Merge 41efa86198
into 96776405a1
This commit is contained in:
commit
cdc3e78bb6
5 changed files with 33 additions and 6 deletions
|
@ -167,6 +167,7 @@ class Keys:
|
||||||
CHAT_TEMPLATE = "tokenizer.chat_template"
|
CHAT_TEMPLATE = "tokenizer.chat_template"
|
||||||
CHAT_TEMPLATE_N = "tokenizer.chat_template.{name}"
|
CHAT_TEMPLATE_N = "tokenizer.chat_template.{name}"
|
||||||
CHAT_TEMPLATES = "tokenizer.chat_templates"
|
CHAT_TEMPLATES = "tokenizer.chat_templates"
|
||||||
|
INVERSE_TEMPLATE = "tokenizer.inverse_template"
|
||||||
# FIM/Infill special tokens constants
|
# FIM/Infill special tokens constants
|
||||||
PREFIX_ID = "tokenizer.ggml.prefix_token_id"
|
PREFIX_ID = "tokenizer.ggml.prefix_token_id"
|
||||||
SUFFIX_ID = "tokenizer.ggml.suffix_token_id"
|
SUFFIX_ID = "tokenizer.ggml.suffix_token_id"
|
||||||
|
|
|
@ -843,6 +843,9 @@ class GGUFWriter:
|
||||||
|
|
||||||
self.add_string(Keys.Tokenizer.CHAT_TEMPLATE, value)
|
self.add_string(Keys.Tokenizer.CHAT_TEMPLATE, value)
|
||||||
|
|
||||||
|
def add_inverse_template(self, value: str) -> None:
|
||||||
|
self.add_string(Keys.Tokenizer.INVERSE_TEMPLATE, value)
|
||||||
|
|
||||||
def add_prefix_token_id(self, id: int) -> None:
|
def add_prefix_token_id(self, id: int) -> None:
|
||||||
self.add_uint32(Keys.Tokenizer.PREFIX_ID, id)
|
self.add_uint32(Keys.Tokenizer.PREFIX_ID, id)
|
||||||
|
|
||||||
|
|
|
@ -21,6 +21,7 @@ class SpecialVocab:
|
||||||
add_special_token: dict[str, bool]
|
add_special_token: dict[str, bool]
|
||||||
special_token_ids: dict[str, int]
|
special_token_ids: dict[str, int]
|
||||||
chat_template: str | Sequence[Mapping[str, str]] | None
|
chat_template: str | Sequence[Mapping[str, str]] | None
|
||||||
|
inverse_template: str | None
|
||||||
|
|
||||||
def __init__(
|
def __init__(
|
||||||
self, path: str | os.PathLike[str], load_merges: bool = False,
|
self, path: str | os.PathLike[str], load_merges: bool = False,
|
||||||
|
@ -33,6 +34,7 @@ class SpecialVocab:
|
||||||
self.load_merges = load_merges
|
self.load_merges = load_merges
|
||||||
self.merges = []
|
self.merges = []
|
||||||
self.chat_template = None
|
self.chat_template = None
|
||||||
|
self.inverse_template = None
|
||||||
if special_token_types is not None:
|
if special_token_types is not None:
|
||||||
self.special_token_types = special_token_types
|
self.special_token_types = special_token_types
|
||||||
else:
|
else:
|
||||||
|
@ -71,6 +73,10 @@ class SpecialVocab:
|
||||||
if not quiet:
|
if not quiet:
|
||||||
logger.info(f'Setting chat_template to {self.chat_template}')
|
logger.info(f'Setting chat_template to {self.chat_template}')
|
||||||
gw.add_chat_template(self.chat_template)
|
gw.add_chat_template(self.chat_template)
|
||||||
|
if self.inverse_template is not None:
|
||||||
|
if not quiet:
|
||||||
|
logger.info(f'Setting inverse_template to {self.inverse_template}')
|
||||||
|
gw.add_inverse_template(self.inverse_template)
|
||||||
|
|
||||||
def _load(self, path: Path) -> None:
|
def _load(self, path: Path) -> None:
|
||||||
self._try_load_from_tokenizer_json(path)
|
self._try_load_from_tokenizer_json(path)
|
||||||
|
@ -159,6 +165,11 @@ class SpecialVocab:
|
||||||
self.chat_template = chat_template
|
self.chat_template = chat_template
|
||||||
else:
|
else:
|
||||||
