tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-27 03:55:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama : improve sep token handling (#14272)

Browse Source
This commit is contained in:
Sigbjørn Skjæret
2025-06-20 14:04:09 +02:00
committed by GitHub
parent e28c1b93fd
commit 88fc854b4b
15 changed files with 161 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
convert_hf_to_gguf.py
View File

@ -2145,7 +2145,6 @@ class Llama4Model(LlamaModel):
def set_vocab(self):
self._set_vocab_gpt2()
self.gguf_writer.add_add_bos_token(True)
def set_gguf_parameters(self):
super().set_gguf_parameters()
@ -3918,9 +3917,6 @@ class BertModel(TextModel):
special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
special_vocab.add_to_gguf(self.gguf_writer)
self.gguf_writer.add_add_bos_token(True)
self.gguf_writer.add_add_eos_token(True)
@ModelBase.register("DistilBertModel", "DistilBertForMaskedLM", "DistilBertForSequenceClassification")
class DistilBertModel(BertModel):
@ -3962,8 +3958,6 @@ class RobertaModel(BertModel):
bpe_tok_path = self.dir_model / "tokenizer.json"
if bpe_tok_path.exists():
self._set_vocab_gpt2()
self.gguf_writer.add_add_bos_token(True)
self.gguf_writer.add_add_eos_token(True)
# we need this to validate the size of the token_type embeddings
# though currently we are passing all zeros to the token_type embeddings
@ -4848,8 +4842,6 @@ class JinaBertV2Model(BertModel):
self.gguf_writer.add_token_type_count(2)
else:
raise NotImplementedError(f'Tokenizer {tokenizer_class} is not supported for JinaBertModel')
self.gguf_writer.add_add_bos_token(True)
self.gguf_writer.add_add_eos_token(True)
@ModelBase.register("OpenELMForCausalLM")
@ -5451,9 +5443,6 @@ class T5Model(TextModel):
special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
special_vocab.add_to_gguf(self.gguf_writer)
self.gguf_writer.add_add_bos_token(False)
self.gguf_writer.add_add_eos_token(True)
def set_gguf_parameters(self):
if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None:
logger.warning("Couldn't find context length in config.json, assuming default value of 512")
@ -5591,9 +5580,6 @@ class T5EncoderModel(TextModel):
special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
special_vocab.add_to_gguf(self.gguf_writer)
self.gguf_writer.add_add_bos_token(False)
self.gguf_writer.add_add_eos_token(True)
def set_gguf_parameters(self):
if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None:
logger.warning("Couldn't find context length in config.json, assuming default value of 512")