diff options
Diffstat (limited to 'llama.cpp/gguf-py/gguf/scripts')
| -rwxr-xr-x | llama.cpp/gguf-py/gguf/scripts/gguf_convert_endian.py | 186 | ||||
| -rwxr-xr-x | llama.cpp/gguf-py/gguf/scripts/gguf_dump.py | 477 | ||||
| -rwxr-xr-x | llama.cpp/gguf-py/gguf/scripts/gguf_editor_gui.py | 1621 | ||||
| -rwxr-xr-x | llama.cpp/gguf-py/gguf/scripts/gguf_hash.py | 102 | ||||
| -rwxr-xr-x | llama.cpp/gguf-py/gguf/scripts/gguf_new_metadata.py | 216 | ||||
| -rwxr-xr-x | llama.cpp/gguf-py/gguf/scripts/gguf_set_metadata.py | 95 |
6 files changed, 2697 insertions, 0 deletions
diff --git a/llama.cpp/gguf-py/gguf/scripts/gguf_convert_endian.py b/llama.cpp/gguf-py/gguf/scripts/gguf_convert_endian.py new file mode 100755 index 0000000..86bf878 --- /dev/null +++ b/llama.cpp/gguf-py/gguf/scripts/gguf_convert_endian.py @@ -0,0 +1,186 @@ +#!/usr/bin/env python3 +from __future__ import annotations + +import logging +import argparse +import os +import sys +from tqdm import tqdm +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.parent / 'gguf-py').exists(): + sys.path.insert(0, str(Path(__file__).parent.parent.parent)) + +import gguf + +logger = logging.getLogger("gguf-convert-endian") + + +def byteswap_noop(tensor, block_offs): + # this function is used when byteswapping is not needed + pass + + +def byteswap_q4_0(tensor, block_offs): + # Each block_q4_0 consists of an f16 delta (scaling factor) followed by 16 int8 quantizations. + + # Byte-Swap f16 sized delta field + delta = tensor.data[block_offs:block_offs + 2].view(dtype=np.uint16) + delta.byteswap(inplace=True) + + +def byteswap_q8_0(tensor, block_offs): + # Each block_q8_0 consists of an f16 delta (scaling factor) followed by 32 int8 quantizations. + + # Byte-Swap f16 sized delta field + delta = tensor.data[block_offs:block_offs + 2].view(dtype=np.uint16) + delta.byteswap(inplace=True) + + +def byteswap_q4_k(tensor, block_offs): + # Each block_q4_k consists of 2 f16 values followed by 140 int8 values. + + # Byte-Swap f16 sized fields + delta = tensor.data[block_offs:block_offs + 2].view(dtype=np.uint16) + delta.byteswap(inplace=True) + + delta = tensor.data[block_offs + 2:block_offs + 4].view(dtype=np.uint16) + delta.byteswap(inplace=True) + + +def byteswap_q6_k(tensor, block_offs): + # Each block_q6_k consists of 208 int8 values followed by 1 f16 value. + + # Byte-Swap f16 sized field + delta = tensor.data[block_offs + 208:block_offs + 210].view(dtype=np.uint16) + delta.byteswap(inplace=True) + + +byteswap_tensors = { + gguf.GGMLQuantizationType.Q4_0: byteswap_q4_0, + gguf.GGMLQuantizationType.Q8_0: byteswap_q8_0, + gguf.GGMLQuantizationType.Q4_K: byteswap_q4_k, + gguf.GGMLQuantizationType.Q6_K: byteswap_q6_k, + gguf.GGMLQuantizationType.MXFP4: byteswap_noop, +} + + +def convert_byteorder(reader: gguf.GGUFReader, args: argparse.Namespace) -> None: + file_endian = reader.endianess.name + if reader.byte_order == 'S': + host_endian = 'BIG' if file_endian == 'LITTLE' else 'LITTLE' + else: + host_endian = file_endian + order = host_endian if args.order == "native" else args.order.upper() + logger.info(f"* Host is {host_endian} endian, GGUF file seems to be {file_endian} endian") + if file_endian == order: + logger.info(f"* File is already {order} endian. Nothing to do.") + sys.exit(0) + logger.info("* Checking tensors for conversion compatibility") + for tensor in reader.tensors: + if tensor.tensor_type not in byteswap_tensors and \ + tensor.tensor_type not in ( + gguf.GGMLQuantizationType.F32, + gguf.GGMLQuantizationType.F16, + gguf.GGMLQuantizationType.BF16, + ): + raise ValueError(f"Cannot handle type {tensor.tensor_type.name} for tensor {repr(tensor.name)}") + logger.info(f"* Preparing to convert from {file_endian} to {order}") + if args.dry_run: + return + logger.warning("*** Warning *** Warning *** Warning **") + logger.warning("* This conversion process may damage the file. Ensure you have a backup.") + if order != host_endian: + logger.warning("* Requested endian differs from host, you will not be able to load the model on this machine.") + logger.warning("* The file will be modified immediately, so if conversion fails or is interrupted") + logger.warning("* 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": + logger.warning("You didn't enter YES. Okay then, see ya!") + sys.exit(0) + logger.info(f"* Converting fields ({len(reader.fields)})") + for idx, field in enumerate(reader.fields.values()): + logger.info(f"- {idx:4}: Converting field {repr(field.name)}, part count: {len(field.parts)}") + for part in field.parts: + part.byteswap(inplace=True) + logger.info(f"* Converting tensors ({len(reader.tensors)})") + + for idx, tensor in enumerate(pbar := tqdm(reader.tensors, desc="Converting tensor")): + log_message = ( + f"Converting tensor {repr(tensor.name)}, " + f"type={tensor.tensor_type.name}, " + f"elements={tensor.n_elements} " + ) + + # Byte-swap each part of the tensor's field + for part in tensor.field.parts: + part.byteswap(inplace=True) + + # Byte-swap tensor data if necessary + if tensor.tensor_type in byteswap_tensors: + # first flatten structure + oldshape = tensor.data.shape + newshape = 1 + for i in tensor.data.shape: + newshape *= i + + tensor.data.resize(newshape) + + block_size = gguf.constants.GGML_QUANT_SIZES[tensor.tensor_type][1] + byteswap_func = byteswap_tensors[tensor.tensor_type] + + n_blocks = len(tensor.data) // block_size + for block_num in (inner_pbar := tqdm(range(n_blocks), desc="Byte-swapping Blocks", leave=False)): + block_offs = block_num * block_size + + byteswap_func(tensor, block_offs) + + if