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

gguf-py: Refactor and allow reading/modifying existing GGUF files (#3981)

Browse source 
* gguf-py: Refactor and add file reading support

* Replay changes from #3871

Credit to @cebtenzzre for that pull

* Various type annotation fixes.

* sort imports with isort (again)

* Fix missing return statement in add_tensor

* style cleanup with flake8

* fix NamedTuple and Enum usage

* Fix an issue with state init in GGUFReader

Move examples to an examples/ directory

Clean up examples

Add an example of modifying keys in a GGUF file

Update documentation with info on examples

Try to support people importing gguf/gguf.py directly

* Damagage is not a word.

* Clean up gguf-py/examples/modify_gguf.py whitespace

Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com>

* Update gguf-py/examples/modify_gguf.py formatting

Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com>

* Update gguf-py/gguf/gguf_reader.py type hint

Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com>

* Make examples executable, formatting changes

* Add more information to GGUFReader and examples comments

* Include a gguf Python package version bump

* Add convert-gguf-endian.py script

* cleanup

* gguf-py : bump minor version

* Reorganize scripts

* Make GGUFReader endian detection less arbitrary

* Add JSON dumping support to gguf-dump.py

Which I kind of regret now

* A few for gguf-dump.py cleanups

* Murder accidental tuple in gguf-py/scripts/gguf-dump.py

Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com>

* cleanup

* constants : remove unneeded type annotations

* fix python 3.8 compat

* Set up gguf- scripts in pyproject.toml

* And include scripts/__init__.py, derp

* convert.py: We can't currently support Q8_0 on big endian.

* gguf-py: SpecialVocab: Always try available sources for special token ids

gguf-py: SpecialVocab: Try to load merges from merges.txt if not in tokenizer.json

gguf-py: SpecialVocab: Add 'add_bos_token' type bools to GGUF metadata
u

* cleanup

* Promote add_X_token to GGUF metadata for BOS and EOS

---------

Co-authored-by: Jared Van Bortel <jared@nomic.ai>
Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com>
This commit is contained in:
Kerfuffle 2023-11-10 22:04:50 -07:00 • committed by GitHub
parent 4a4fd3eefa
commit 34b0a08207
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
20 changed files with 1982 additions and 1176 deletions
Download patch file Download diff file Expand all files Collapse all files

90
gguf-py/scripts/gguf-set-metadata.py Executable file
View file

@ -0,0 +1,90 @@
#!/usr/bin/env python3
import argparse
import os
import sys
from pathlib import Path
# Necessary to load the local gguf package
if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / 'gguf-py').exists():
sys.path.insert(0, str(Path(__file__).parent.parent))
from gguf import GGUFReader # noqa: E402
def minimal_example(filename: str) -> None:
reader = GGUFReader(filename, 'r+')
field = reader.fields['tokenizer.ggml.bos_token_id']
if field is None:
return
part_index = field.data[0]
field.parts[part_index][0] = 2 # Set tokenizer.ggml.bos_token_id to 2
#
# So what's this field.data thing? It's helpful because field.parts contains
# _every_ part of the GGUF field. For example, tokenizer.ggml.bos_token_id consists
# of:
#
# Part index 0: Key length (27)
# Part index 1: Key data ("tokenizer.ggml.bos_token_id")
# Part index 2: Field type (4, the id for GGUFValueType.UINT32)
# Part index 3: Field value
#
# Note also that each part is an NDArray slice, so even a part that
# is only a single value like the key length will be a NDArray of
# the key length type (numpy.uint32).
#
# The .data attribute in the Field is a list of relevant part indexes
# and doesn't contain internal GGUF details like the key length part.
# In this case, .data will be [3] - just the part index of the
# field value itself.
def set_metadata(reader: GGUFReader, args: argparse.Namespace) -> None:
field = reader.get_field(args.key)
if field is None:
print(f'! Field {repr(args.key)} not found', file = sys.stderr)
sys.exit(1)
# Note that field.types is a list of types. This is because the GGUF
# format supports arrays. For example, an array of UINT32 would
# look like [GGUFValueType.ARRAY, GGUFValueType.UINT32]
handler = reader.gguf_scalar_to_np.get(field.types[0]) if field.types else None
if handler is None:
print(
f'! This tool only supports changing simple values, {repr(args.key)} has unsupported type {field.types}',
file = sys.stderr,
)
sys.exit(1)
current_value = field.parts[field.data[0]][0]
new_value = handler(args.value)
print(f'* Preparing to change field {repr(args.key)} from {current_value} to {new_value}')
if current_value == new_value:
print(f'- Key {repr(args.key)} already set to requested value {current_value}')
sys.exit(0)
if args.dry_run:
sys.exit(0)
if not args.force:
print('*** Warning *** Warning *** Warning **')
print('* Changing fields in a GGUF file can make it unusable. Proceed at your own risk.')
print('* Enter exactly YES if you are positive you want to proceed:')
response = input('YES, I am sure> ')
if response != 'YES':
print("You didn't enter YES. Okay then, see ya!")
sys.exit(0)
field.parts[field.data[0]][0] = new_value
print('* Field changed. Successful completion.')
def main() -> None:
parser = argparse.ArgumentParser(description="Set a simple value in GGUF file metadata")
parser.add_argument("model", type=str, help="GGUF format model filename")
parser.add_argument("key", type=str, help="Metadata key to set")
parser.add_argument("value", type=str, help="Metadata value to set")
parser.add_argument("--dry-run", action="store_true", help="Don't actually change anything")
parser.add_argument("--force", action="store_true", help="Change the field without confirmation")
args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"])
print(f'* Loading: {args.model}')
reader = GGUFReader(args.model, 'r' if args.dry_run else 'r+')
set_metadata(reader, args)
if __name__ == '__main__':
main()