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

Add --pre-tokenizer option to convert

Browse source
This commit is contained in:
Galunid 2024-06-16 20:18:27 +02:00
parent bc6c457fa3
commit f05a0e0a00
2 changed files with 12 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

3
convert-hf-to-gguf-update.py
View file

@ -173,6 +173,9 @@ for model in models:
src_func = f""" src_func = f"""
def get_vocab_base_pre(self, tokenizer) -> str: def get_vocab_base_pre(self, tokenizer) -> str:
if self.pre_tokenizer is not None:
return self.pre_tokenizer
# encoding this string and hashing the resulting tokens would (hopefully) give us a unique identifier that # encoding this string and hashing the resulting tokens would (hopefully) give us a unique identifier that
# is specific for the BPE pre-tokenizer used by the model # is specific for the BPE pre-tokenizer used by the model
# we will use this unique identifier to write a "tokenizer.ggml.pre" entry in the GGUF file which we can # we will use this unique identifier to write a "tokenizer.ggml.pre" entry in the GGUF file which we can

11
convert-hf-to-gguf.py
View file

@ -65,7 +65,7 @@ class Model:
# subclasses should define this! # subclasses should define this!
model_arch: gguf.MODEL_ARCH model_arch: gguf.MODEL_ARCH
def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool, use_temp_file: bool, eager: bool, model_name: str | None): def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool, use_temp_file: bool, eager: bool, model_name: str | None, pre_tokenizer: str | None):
if type(self) is Model: if type(self) is Model:
raise TypeError(f"{type(self).__name__!r} should not be directly instantiated") raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
self.dir_model = dir_model self.dir_model = dir_model
@ -73,6 +73,7 @@ class Model:
self.is_big_endian = is_big_endian self.is_big_endian = is_big_endian
self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
self.use_temp_file = use_temp_file self.use_temp_file = use_temp_file
self.pre_tokenizer = pre_tokenizer
self.lazy = not eager self.lazy = not eager
self.model_name = model_name self.model_name = model_name
self.part_names = Model.get_model_part_names(self.dir_model, "model", ".safetensors") self.part_names = Model.get_model_part_names(self.dir_model, "model", ".safetensors")
@ -405,6 +406,8 @@ class Model:
# ref: https://github.com/ggerganov/llama.cpp/pull/6920 # ref: https://github.com/ggerganov/llama.cpp/pull/6920
# Marker: Start get_vocab_base_pre # Marker: Start get_vocab_base_pre
def get_vocab_base_pre(self, tokenizer) -> str: def get_vocab_base_pre(self, tokenizer) -> str:
if self.pre_tokenizer is not None:
return self.pre_tokenizer
# encoding this string and hashing the resulting tokens would (hopefully) give us a unique identifier that # encoding this string and hashing the resulting tokens would (hopefully) give us a unique identifier that
# is specific for the BPE pre-tokenizer used by the model # is specific for the BPE pre-tokenizer used by the model
# we will use this unique identifier to write a "tokenizer.ggml.pre" entry in the GGUF file which we can # we will use this unique identifier to write a "tokenizer.ggml.pre" entry in the GGUF file which we can
@ -2800,6 +2803,10 @@ def parse_args() -> argparse.Namespace:
"--model-name", type=str, default=None, "--model-name", type=str, default=None,
help="name of the model", help="name of the model",
) )
parser.add_argument(
"--pre-tokenizer", type=str, default=None,
help="overwrite pre-tokenizer, if not specified this script will try to detect it automatically"
)
parser.add_argument( parser.add_argument(
"--verbose", action="store_true", "--verbose", action="store_true",
help="increase output verbosity", help="increase output verbosity",
@ -2857,7 +2864,7 @@ def main() -> None:
logger.error(f"Model {hparams['architectures'][0]} is not supported") logger.error(f"Model {hparams['architectures'][0]} is not supported")
sys.exit(1) sys.exit(1)
model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian, args.use_temp_file, args.no_lazy, args.model_name) model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian, args.use_temp_file, args.no_lazy, args.model_name, args.pre_tokenizer)
logger.info("Set model parameters") logger.info("Set model parameters")
model_instance.set_gguf_parameters() model_instance.set_gguf_parameters()