block_num % 100000 == 0: + inner_pbar.set_description(f"Byte-swapping Blocks [{(n_blocks - block_num) // n_blocks}]") + + # restore old shape in case it's ever used + tensor.data.resize(oldshape) + elif tensor.tensor_type == gguf.GGMLQuantizationType.BF16: + # Special case for BF16 + # It is 2-bytes data, but by default view loads it as 1-byte data. + # Change to correct view before byteswapping. + tensor.data.view(dtype=np.uint16).byteswap(inplace=True) + else: + # Handle other tensor types + tensor.data.byteswap(inplace=True) + + pbar.set_description(log_message) + + logger.info("* 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", + ) + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") + + args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"]) + + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + + logger.info(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/llama.cpp/gguf-py/gguf/scripts/gguf_dump.py b/llama.cpp/gguf-py/gguf/scripts/gguf_dump.py new file mode 100755 index 0000000..8177dff --- /dev/null +++ b/llama.cpp/gguf-py/gguf/scripts/gguf_dump.py @@ -0,0 +1,477 @@ +#!/usr/bin/env python3 +from __future__ import annotations + +import logging +import argparse +import os +import re +import sys +from pathlib import Path +from typing import Any + +# Necessary to load the local gguf package +if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent.parent / 'gguf-py').exists(): + sys.path.insert(0, str(Path(__file__).parent.parent.parent)) + +from gguf import GGUFReader, GGUFValueType, ReaderTensor # noqa: E402 + +logger = logging.getLogger("gguf-dump") + + +def get_file_host_endian(reader: GGUFReader) -> tuple[str, str]: + file_endian = reader.endianess.name + if reader.byte_order == 'S': + host_endian = 'BIG' if file_endian == 'LITTLE' else 'LITTLE' + else: + host_endian = file_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.') # noqa: NP100 + print(f'* Dumping {len(reader.fields)} key/value pair(s)') # noqa: NP100 + 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) + + log_message = f' {n:5}: {pretty_type:10} | {len(field.data):8} | {field.name}' + if field.types: + curr_type = field.types[0] + if curr_type == GGUFValueType.STRING: + content = field.contents() + if len(content) > 60: + content = content[:57] + '...' + log_message += ' = {0}'.format(repr(content)) + elif curr_type in reader.gguf_scalar_to_np: + log_message += ' = {0}'.format(field.contents()) + else: + content = repr(field.contents(slice(6))) + if len(field.data) > 6: + content = content[:-1] + ', ...]' + log_message += ' = {0}'.format(content) + print(log_message) # noqa: NP100 + if args.no_tensors: + return + print(f'* Dumping {len(reader.tensors)} tensor(s)') # noqa: NP100 + 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}') # noqa: NP100 + + +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 + curr["value"] = field.contents() + else: + curr["value"] = field.contents() + if not args.no_tensors: + 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 markdown_table_with_alignment_support(header_map: list[dict[str, str]], data: list[dict[str, Any]]): + # JSON to Markdown table formatting: https://stackoverflow.com/a/72983854/2850957 + + # Alignment Utility Function + def strAlign(padding: int, alignMode: str | None, strVal: str): + if alignMode == 'center': + return strVal.center(padding) + elif alignMode == 'right': + return strVal.rjust(padding - 1) + ' ' + elif alignMode == 'left': + return ' ' + strVal.ljust(padding - 1) + else: # default left + return ' ' + strVal.ljust(padding - 1) + + def dashAlign(padding: int, alignMode: str | None): + if alignMode == 'center': + return ':' + '-' * (padding - 2) + ':' + elif alignMode == 'right': + return '-' * (padding - 1) + ':' + elif alignMode == 'left': + return ':' + '-' * (padding - 1) + else: # default left + return '-' * (padding) + + # Calculate Padding For Each Column Based On Header and Data Length + rowsPadding = {} + for index, columnEntry in enumerate(header_map): + padCount = max([len(str(v)) for d in data for k, v in d.items() if k == columnEntry['key_name']], default=0) + 2 + headerPadCount = len(columnEntry['header_name']) + 2 + rowsPadding[index] = headerPadCount if padCount <= headerPadCount else padCount + + # Render Markdown Header + rows = [] + rows.append('|'.join(strAlign(rowsPadding[index], columnEntry.get('align'), str(columnEntry['header_name'])) for index, columnEntry in enumerate(header_map))) + rows.append('|'.join(dashAlign(rowsPadding[index], columnEntry.get('align')) for index, columnEntry in enumerate(header_map))) + + # Render Tabular Data + for item in data: + rows.append('|'.join(strAlign(rowsPadding[index], columnEntry.get('align'), str(item[columnEntry['key_name']])) for index, columnEntry in enumerate(header_map))) + + # Convert Tabular String Rows Into String + tableString = "" + for row in rows: + tableString += f'|{row}|\n' + + return tableString + + +def element_count_rounded_notation(count: int) -> str: + if count > 1e15 : + # Quadrillion + scaled_amount = count * 1e-15 + scale_suffix = "Q" + elif count > 1e12 : + # Trillions + scaled_amount = count * 1e-12 + scale_suffix = "T" + elif count > 1e9 : + # Billions + scaled_amount = count * 1e-9 + scale_suffix = "B" + elif count > 1e6 : + # Millions + scaled_amount = count * 1e-6 + scale_suffix = "M" + elif count > 1e3 : + # Thousands + scaled_amount = count * 1e-3 + scale_suffix = "K" + else: + # Under Thousands + scaled_amount = count + scale_suffix = "" + return f"{'~' if count > 1e3 else ''}{round(scaled_amount)}{scale_suffix}" + + +def