logger.warning(f'Bad type for chat_template field in {tokenizer_config_file!r} - ignoring')
|
logger.warning(f'Bad type for chat_template field in {tokenizer_config_file!r} - ignoring')
|
||||||
|
inverse_template = tokenizer_config.get('inverse_template')
|
||||||
|
if inverse_template is None or isinstance(inverse_template, str):
|
||||||
|
self.inverse_template = inverse_template
|
||||||
|
else:
|
||||||
|
logger.warning(f'Bad type for inverse_template field in {tokenizer_config_file!r} - ignoring')
|
||||||
for typ in self.special_token_types:
|
for typ in self.special_token_types:
|
||||||
add_entry = tokenizer_config.get(f'add_{typ}_token')
|
add_entry = tokenizer_config.get(f'add_{typ}_token')
|
||||||
if isinstance(add_entry, bool):
|
if isinstance(add_entry, bool):
|
||||||
|
|
|
@ -1,6 +1,6 @@
|
||||||
[tool.poetry]
|
[tool.poetry]
|
||||||
name = "gguf"
|
name = "gguf"
|
||||||
version = "0.10.0"
|
version = "0.11.0"
|
||||||
description = "Read and write ML models in GGUF for GGML"
|
description = "Read and write ML models in GGUF for GGML"
|
||||||
authors = ["GGML <ggml@ggml.ai>"]
|
authors = ["GGML <ggml@ggml.ai>"]
|
||||||
packages = [
|
packages = [
|
||||||
|
|
|
@ -85,7 +85,7 @@ def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new
|
||||||
continue
|
continue
|
||||||
|
|
||||||
# Skip old chat templates if we have new ones
|
# Skip old chat templates if we have new ones
|
||||||
if field.name.startswith(gguf.Keys.Tokenizer.CHAT_TEMPLATE) and gguf.Keys.Tokenizer.CHAT_TEMPLATE in new_metadata:
|
if (field.name.startswith(gguf.Keys.Tokenizer.CHAT_TEMPLATE) and gguf.Keys.Tokenizer.CHAT_TEMPLATE in new_metadata) or (field.name.startswith(gguf.Keys.Tokenizer.INVERSE_TEMPLATE) and gguf.Keys.Tokenizer.INVERSE_TEMPLATE in new_metadata):
|
||||||
logger.debug(f'Skipping {field.name}')
|
logger.debug(f'Skipping {field.name}')
|
||||||
continue
|
continue
|
||||||
|
|
||||||
|
@ -110,6 +110,11 @@ def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new
|
||||||
writer.add_chat_template(new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE].value)
|
writer.add_chat_template(new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE].value)
|
||||||
del new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE]
|
del new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE]
|
||||||
|
|
||||||
|
if gguf.Keys.Tokenizer.INVERSE_TEMPLATE in new_metadata:
|
||||||
|
logger.debug('Adding inverse template')
|
||||||
|
writer.add_inverse_template(new_metadata[gguf.Keys.Tokenizer.INVERSE_TEMPLATE].value)
|
||||||
|
del new_metadata[gguf.Keys.Tokenizer.INVERSE_TEMPLATE]
|
||||||
|
|
||||||
for key, val in new_metadata.items():
|
for key, val in new_metadata.items():
|
||||||
logger.debug(f'Adding {key}: "{val.value}" {val.description}')
|
logger.debug(f'Adding {key}: "{val.value}" {val.description}')
|
||||||
writer.add_key_value(key, val.value, val.type)
|
writer.add_key_value(key, val.value, val.type)
|
||||||
|
@ -143,7 +148,8 @@ def main() -> None:
|
||||||
parser.add_argument("--general-name", type=str, help="The models general.name", metavar='"name"')
|
parser.add_argument("--general-name", type=str, help="The models general.name", metavar='"name"')
|
||||||
parser.add_argument("--general-description", type=str, help="The models general.description", metavar='"Description ..."')
