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

* 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>

4 files changed, 331 insertions, 0 deletions

diff --git a/gguf-py/scripts/__init__.py b/gguf-py/scripts/__init__.py
new file mode 100644
index 00000000..77132db7
--- /dev/null
+++ b/gguf-py/scripts/__init__.py
@@ -0,0 +1,12 @@
+import os
+
+from importlib import import_module
+
+
+os.environ["NO_LOCAL_GGUF"] = "TRUE"
+
+gguf_convert_endian_entrypoint = import_module("scripts.gguf-convert-endian").main
+gguf_dump_entrypoint           = import_module("scripts.gguf-dump").main
+gguf_set_metadata_entrypoint   = import_module("scripts.gguf-set-metadata").main
+
+del import_module, os
diff --git a/gguf-py/scripts/gguf-convert-endian.py b/gguf-py/scripts/gguf-convert-endian.py
new file mode 100755
index 00000000..b79d86e0
--- /dev/null
+++ b/gguf-py/scripts/gguf-convert-endian.py
@@ -0,0 +1,113 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import argparse
+import os
+import sys
+from pathlib import Path
+
+import numpy as np
+
+# 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))
+
+import gguf
+
+
+def convert_byteorder(reader: gguf.GGUFReader, args: argparse.Namespace) -> None:
+    if np.uint32(1) == np.uint32(1).newbyteorder("<"):
+        # Host is little endian
+        host_endian = "little"
+        swapped_endian = "big"
+    else:
+        # Sorry PDP or other weird systems that don't use BE or LE.
+        host_endian = "big"
+        swapped_endian = "little"
+    if reader.byte_order == "S":
+        file_endian = swapped_endian
+    else:
+        file_endian = host_endian
+    if args.order == "native":
+        order = host_endian
+    print(f"* Host is {host_endian.upper()} endian, GGUF file seems to be {file_endian.upper()} endian")
+    if file_endian == order:
+        print(f"* File is already {order.upper()} endian. Nothing to do.")
+        sys.exit(0)
+    print("* Checking tensors for conversion compatibility")
+    for tensor in reader.tensors:
+        if tensor.tensor_type not in (
+            gguf.GGMLQuantizationType.F32,
+            gguf.GGMLQuantizationType.F16,
+            gguf.GGMLQuantizationType.Q8_0,
+        ):
+            raise ValueError(f"Cannot handle type {tensor.tensor_type.name} for tensor {repr(tensor.name)}")
+    print(f"* Preparing to convert from {file_endian.upper()} to {order.upper()}")
+    if args.dry_run:
+        return
+    print("\n*** Warning *** Warning *** Warning **")
+    print("* This conversion process may damage the file. Ensure you have a backup.")
+    if order != host_endian:
+        print("* Requested endian differs from host, you will not be able to load the model on this machine.")
+    print("* The file will be modified immediately, so if conversion fails or is interrupted")
+    print("* the file will be corrupted. 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)
+    print(f"\n* Converting fields ({len(reader.fields)})")
+    for idx, field in enumerate(reader.fields.values()):
+        print(f"- {idx:4}: Converting field {repr(field.name)}, part count: {len(field.parts)}")
+        for part in field.parts:
+            part.byteswap(inplace=True)
+    print(f"\n* Converting tensors ({len(reader.tensors)})")
+    for idx, tensor in enumerate(reader.tensors):
+        print(
+            f"  - {idx:4}: Converting tensor {repr(tensor.name)}, type={tensor.tensor_type.name}, "
+            f"elements={tensor.n_elements}... ",
+            end="",
+        )
+        tensor_type = tensor.tensor_type
+        for part in tensor.field.parts:
+            part.byteswap(inplace=True)
+        if tensor_type != gguf.GGMLQuantizationType.Q8_0:
+            tensor.data.byteswap(inplace=True)
+            print()
+            continue
+        # A Q8_0 block consists of a f16 delta followed by 32 int8 quants, so 34 bytes
+        block_size = 34
+        n_blocks = len(tensor.data) // block_size
+        for block_num in range(n_blocks):
+            block_offs = block_num * block_size
+            # I know I said f16, but it doesn't matter here - any simple 16 bit type works.
+            delta = tensor.data[block_offs:block_offs + 2].view(dtype=np.uint16)
+            delta.byteswap(inplace=True)
+            if block_num % 100000 == 0:
+                print(f"[{(n_blocks - block_num) // 1000}K]", end="")
+                sys.stdout.flush()
+        print()
+    print("* Completion")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Convert GGUF file byte order")
+    parser.add_argument(
+        "model", type=str,
+        help="GGUF format model filename",
+    )
+    parser.add_argument(
+        "order", type=str, choices=['big', 'little', 'native'],
+        help="Requested byte order",
+    )
+    parser.add_argument(
+        "--dry-run", action="store_true",
+        help="Don't actually change anything",
+    )
+    args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"])
+    print(f'* Loading: {args.model}')
+    reader = gguf.GGUFReader(args.model, 'r' if args.dry_run else 'r+')