translate_tensor_name(name): + words = name.split(".") + + # Source: https://github.com/ggml-org/ggml/blob/master/docs/gguf.md#standardized-tensor-names + abbreviation_dictionary = { + 'token_embd': 'Token embedding', + 'pos_embd': 'Position embedding', + 'output_norm': 'Output normalization', + 'output': 'Output', + 'attn_norm': 'Attention normalization', + 'attn_norm_2': 'Attention normalization', + 'attn_qkv': 'Attention query-key-value', + 'attn_q': 'Attention query', + 'attn_k': 'Attention key', + 'attn_v': 'Attention value', + 'attn_output': 'Attention output', + 'ffn_norm': 'Feed-forward network normalization', + 'ffn_up': 'Feed-forward network "up"', + 'ffn_gate': 'Feed-forward network "gate"', + 'ffn_down': 'Feed-forward network "down"', + 'ffn_gate_inp': 'Expert-routing layer for the Feed-forward network in Mixture of Expert models', + 'ffn_gate_exp': 'Feed-forward network "gate" layer per expert in Mixture of Expert models', + 'ffn_down_exp': 'Feed-forward network "down" layer per expert in Mixture of Expert models', + 'ffn_up_exp': 'Feed-forward network "up" layer per expert in Mixture of Expert models', + 'ssm_in': 'State space model input projections', + 'ssm_conv1d': 'State space model rolling/shift', + 'ssm_x': 'State space model selective parametrization', + 'ssm_a': 'State space model state compression', + 'ssm_d': 'State space model skip connection', + 'ssm_dt': 'State space model time step', + 'ssm_out': 'State space model output projection', + 'blk': 'Block', + 'enc': 'Encoder', + 'dec': 'Decoder', + } + + expanded_words = [] + for word in words: + word_norm = word.strip().lower() + if word_norm in abbreviation_dictionary: + expanded_words.append(abbreviation_dictionary[word_norm].title()) + else: + expanded_words.append(word.title()) + + return ' '.join(expanded_words) + + +def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None: + host_endian, file_endian = get_file_host_endian(reader) + markdown_content = "" + markdown_content += f'# {args.model} - GGUF Internal File Dump\n\n' + markdown_content += f'- Endian: {file_endian} endian\n' + markdown_content += '\n' + markdown_content += '## Key Value Metadata Store\n\n' + markdown_content += f'There are {len(reader.fields)} key-value pairs in this file\n' + markdown_content += '\n' + total_model_bytes = 0 + total_model_elements = 0 + + kv_dump_table: list[dict[str, str | int]] = [] + 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) + + def escape_markdown_inline_code(value_string): + # Find the longest contiguous sequence of backticks in the string then + # wrap string with appropriate number of backticks required to escape it + max_backticks = max((len(match.group(0)) for match in re.finditer(r'`+', value_string)), default=0) + inline_code_marker = '`' * (max_backticks + 1) + + # If the string starts or ends with a backtick, add a space at the beginning and end + if value_string.startswith('`') or value_string.endswith('`'): + value_string = f" {value_string} " + + return f"{inline_code_marker}{value_string}{inline_code_marker}" + + total_elements = len(field.data) + value = "" + if len(field.types) == 1: + curr_type = field.types[0] + if curr_type == GGUFValueType.STRING: + truncate_length = 60 + value_string = str(bytes(field.parts[-1]), encoding='utf-8') + if len(value_string) > truncate_length: + head = escape_markdown_inline_code(value_string[:truncate_length // 2]) + tail = escape_markdown_inline_code(value_string[-truncate_length // 2:]) + value = "{head}...{tail}".format(head=head, tail=tail) + else: + value = escape_markdown_inline_code(value_string) + elif curr_type in reader.gguf_scalar_to_np: + value = str(field.parts[-1][0]) + else: + if field.types[0] == GGUFValueType.ARRAY: + curr_type = field.types[1] + array_elements = [] + + if curr_type == GGUFValueType.STRING: + render_element = min(5, total_elements) + for element_pos in range(render_element): + truncate_length = 30 + value_string = str(bytes(field.parts[-1 - (total_elements - element_pos - 1) * 2]), encoding='utf-8') + if len(value_string) > truncate_length: + head = escape_markdown_inline_code(value_string[:truncate_length // 2]) + tail = escape_markdown_inline_code(value_string[-truncate_length // 2:]) + value = "{head}...{tail}".format(head=head, tail=tail) + else: + value = escape_markdown_inline_code(value_string) + array_elements.append(value) + + elif curr_type in reader.gguf_scalar_to_np: + render_element = min(7, total_elements) + for element_pos in range(render_element): + array_elements.append(str(field.parts[-1 - (total_elements - element_pos - 1)][0])) + + value = f'[ {", ".join(array_elements).strip()}{", ..." if total_elements > len(array_elements) else ""} ]' + + kv_dump_table.append({"n":n, "pretty_type":pretty_type, "total_elements":total_elements, "field_name":field.name, "value":value}) + + kv_dump_table_header_map = [ + {'key_name':'n', 'header_name':'POS', 'align':'right'}, + {'key_name':'pretty_type', 'header_name':'TYPE', 'align':'left'}, + {'key_name':'total_elements', 'header_name':'Count', 'align':'right'}, + {'key_name':'field_name', 'header_name':'Key', 'align':'left'}, + {'key_name':'value', 'header_name':'Value', 'align':'left'}, + ] + + markdown_content += markdown_table_with_alignment_support(kv_dump_table_header_map, kv_dump_table) + + markdown_content += "\n" + + if not args.no_tensors: + # Group tensors by their prefix and maintain order + tensor_prefix_order: list[str] = [] + tensor_name_to_key: dict[str, int] = {} + tensor_groups: dict[str, list[ReaderTensor]] = {} + total_elements = sum(tensor.n_elements for tensor in reader.tensors) + + # Parsing Tensors Record + for key, tensor in enumerate(reader.tensors): + tensor_components = tensor.name.split('.') + + # Classify Tensor Group + tensor_group_name = "base" + if tensor_components[0] == 'blk': + tensor_group_name = f"{tensor_components[0]}.