tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-07-28 21:23:55 -04:00
Code Issues Packages Projects Releases Wiki Activity

llama : the WPM vocabs use the CLS token as BOS (#10930)

Browse Source 
* llama : the WPM vocabs use the CLS token as BOS

ggml-ci

* llama : add comment
This commit is contained in:
Georgi Gerganov
2024-12-24 09:44:20 +02:00
committed by GitHub
parent 60cfa728e2
commit 30caac3a68
2 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/llama-vocab.cpp
View File

@@ -1657,7 +1657,7 @@ bool llama_token_is_control_impl(const struct llama_vocab & vocab, llama_token t
}
llama_token llama_token_bos_impl(const struct llama_vocab & vocab) {
return vocab.special_bos_id;
return vocab.type != LLAMA_VOCAB_TYPE_WPM ? vocab.special_bos_id : vocab.special_cls_id;
}
llama_token llama_token_eos_impl(const struct llama_vocab & vocab) {

2
src/llama-vocab.h
View File

@@ -45,7 +45,7 @@ struct llama_vocab {
id special_unk_id = 0;
id special_sep_id = LLAMA_TOKEN_NULL;
id special_pad_id = LLAMA_TOKEN_NULL;
id special_cls_id = LLAMA_TOKEN_NULL;
id special_cls_id = LLAMA_TOKEN_NULL; // TODO: revisit if this is really needed https://github.com/ggerganov/llama.cpp/pull/10930
id special_mask_id = LLAMA_TOKEN_NULL;
id linefeed_id = 13;