|
parser.add_argument("--general-description", type=str, help="The models general.description", metavar='"Description ..."')
|
||||||
parser.add_argument("--chat-template", type=str, help="Chat template string (or JSON string containing templates)", metavar='"{% ... %} ..."')
|
parser.add_argument("--chat-template", type=str, help="Chat template string (or JSON string containing templates)", metavar='"{% ... %} ..."')
|
||||||
parser.add_argument("--chat-template-config", type=Path, help="Config file containing chat template(s)", metavar='tokenizer_config.json')
|
parser.add_argument("--inverse-template", type=str, help="Inverse template string", metavar='"{% ... %} ..."')
|
||||||
|
parser.add_argument("--chat-template-config", type=Path, help="Config file containing chat and/or inverse template(s)", metavar='tokenizer_config.json')
|
||||||
parser.add_argument("--pre-tokenizer", type=str, help="The models tokenizer.ggml.pre", metavar='"pre tokenizer"')
|
parser.add_argument("--pre-tokenizer", type=str, help="The models tokenizer.ggml.pre", metavar='"pre tokenizer"')
|
||||||
parser.add_argument("--remove-metadata", action="append", type=str, help="Remove metadata (by key name) from output model", metavar='general.url')
|
parser.add_argument("--remove-metadata", action="append", type=str, help="Remove metadata (by key name) from output model", metavar='general.url')
|
||||||
parser.add_argument("--special-token", action="append", type=str, help="Special token by value", nargs=2, metavar=(' | '.join(token_names.keys()), '"<token>"'))
|
parser.add_argument("--special-token", action="append", type=str, help="Special token by value", nargs=2, metavar=(' | '.join(token_names.keys()), '"<token>"'))
|
||||||
|
@ -166,12 +172,18 @@ def main() -> None:
|
||||||
if args.chat_template:
|
if args.chat_template:
|
||||||
new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, json.loads(args.chat_template) if args.chat_template.startswith('[') else args.chat_template)
|
new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, json.loads(args.chat_template) if args.chat_template.startswith('[') else args.chat_template)
|
||||||
|
|
||||||
|
if args.inverse_template:
|
||||||
|
new_metadata[gguf.Keys.Tokenizer.INVERSE_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, args.inverse_template)
|
||||||
|
|
||||||
if args.chat_template_config:
|
if args.chat_template_config:
|
||||||
with open(args.chat_template_config, 'r') as fp:
|
with open(args.chat_template_config, 'r') as fp:
|
||||||
config = json.load(fp)
|
config = json.load(fp)
|
||||||
template = config.get('chat_template')
|
chat_template = config.get('chat_template')
|
||||||
if template:
|
inverse_template = config.get('inverse_template')
|
||||||
new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, template)
|
if chat_template:
|
||||||
|
new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, chat_template)
|
||||||
|
if inverse_template:
|
||||||
|
new_metadata[gguf.Keys.Tokenizer.INVERSE_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, inverse_template)
|
||||||
|
|
||||||
if args.pre_tokenizer:
|
if args.pre_tokenizer:
|
||||||
new_metadata[gguf.Keys.Tokenizer.PRE] = MetadataDetails(gguf.GGUFValueType.STRING, args.pre_tokenizer)
|
new_metadata[gguf.Keys.Tokenizer.PRE] = MetadataDetails(gguf.GGUFValueType.STRING, args.pre_tokenizer)
|
||||||
|
|
Loading…
Add table
Add a link
Reference in a new issue