{tensor_components[1]}" + elif tensor_components[0] in ['enc', 'dec'] and tensor_components[1] == 'blk': + tensor_group_name = f"{tensor_components[0]}.{tensor_components[1]}.{tensor_components[2]}" + elif tensor_components[0] in ['enc', 'dec']: + tensor_group_name = f"{tensor_components[0]}" + + # Check if new Tensor Group + if tensor_group_name not in tensor_groups: + tensor_groups[tensor_group_name] = [] + tensor_prefix_order.append(tensor_group_name) + + # Record Tensor and Tensor Position + tensor_groups[tensor_group_name].append(tensor) + tensor_name_to_key[tensor.name] = key + + # Tensors Mapping Dump + markdown_content += f'## Tensors Overview {element_count_rounded_notation(total_elements)} Elements\n\n' + markdown_content += f'Total number of elements in all tensors: {total_elements} Elements\n' + markdown_content += '\n' + + for group in tensor_prefix_order: + tensors = tensor_groups[group] + group_elements = sum(tensor.n_elements for tensor in tensors) + markdown_content += f"- [{translate_tensor_name(group)} Tensor Group - {element_count_rounded_notation(group_elements)} Elements](#{group.replace('.', '_')})\n" + + markdown_content += "\n" + + markdown_content += "### Tensor Data Offset\n" + markdown_content += '\n' + markdown_content += 'This table contains the offset and data segment relative to start of file\n' + markdown_content += '\n' + + tensor_mapping_table: list[dict[str, str | int]] = [] + for key, tensor in enumerate(reader.tensors): + data_offset_pretty = '{0:#16x}'.format(tensor.data_offset) + data_size_pretty = '{0:#16x}'.format(tensor.n_bytes) + tensor_mapping_table.append({"t_id":key, "layer_name":tensor.name, "data_offset":data_offset_pretty, "data_size":data_size_pretty}) + + tensors_mapping_table_header_map = [ + {'key_name':'t_id', 'header_name':'T_ID', 'align':'right'}, + {'key_name':'layer_name', 'header_name':'Tensor Layer Name', 'align':'left'}, + {'key_name':'data_offset', 'header_name':'Data Offset (B)', 'align':'right'}, + {'key_name':'data_size', 'header_name':'Data Size (B)', 'align':'right'}, + ] + + markdown_content += markdown_table_with_alignment_support(tensors_mapping_table_header_map, tensor_mapping_table) + markdown_content += "\n" + + for group in tensor_prefix_order: + tensors = tensor_groups[group] + group_elements = sum(tensor.n_elements for tensor in tensors) + group_percentage = group_elements / total_elements * 100 + total_group_bytes = 0 + total_group_elements = 0 + markdown_content += f"### <a name=\"{group.replace('.', '_')}\">{translate_tensor_name(group)} Tensor Group : {element_count_rounded_notation(group_elements)} Elements</a>\n\n" + + # Precalculate column sizing for visual consistency + prettify_element_est_count_size: int = 1 + prettify_element_count_size: int = 1 + prettify_dimension_max_widths: dict[int, int] = {} + for tensor in tensors: + prettify_element_est_count_size = max(prettify_element_est_count_size, len(str(element_count_rounded_notation(tensor.n_elements)))) + prettify_element_count_size = max(prettify_element_count_size, len(str(tensor.n_elements))) + for i, dimension_size in enumerate(list(tensor.shape) + [1] * (4 - len(tensor.shape))): + prettify_dimension_max_widths[i] = max(prettify_dimension_max_widths.get(i,1), len(str(dimension_size))) + + # Generate Tensor Layer Table Content + tensor_dump_table: list[dict[str, str | int]] = [] + for tensor in tensors: + human_friendly_name = translate_tensor_name(tensor.name.replace(".weight", ".(W)").replace(".bias", ".(B)")) + pretty_dimension = ' x '.join(f'{str(d):>{prettify_dimension_max_widths[i]}}' for i, d in enumerate(list(tensor.shape) + [1] * (4 - len(tensor.shape)))) + element_count_est = f"({element_count_rounded_notation(tensor.n_elements):>{prettify_element_est_count_size}})" + element_count_string = f"{element_count_est} {tensor.n_elements:>{prettify_element_count_size}}" + type_name_string = f"{tensor.tensor_type.name}" + if tensor.n_elements > 0: + bpw = (tensor.n_bytes * 8) / tensor.n_elements + else: + bpw = float('nan') + tensor_dump_table.append({"t_id":tensor_name_to_key[tensor.name], "layer_name":tensor.name, "human_layer_name":human_friendly_name, "element_count":element_count_string, "pretty_dimension":pretty_dimension, "tensor_type":type_name_string, "bpw": f"{bpw:.4f}"}) + total_group_bytes += tensor.n_bytes + total_group_elements += tensor.n_elements + + tensor_dump_table_header_map = [ + {'key_name':'t_id', 'header_name':'T_ID', 'align':'right'}, + {'key_name':'layer_name', 'header_name':'Tensor Layer Name', 'align':'left'}, + {'key_name':'human_layer_name', 'header_name':'Human Friendly Tensor Layer Name', 'align':'left'}, + {'key_name':'element_count', 'header_name':'Elements', 'align':'left'}, + {'key_name':'pretty_dimension', 'header_name':'Shape', 'align':'left'}, + {'key_name':'tensor_type', 'header_name':'Type', 'align':'left'}, + {'key_name':'bpw', 'header_name':'BPW', 'align':'right'}, + ] + + markdown_content += markdown_table_with_alignment_support(tensor_dump_table_header_map, tensor_dump_table) + + markdown_content += "\n" + markdown_content += f"- Total elements in {group}: ({element_count_rounded_notation(group_elements):>4}) {group_elements}\n" + markdown_content += f"- Percentage of