LLM inference in C/C++
C++ ggml
vocab : honor BOS/EOS settings for PLaMo-2 and PLaMo-3 (#29734)
Toki Nasindb33d3cb89d8
diff --git a/conversion/base.py b/conversion/base.pyindex 35f564e40c..a39a728f5f 100644--- a/conversion/base.py+++ b/conversion/base.py@@ -2566,6 +2566,11 @@ class TextModel(ModelBase):self.gguf_writer.add_add_space_prefix(False)+ if (add_bos := tokenizer_config.get("add_bos_token")) is not None:+ self.gguf_writer.add_add_bos_token(add_bos)+ if (add_eos := tokenizer_config.get("add_eos_token")) is not None:+ self.gguf_writer.add_add_eos_token(add_eos)+class MmprojModel(ModelBase):model_type = ModelType.MMPROJdiff --git a/src/llama-vocab.cpp b/src/llama-vocab.cppindex 93468b6b4c..e5703f3e81 100644--- a/src/llama-vocab.cpp+++ b/src/llama-vocab.cpp@@ -3599,6 +3599,10 @@ std::vector<llama_token> llama_vocab::impl::tokenize(} break;case LLAMA_VOCAB_TYPE_PLAMO2:{+ if (add_special && add_bos) {+ GGML_ASSERT(special_bos_id != LLAMA_TOKEN_NULL);+ output.push_back(special_bos_id);+ }llm_tokenizer_plamo2_session session(*static_cast<const llm_tokenizer_plamo2 *>(tokenizer.get()));for (const auto & fragment : fragment_buffer) {if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {@@ -3613,6 +3617,18 @@ std::vector<llama_token> llama_vocab::impl::tokenize(output.push_back(fragment.token);}}++ if (add_special && add_bos && output.size() >= 2 && output[1] == special_bos_id) {+ LLAMA_LOG_WARN(+ "%s: Added a BOS token to the prompt as specified by the model but the prompt "+ "also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. "+ "Are you sure this is what you want?\n", __FUNCTION__);+ }++ if (add_special && add_eos) {+ GGML_ASSERT(special_eos_id != LLAMA_TOKEN_NULL);+ output.push_back(special_eos_id);+ }} break;case LLAMA_VOCAB_TYPE_TEST:{