+    convert_byteorder(reader, args)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/gguf-py/scripts/gguf-dump.py b/gguf-py/scripts/gguf-dump.py
new file mode 100755
index 00000000..5141873d
--- /dev/null
+++ b/gguf-py/scripts/gguf-dump.py
@@ -0,0 +1,116 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import argparse
+import os
+import sys
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+
+# 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, GGUFValueType  # noqa: E402
+
+
+def get_file_host_endian(reader: GGUFReader) -> tuple[str, str]:
+    host_endian = 'LITTLE' if np.uint32(1) == np.uint32(1).newbyteorder("<") else 'BIG'
+    if reader.byte_order == 'S':
+        file_endian = 'BIG' if host_endian == 'LITTLE' else 'LITTLE'
+    else:
+        file_endian = host_endian
+    return (host_endian, file_endian)
+
+
+# For more information about what field.parts and field.data represent,
+# please see the comments in the modify_gguf.py example.
+def dump_metadata(reader: GGUFReader, args: argparse.Namespace) -> None:
+    host_endian, file_endian = get_file_host_endian(reader)
+    print(f'* File is {file_endian} endian, script is running on a {host_endian} endian host.')
+    print(f'\n* Dumping {len(reader.fields)} key/value pair(s)')
+    for n, field in enumerate(reader.fields.values(), 1):
+        if not field.types:
+            pretty_type = 'N/A'
+        elif field.types[0] == GGUFValueType.ARRAY:
+            nest_count = len(field.types) - 1
+            pretty_type = '[' * nest_count + str(field.types[-1].name) + ']' * nest_count
+        else:
+            pretty_type = str(field.types[-1].name)
+        print(f'  {n:5}: {pretty_type:10} | {len(field.data):8} | {field.name}', end = '')
+        if len(field.types) == 1:
+            curr_type = field.types[0]
+            if curr_type == GGUFValueType.STRING:
+                print(' = {0}'.format(repr(str(bytes(field.parts[-1]), encoding='utf8')[:60])), end = '')
+            elif field.types[0] in reader.gguf_scalar_to_np:
+                print(' = {0}'.format(field.parts[-1][0]), end = '')
+        print()
+    if args.no_tensors:
+        return
+    print(f'\n* Dumping {len(reader.tensors)} tensor(s)')
+    for n, tensor in enumerate(reader.tensors, 1):
+        prettydims = ', '.join('{0:5}'.format(d) for d in list(tensor.shape) + [1] * (4 - len(tensor.shape)))
+        print(f'  {n:5}: {tensor.n_elements:10} | {prettydims} | {tensor.tensor_type.name:7} | {tensor.name}')
+
+
+def dump_metadata_json(reader: GGUFReader, args: argparse.Namespace) -> None:
+    import json
+    host_endian, file_endian = get_file_host_endian(reader)
+    metadata: dict[str, Any] = {}
+    tensors: dict[str, Any] = {}
+    result = {
+        "filename": args.model,
+        "endian": file_endian,
+        "metadata": metadata,
+        "tensors": tensors,
+    }
+    for idx, field in enumerate(reader.fields.values()):
+        curr: dict[str, Any] = {
+            "index": idx,
+            "type": field.types[0].name if field.types else 'UNKNOWN',
+            "offset": field.offset,
+        }
+        metadata[field.name] = curr
+        if field.types[:1] == [GGUFValueType.ARRAY]:
+            curr["array_types"] = [t.name for t in field.types][1:]
+            if not args.json_array:
+                continue
+            itype = field.types[-1]
+            if itype == GGUFValueType.STRING:
+                curr["value"] = [str(bytes(field.parts[idx]), encoding="utf-8") for idx in field.data]
+            else:
+                curr["value"] = [pv for idx in field.data for pv in field.parts[idx].tolist()]
+        elif field.types[0] == GGUFValueType.STRING:
+            curr["value"] = str(bytes(field.parts[-1]), encoding="utf-8")
+        else:
+            curr["value"] = field.parts[-1].tolist()[0]
+    for idx, tensor in enumerate(reader.tensors):
+        tensors[tensor.name] = {
+            "index": idx,
+            "shape": tensor.shape.tolist(),
+            "type": tensor.tensor_type.name,
+            "offset": tensor.field.offset,
+        }
+    json.dump(result, sys.stdout)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Dump GGUF file metadata")
+    parser.add_argument("model",           type=str,            help="GGUF format model filename")
+    parser.add_argument("--no-tensors", action="store_true", help="Don't dump tensor metadata")
+    parser.add_argument("--json",       action="store_true", help="Produce JSON output")
+    parser.add_argument("--json-array", action="store_true", help="Include full array values in JSON output (long)")
+    args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"])
+    if not args.json:
+        print(f'* Loading: {args.model}')
+    reader = GGUFReader(args.model, 'r')
+    if args.json:
+        dump_metadata_json(reader, args)
+    else:
+        dump_metadata(reader, args)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/gguf-py/scripts/gguf-set-metadata.py b/gguf-py/scripts/gguf-set-metadata.py
new file mode 100755
index 00000000..3ebdfa89
--- /dev/null
+++ b/gguf-py/scripts/gguf-set-metadata.py
@@ -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()