total elements: {group_percentage:.2f}%\n" + if total_group_elements > 0: + total_group_bpw = (total_group_bytes * 8) / total_group_elements + markdown_content += f"- Bits per Weight (BPW) for {group}: {total_group_bpw:.4f} bits\n" + else: + markdown_content += f"- Bits per Weight (BPW) for {group}: undefined (no elements)\n" + markdown_content += "\n\n" + total_model_bytes += total_group_bytes + total_model_elements += total_group_elements + + if total_model_elements > 0: + total_model_bpw = (total_model_bytes * 8) / total_model_elements + markdown_content += f"Total BPW for {os.path.basename(args.model)}: {total_model_bpw:.4f} bits" + else: + markdown_content += f"Total BPW for {os.path.basename(args.model)}: undefined (no elements)" + print(markdown_content) # noqa: NP100 + + +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)") + parser.add_argument("--data-offset", action="store_true", help="Start of data offset") + parser.add_argument("--data-alignment", action="store_true", help="Data alignment applied globally to data field") + parser.add_argument("--markdown", action="store_true", help="Produce markdown output") + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") + + args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"]) + + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + + if not args.json and not args.markdown and not args.data_offset and not args.data_alignment: + logger.info(f'* Loading: {args.model}') + + reader = GGUFReader(args.model, 'r') + + if args.json: + dump_metadata_json(reader, args) + elif args.markdown: + dump_markdown_metadata(reader, args) + elif args.data_offset: + print(reader.data_offset) # noqa: NP100 + elif args.data_alignment: + print(reader.alignment) # noqa: NP100 + else: + dump_metadata(reader, args) + + +if __name__ == '__main__': + main() diff --git a/llama.cpp/gguf-py/gguf/scripts/gguf_editor_gui.py b/llama.cpp/gguf-py/gguf/scripts/gguf_editor_gui.py new file mode 100755 index 0000000..293316a --- /dev/null +++ b/llama.cpp/gguf-py/gguf/scripts/gguf_editor_gui.py @@ -0,0 +1,1621 @@ +#!/usr/bin/env python3 +from __future__ import annotations + +import logging +import argparse +import os +import sys +import numpy +import enum +from pathlib import Path +from typing import Any, Optional, Tuple, Type +import warnings + +import numpy as np +from PySide6.QtWidgets import ( + QApplication, QMainWindow, QWidget, QVBoxLayout, QHBoxLayout, + QPushButton, QLabel, QLineEdit, QFileDialog, QTableWidget, + QTableWidgetItem, QComboBox, QMessageBox, QTabWidget, + QTextEdit, QFormLayout, + QHeaderView, QDialog, QDialogButtonBox +) +from PySide6.QtCore import Qt + +# Necessary to load the local gguf package +if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent.parent / 'gguf-py').exists(): + sys.path.insert(0, str(Path(__file__).parent.parent.parent)) + +import gguf +from gguf import GGUFReader, GGUFWriter, GGUFValueType, ReaderField +from gguf.constants import TokenType, RopeScalingType, PoolingType, GGMLQuantizationType + +logger = logging.getLogger("gguf-editor-gui") + +# Map of key names to enum types for automatic enum interpretation +KEY_TO_ENUM_TYPE = { + gguf.Keys.Tokenizer.TOKEN_TYPE: TokenType, + gguf.Keys.Rope.SCALING_TYPE: RopeScalingType, + gguf.Keys.LLM.POOLING_TYPE: PoolingType, + gguf.Keys.General.FILE_TYPE: GGMLQuantizationType, +} + +# Define the tokenizer keys that should be edited together +TOKENIZER_LINKED_KEYS = [ + gguf.Keys.Tokenizer.LIST, + gguf.Keys.Tokenizer.TOKEN_TYPE, + gguf.Keys.Tokenizer.SCORES +] + + +class TokenizerEditorDialog(QDialog): + def __init__(self, tokens, token_types, scores, parent=None): + super().__init__(parent) + self.setWindowTitle("Edit Tokenizer Data") + self.resize(900, 600) + + self.tokens = tokens.copy() if tokens else [] + self.token_types = token_types.copy() if token_types else [] + self.scores = scores.copy() if scores else [] + + # Ensure all arrays have the same length + max_len = max(len(self.tokens), len(self.token_types), len(self.scores)) + if len(self.tokens) < max_len: + self.tokens.extend([""] * (max_len - len(self.tokens))) + if len(self.token_types) < max_len: + self.token_types.extend([0] * (max_len - len(self.token_types))) + if len(self.scores) < max_len: + self.scores.extend([0.0] * (max_len - len(self.scores))) + + layout = QVBoxLayout(self) + + # Add filter controls + filter_layout = QHBoxLayout() + filter_layout.addWidget(QLabel("Filter:")) + self.filter_edit = QLineEdit() + self.filter_edit.setPlaceholderText("Type to filter tokens...") + self.filter_edit.textChanged.connect(self.apply_filter) + filter_layout.addWidget(self.filter_edit) + + # Add page controls + self.page_size = 100 # Show 100 items per page + self.current_page = 0 + self.total_pages = max(1, (len(self.tokens) + self.page_size - 1) // self.page_size) + + self.page_label = QLabel(f"Page 1 of {self.total_pages}") + filter_layout.addWidget(self.page_label) + + prev_page = QPushButton("Previous") + prev_page.clicked.connect(self.previous_page) + filter_layout.addWidget(prev_page) + + next_page = QPushButton("Next") + next_page.clicked.connect(self.next_page) + filter_layout.addWidget(next_page) + + layout.addLayout(filter_layout) + + # Tokenizer data table + self.tokens_table = QTableWidget() + self.tokens_table.setColumnCount(4) + self.tokens_table.setHorizontalHeaderLabels(["Index", "Token", "Type", "Score"]) + self.tokens_table.horizontalHeader().setSectionResizeMode(0, QHeaderView.ResizeMode.ResizeToContents) + self.tokens_table.horizontalHeader().setSectionResizeMode(1, QHeaderView.ResizeMode.Stretch) + self.tokens_table.horizontalHeader().setSectionResizeMode(2, QHeaderView.ResizeMode.ResizeToContents) + self.tokens_table.horizontalHeader().setSectionResizeMode(3, QHeaderView.ResizeMode.ResizeToContents) + + layout.addWidget(self.tokens_table) + + # Controls + controls_layout = QHBoxLayout() + + add_button = QPushButton("Add Token") + add_button.clicked.connect(self.add_token) + controls_layout.addWidget(add_button) + + remove_button = QPushButton("Remove Selected") + remove_button.clicked.connect(self.remove_selected) + controls_layout.addWidget(remove_button) + + controls_layout.addStretch() + + layout.addLayout(controls_layout) + + # Buttons + buttons = QDialogButtonBox(QDialogButtonBox.StandardButton.Ok | QDialogButtonBox.StandardButton.Cancel) + buttons.accepted.connect(self.accept) + buttons.rejected.connect(self.reject) + layout.addWidget(buttons) + + # Initialize the filtered values + self.filtered_indices = list(range(len(self.tokens))) + + # Load data for the first page + self.load_page() + + def apply_filter(self): + """Filter the tokens based on the search text.""" + filter_text = self.filter_edit.text().lower() + + if not filter_text: + # No filter, show all values + self.filtered_indices = list(range(len(self.tokens))) + else: + # Apply filter + self.filtered_indices = [] + for i, token in enumerate(self.tokens): + if filter_text in str(token).lower(): + self.filtered_indices.append(i) + + # Reset to first page and reload + self.total_pages = max(1, (len(self.filtered_indices) + self.page_size - 1) // self.page_size) + self.current_page = 0 + self.page_label.setText(f"Page 1 of {self.total_pages}") + self.load_page() + + def previous_page(self): + """Go to the previous page of results.""" + if self.current_page > 0: + self.current_page -= 1 + self.page_label.setText(f"Page {self.current_page + 1} of {self.total_pages}") + self.load_page() + + def next_page(self): + """Go to the next page of results.""" + if self.current_page < self.total_pages - 1: + self.current_page += 1 + self.page_label.setText(f"Page {self.current_page + 1} of {self.total_pages}") + self.load_page() + + def load_page(self): + """Load the current page of tokenizer data.""" + self.tokens_table.setRowCount(0) # Clear the table + + # Calculate start and end indices for the current page + start_idx = self.current_page * self.page_size + end_idx = min(start_idx + self.page_size, len(self.filtered_indices)) + + # Pre-allocate rows for better performance + self.tokens_table.setRowCount(end_idx - start_idx) + + for row, i in enumerate(range(start_idx, end_idx)): + orig_idx = self.filtered_indices[i] + + # Index + index_item = QTableWidgetItem(str(orig_idx)) + index_item.setData(Qt.ItemDataRole.UserRole, orig_idx) # Store original index + index_item.setFlags(index_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.tokens_table.setItem(row, 0, index_item) + + # Token + token_item = QTableWidgetItem(str(self.tokens[orig_idx])) + self.tokens_table.setItem(row, 1, token_item) + + # Token Type + token_type = self.token_types[orig_idx] if orig_idx < len(self.token_types) else 0 + try: + enum_val = TokenType(token_type) + display_text = f"{enum_val.name} ({token_type})" + except (ValueError, KeyError): + display_text = f"Unknown ({token_type})" + + type_item = QTableWidgetItem(display_text) + type_item.setData(Qt.ItemDataRole.UserRole, token_type) + + # Make type cell editable with a double-click handler + type_item.setFlags(type_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.tokens_table.setItem(row, 2, type_item) + + # Score + score = self.scores[orig_idx] if orig_idx < len(self.scores) else 0.0 + score_item = QTableWidgetItem(str(score)) + self.tokens_table.setItem(row, 3, score_item) + + # Connect double-click handler for token type cells + self.tokens_table.cellDoubleClicked.connect(self.handle_cell_double_click) + + def handle_cell_double_click(self, row, column): + """Handle double-click on a cell, specifically for token type editing.""" + if column == 2: # Token Type column + orig_item = self.tokens_table.item(row, 0) + if orig_item: + orig_idx = orig_item.data(Qt.ItemDataRole.UserRole) + self.edit_token_type(row, orig_idx) + + def edit_token_type(self, row, orig_idx): + """Edit a token type using a dialog with a dropdown of all enum options.""" + current_value = self.token_types[orig_idx] if orig_idx < len(self.token_types) else 0 + + # Create a dialog with enum options + dialog = QDialog(self) + dialog.setWindowTitle("Select Token Type") + layout = QVBoxLayout(dialog) + + combo = QComboBox() + for enum_val in TokenType: + combo.addItem(f"{enum_val.name} ({enum_val.value})", enum_val.value) + + # Set current value + try: + if isinstance(current_value, int): + enum_val = TokenType(current_value) + combo.setCurrentText(f"{enum_val.name} ({current_value})") + except (ValueError, KeyError): + pass + + layout.addWidget(combo) + + buttons = QDialogButtonBox(QDialogButtonBox.StandardButton.Ok | QDialogButtonBox.StandardButton.Cancel) + buttons.accepted.connect(dialog.accept) + buttons.rejected.connect(dialog.reject) + layout.addWidget(buttons) + + if dialog.exec() == QDialog.DialogCode.Accepted: + # Get the selected value + new_value = combo.currentData() + enum_val = TokenType(new_value) + display_text = f"{enum_val.name} ({new_value})" + + # Update the display + type_item = self.tokens_table.item(row, 2) + if type_item: + type_item.setText(display_text) + type_item.setData(Qt.ItemDataRole.UserRole, new_value) + + # Update the actual value + self.token_types[orig_idx] = new_value + + def add_token(self): + """Add a new token to the end of the list.""" + # Add to the end of the arrays + self.tokens.append("") + self.token_types.append(0) # Default to normal token + self.scores.append(0.0) + + orig_idx = len(self.tokens) - 1 + + # Add to filtered indices if it matches the current filter + filter_text = self.filter_edit.text().lower() + if not filter_text or filter_text in "": + self.filtered_indices.append(orig_idx) + + # Update pagination + self.total_pages = max(1, (len(self.filtered_indices) + self.page_size - 1) // self.page_size) + + # Go to the last page to show the new item + self.current_page = self.total_pages - 1 + self.page_label.setText(f"Page {self.current_page + 1} of {self.total_pages}") + + # Reload the page + self.load_page() + + def remove_selected(self): + """Remove selected tokens from all arrays.""" + selected_rows = [] + for item in self.tokens_table.selectedItems(): + row = item.row() + if row not in selected_rows: + selected_rows.append(row) + + if not selected_rows: + return + + # Get original indices in descending order to avoid index shifting + orig_indices = [] + for row in selected_rows: + orig_item = self.tokens_table.item(row, 0) + if orig_item: + orig_indices.append(orig_item.data(Qt.ItemDataRole.UserRole)) + orig_indices.sort(reverse=True) + + # Remove from all arrays + for idx in orig_indices: + if idx < len(self.tokens): + del self.tokens[idx] + if idx < len(self.token_types): + del self.token_types[idx] + if idx < len(self.scores): + del self.scores[idx] + + # Rebuild filtered_indices + self.filtered_indices = [] + filter_text = self.filter_edit.text().lower() + + for i, token in enumerate(self.tokens): + if not filter_text or filter_text in str(token).lower(): + self.filtered_indices.append(i) + + # Update pagination + self.total_pages = max(1, (len(self.filtered_indices) + self.page_size - 1) // self.page_size) + self.current_page = min(self.current_page, self.total_pages - 1) + self.page_label.setText(f"Page {self.current_page + 1} of {self.total_pages}") + + # Reload the page + self.load_page() + + def get_data(self): + """Return the edited tokenizer data.""" + return self.tokens, self.token_types, self.scores + + +class ArrayEditorDialog(QDialog): + def __init__(self, array_values, element_type, key=None, parent=None): + super().__init__(parent) + self.setWindowTitle("Edit Array Values") + self.resize(700, 500) + + self.array_values = array_values + self.element_type = element_type + self.key = key + + # Get enum type for this array if applicable + self.enum_type = None + if key in KEY_TO_ENUM_TYPE and element_type == GGUFValueType.INT32: + self.enum_type = KEY_TO_ENUM_TYPE[key] + + layout = QVBoxLayout(self) + + # Add enum type information if applicable + if self.enum_type is not None: + enum_info_layout = QHBoxLayout() + enum_label = QLabel(f"Editing {self.enum_type.__name__} values:") + enum_info_layout.addWidget(enum_label) + + # Add a legend for the enum values + enum_values = ", ".join([f"{e.name}={e.value}" for e in self.enum_type]) + enum_values_label = QLabel(f"Available values: {enum_values}") + enum_values_label.setWordWrap(True) + enum_info_layout.addWidget(enum_values_label, 1) + + layout.addLayout(enum_info_layout) + + # Add search/filter controls + filter_layout = QHBoxLayout() + filter_layout.addWidget(QLabel("Filter:")) + self.filter_edit = QLineEdit() + self.filter_edit.setPlaceholderText("Type to filter values...") + self.filter_edit.textChanged.connect(self.apply_filter) + filter_layout.addWidget(self.filter_edit) + + # Add page controls for large arrays + self.page_size = 100 # Show 100 items per page + self.current_page = 0 + self.total_pages = max(1, (len(array_values) + self.page_size - 1) // self.page_size) + + self.page_label = QLabel(f"Page 1 of {self.total_pages}") + filter_layout.addWidget(self.page_label) + + prev_page = QPushButton("Previous") + prev_page.clicked.connect(self.previous_page) + filter_layout.addWidget(prev_page) + + next_page = QPushButton("Next") + next_page.clicked.connect(self.next_page) + filter_layout.addWidget(next_page) + + layout.addLayout(filter_layout) + + # Array items table + self.items_table = QTableWidget() + + # Set up columns based on whether we have an enum type + if self.enum_type is not None: + self.items_table.setColumnCount(3) + self.items_table.setHorizontalHeaderLabels(["Index", "Value", "Actions"]) + self.items_table.horizontalHeader().setSectionResizeMode(0, QHeaderView.ResizeMode.ResizeToContents) + self.items_table.horizontalHeader().setSectionResizeMode(1, QHeaderView.ResizeMode.Stretch) + self.items_table.horizontalHeader().setSectionResizeMode(2, QHeaderView.ResizeMode.ResizeToContents) + else: + self.items_table.setColumnCount(2) + self.items_table.setHorizontalHeaderLabels(["Index", "Value"]) + self.items_table.horizontalHeader().setSectionResizeMode(0, QHeaderView.ResizeMode.ResizeToContents) + self.items_table.horizontalHeader().setSectionResizeMode(1, QHeaderView.ResizeMode.Stretch) + + layout.addWidget(self.items_table) + + # Controls + controls_layout = QHBoxLayout() + + add_button = QPushButton("Add Item") + add_button.clicked.connect(self.add_item) + controls_layout.addWidget(add_button) + + remove_button = QPushButton("Remove Selected") + remove_button.clicked.connect(self.remove_selected) + controls_layout.addWidget(remove_button) + + # Add bulk edit button for enum arrays + if self.enum_type is not None: + bulk_edit_button = QPushButton("Bulk Edit