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

BERT tokenizer fixes (#6498)

Browse source 
Key changes:
* BERT conversion: fix abuse of LlamaHfVocab, do not set BOS or EOS
* Nomic Embed conversion: pad vocab instead of slicing embedding tensor
* llama_tokenize: handle added special tokens like HF does
This commit is contained in:
Jared Van Bortel 2024-04-09 13:44:08 -04:00 • committed by GitHub
parent c4a3a4ff47
commit 1b67731e18
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
20 changed files with 221 additions and 194 deletions
Download patch file Download diff file Expand all files Collapse all files

5
examples/lookahead/lookahead.cpp
View file

@ -64,13 +64,10 @@ int main(int argc, char ** argv) {
std::tie(model, ctx) = llama_init_from_gpt_params(params);
// Tokenize the prompt
const bool add_bos = llama_should_add_bos_token(model);
LOG("add_bos tgt: %d\n", add_bos);
std::vector<llama_token> inp;
std::vector<llama_token> all;
inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
inp = ::llama_tokenize(ctx, params.prompt, true, true);
all = inp;
const int max_context_size = llama_n_ctx(ctx);