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

llama : infill sampling handle very long tokens (#9924)

Browse source 
* llama : infill sampling handle very long tokens

ggml-ci

* cont : better indices

ggml-ci
This commit is contained in:
Georgi Gerganov 2024-10-17 22:32:47 +03:00 • committed by GitHub
parent 3752217ed5
commit 99bd4ac28c
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 35 additions and 43 deletions
Download patch file Download diff file Expand all files Collapse all files

17
src/llama-vocab.cpp
View file

@ -1858,23 +1858,6 @@ int32_t llama_token_to_piece_impl(const struct llama_vocab & vocab, llama_token
return 0;
}
bool llama_token_is_prefix_impl(
const struct llama_vocab & vocab,
llama_token token0,
llama_token token1) {
char text_buf_0[128];
char text_buf_1[128];
const int32_t len0 = llama_token_to_piece_impl(vocab, token0, text_buf_0, sizeof(text_buf_0) - 1, 0, false);
const int32_t len1 = llama_token_to_piece_impl(vocab, token1, text_buf_1, sizeof(text_buf_1) - 1, 0, false);
if (len0 <= 0 || len1 <= 0) {
return false;
}
return len0 <= len1 && memcmp(text_buf_0, text_buf_1, len0) == 0;
}
int32_t llama_detokenize_impl(
const struct llama_vocab & vocab,
const llama_token * tokens,