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

llama : sanitize invalid tokens (#9357)

Browse source 
* common : do not add null tokens during warmup

ggml-ci

* llama : check that the input tokens are valid

ggml-ci

* tests : fix batch size of bert model

ggml-ci
This commit is contained in:
Georgi Gerganov 2024-09-08 00:33:13 +03:00 • committed by GitHub
parent e536426ded
commit faf69d4237
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 26 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

9
common/common.cpp
View file

@ -2690,10 +2690,15 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
llama_token bos = llama_token_bos(model);
llama_token eos = llama_token_eos(model);
// some models (e.g. T5) don't have a BOS token
if (bos != -1) {
if (bos != LLAMA_TOKEN_NULL) {
tmp.push_back(bos);
}
tmp.push_back(eos);
if (eos != LLAMA_TOKEN_NULL) {
tmp.push_back(eos);
}
if (tmp.empty()) {
tmp.push_back(0);
}
if (llama_model_has_encoder(model)) {
llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size(), 0, 0));