1#!/usr/bin/env python3
2import logging
3import argparse
4import os
5import sys
6from pathlib import Path
7
8# Necessary to load the local gguf package
9if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent.parent / 'gguf-py').exists():
10 sys.path.insert(0, str(Path(__file__).parent.parent.parent))
11
12from gguf import GGUFReader # noqa: E402
13
14logger = logging.getLogger("gguf-set-metadata")
15
16
17def minimal_example(filename: str) -> None:
18 reader = GGUFReader(filename, 'r+')
19 field = reader.fields['tokenizer.ggml.bos_token_id']
20 if field is None:
21 return
22 part_index = field.data[0]
23 field.parts[part_index][0] = 2 # Set tokenizer.ggml.bos_token_id to 2
24 #
25 # So what's this field.data thing? It's helpful because field.parts contains
26 # _every_ part of the GGUF field. For example, tokenizer.ggml.bos_token_id consists
27 # of:
28 #
29 # Part index 0: Key length (27)
30 # Part index 1: Key data ("tokenizer.ggml.bos_token_id")
31 # Part index 2: Field type (4, the id for GGUFValueType.UINT32)
32 # Part index 3: Field value
33 #
34 # Note also that each part is an NDArray slice, so even a part that
35 # is only a single value like the key length will be a NDArray of
36 # the key length type (numpy.uint32).
37 #
38 # The .data attribute in the Field is a list of relevant part indexes
39 # and doesn't contain internal GGUF details like the key length part.
40 # In this case, .data will be [3] - just the part index of the
41 # field value itself.
42
43
44def set_metadata(reader: GGUFReader, args: argparse.Namespace) -> None:
45 field = reader.get_field(args.key)
46 if field is None:
47 logger.error(f'! Field {repr(args.key)} not found')
48 sys.exit(1)
49 # Note that field.types is a list of types. This is because the GGUF
50 # format supports arrays. For example, an array of UINT32 would
51 # look like [GGUFValueType.ARRAY, GGUFValueType.UINT32]
52 handler = reader.gguf_scalar_to_np.get(field.types[0]) if field.types else None
53 if handler is None:
54 logger.error(f'! This tool only supports changing simple values, {repr(args.key)} has unsupported type {field.types}')
55 sys.exit(1)
56 current_value = field.parts[field.data[0]][0]
57 new_value = handler(args.value)
58 logger.info(f'* Preparing to change field {repr(args.key)} from {current_value} to {new_value}')
59 if current_value == new_value:
60 logger.info(f'- Key {repr(args.key)} already set to requested value {current_value}')
61 sys.exit(0)
62 if args.dry_run:
63 sys.exit(0)
64 if not args.force:
65 logger.warning('*** Warning *** Warning *** Warning **')
66 logger.warning('* Changing fields in a GGUF file can make it unusable. Proceed at your own risk.')
67 logger.warning('* Enter exactly YES if you are positive you want to proceed:')
68 response = input('YES, I am sure> ')
69 if response != 'YES':
70 logger.info("You didn't enter YES. Okay then, see ya!")
71 sys.exit(0)
72 field.parts[field.data[0]][0] = new_value
73 logger.info('* Field changed. Successful completion.')
74
75
76def main() -> None:
77 parser = argparse.ArgumentParser(description="Set a simple value in GGUF file metadata")
78 parser.add_argument("model", type=str, help="GGUF format model filename")
79 parser.add_argument("key", type=str, help="Metadata key to set")
80 parser.add_argument("value", type=str, help="Metadata value to set")
81 parser.add_argument("--dry-run", action="store_true", help="Don't actually change anything")
82 parser.add_argument("--force", action="store_true", help="Change the field without confirmation")
83 parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
84
85 args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"])
86
87 logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
88
89 logger.info(f'* Loading: {args.model}')
90 reader = GGUFReader(args.model, 'r' if args.dry_run else 'r+')
91 set_metadata(reader, args)
92
93
94if __name__ == '__main__':
95 main()