd_inp;
if (llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM) {
// Add a space in front of the first character to match OG llama tokenizer behavior
params.prompt.insert(0, 1, ' ');
}
if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
embd_inp = ::llama_tokenize(ctx, params.prompt, is_spm);
embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
} else {
embd_inp = session_tokens;
}
@ -210,9 +217,9 @@ int main(int argc, char ** argv) {
int original_prompt_len = 0;
if (ctx_guidance) {
params.cfg_negative_prompt.insert(0, 1, ' ');
guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, is_spm);
guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, add_bos);
std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, is_spm);
std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
original_prompt_len = original_inp.size();
guidance_offset = (int)guidance_inp.size() - original_prompt_len;
}
@ -259,7 +266,7 @@ int main(int argc, char ** argv) {
}
// prefix & suffix for instruct mode
const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", is_spm);
const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos);
const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n", false);
// in instruct mode, we inject a prefix and a suffix to each input by the user