From 1b67731e184e27a465b8c5476061294a4af668ea Mon Sep 17 00:00:00 2001 From: Jared Van Bortel Date: Tue, 9 Apr 2024 13:44:08 -0400 Subject: BERT tokenizer fixes (#6498) 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 --- examples/infill/infill.cpp | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) (limited to 'examples/infill/infill.cpp') diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp index 91c39c5a..c69dcd06 100644 --- a/examples/infill/infill.cpp +++ b/examples/infill/infill.cpp @@ -239,6 +239,7 @@ int main(int argc, char ** argv) { LOG_TEE("%s\n", get_system_info(params).c_str()); } const bool add_bos = llama_should_add_bos_token(model); + GGML_ASSERT(llama_add_eos_token(model) != 1); LOG("add_bos: %d\n", add_bos); bool suff_rm_leading_spc = params.escape; @@ -279,10 +280,10 @@ int main(int argc, char ** argv) { if (ctx_guidance) { LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(sparams.cfg_negative_prompt)); - guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, add_bos); + guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, true); LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp).c_str()); - std::vector original_inp = ::llama_tokenize(ctx, params.prompt, add_bos); + std::vector original_inp = ::llama_tokenize(ctx, params.prompt, true); LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp).c_str()); original_prompt_len = original_inp.size(); -- cgit v1.2.3