Selected") + bulk_edit_button.clicked.connect(self.bulk_edit_selected) + controls_layout.addWidget(bulk_edit_button) + + controls_layout.addStretch() + + layout.addLayout(controls_layout) + + # Buttons + buttons = QDialogButtonBox(QDialogButtonBox.StandardButton.Ok | QDialogButtonBox.StandardButton.Cancel) + buttons.accepted.connect(self.accept) + buttons.rejected.connect(self.reject) + layout.addWidget(buttons) + + # Initialize the filtered values + self.filtered_indices = list(range(len(self.array_values))) + + # Load array values for the first page + self.load_page() + + def apply_filter(self): + """Filter the array values based on the search text.""" + filter_text = self.filter_edit.text().lower() + + if not filter_text: + # No filter, show all values + self.filtered_indices = list(range(len(self.array_values))) + else: + # Apply filter + self.filtered_indices = [] + for i, value in enumerate(self.array_values): + # For enum values, search in both name and value + if self.enum_type is not None and isinstance(value, int): + try: + enum_val = self.enum_type(value) + display_text = f"{enum_val.name} ({value})".lower() + if filter_text in display_text: + self.filtered_indices.append(i) + except (ValueError, KeyError): + # If not a valid enum value, just check the raw value + if filter_text in str(value).lower(): + self.filtered_indices.append(i) + else: + # For non-enum values, just check the string representation + if filter_text in str(value).lower(): + self.filtered_indices.append(i) + + # Reset to first page and reload + self.total_pages = max(1, (len(self.filtered_indices) + self.page_size - 1) // self.page_size) + self.current_page = 0 + self.page_label.setText(f"Page 1 of {self.total_pages}") + self.load_page() + + def previous_page(self): + """Go to the previous page of results.""" + if self.current_page > 0: + self.current_page -= 1 + self.page_label.setText(f"Page {self.current_page + 1} of {self.total_pages}") + self.load_page() + + def next_page(self): + """Go to the next page of results.""" + if self.current_page < self.total_pages - 1: + self.current_page += 1 + self.page_label.setText(f"Page {self.current_page + 1} of {self.total_pages}") + self.load_page() + + def load_page(self): + """Load the current page of array values.""" + self.items_table.setRowCount(0) # Clear the table + + # Calculate start and end indices for the current page + start_idx = self.current_page * self.page_size + end_idx = min(start_idx + self.page_size, len(self.filtered_indices)) + + # Pre-allocate rows for better performance + self.items_table.setRowCount(end_idx - start_idx) + + for row, i in enumerate(range(start_idx, end_idx)): + orig_idx = self.filtered_indices[i] + value = self.array_values[orig_idx] + + # Index + index_item = QTableWidgetItem(str(orig_idx)) + index_item.setData(Qt.ItemDataRole.UserRole, orig_idx) # Store original index + index_item.setFlags(index_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.items_table.setItem(row, 0, index_item) + + # Value + if self.enum_type is not None: + # Display enum value and name + try: + if isinstance(value, (int, numpy.signedinteger)): + enum_val = self.enum_type(value) + display_text = f"{enum_val.name} ({value})" + else: + display_text = str(value) + except (ValueError, KeyError): + display_text = f"Unknown ({value})" + + # Store the enum value in the item + value_item = QTableWidgetItem(display_text) + value_item.setData(Qt.ItemDataRole.UserRole, value) + value_item.setFlags(value_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.items_table.setItem(row, 1, value_item) + + # Add an edit button in a separate column + edit_button = QPushButton("Edit") + edit_button.setProperty("row", row) + edit_button.clicked.connect(self.edit_array_enum_value) + + # Create a widget to hold the button + button_widget = QWidget() + button_layout = QHBoxLayout(button_widget) + button_layout.setContentsMargins(2, 2, 2, 2) + button_layout.addWidget(edit_button) + button_layout.addStretch() + + self.items_table.setCellWidget(row, 2, button_widget) + else: + value_item = QTableWidgetItem(str(value)) + self.items_table.setItem(row, 1, value_item) + + def edit_array_enum_value(self): + """Handle editing an enum value in the array editor.""" + button = self.sender() + row = button.property("row") + + # Get the original index from the table item + orig_item = self.items_table.item(row, 0) + new_item = self.items_table.item(row, 1) + if orig_item and new_item and self.enum_type and self.edit_enum_value(row, self.enum_type): + orig_idx = orig_item.data(Qt.ItemDataRole.UserRole) + new_value = new_item.data(Qt.ItemDataRole.UserRole) + # Update the stored value in the array + if isinstance(new_value, (int, float, str, bool)): + self.array_values[orig_idx] = new_value + + def bulk_edit_selected(self): + """Edit multiple enum values at once.""" + if not self.enum_type: + return + + selected_rows = set() + for item in self.items_table.selectedItems(): + selected_rows.add(item.row()) + + if not selected_rows: + QMessageBox.information(self, "No Selection", "Please select at least one row to edit.") + return + + # Create a dialog with enum options + dialog = QDialog(self) + dialog.setWindowTitle(f"Bulk Edit {self.enum_type.__name__} Values") + layout = QVBoxLayout(dialog) + + layout.addWidget(QLabel(f"Set {len(selected_rows)} selected items to:")) + + combo = QComboBox() + for enum_val in self.enum_type: + combo.addItem(f"{enum_val.name} ({enum_val.value})", enum_val.value) + + layout.addWidget(combo) + + buttons = QDialogButtonBox(QDialogButtonBox.StandardButton.Ok | QDialogButtonBox.StandardButton.Cancel) + buttons.accepted.connect(dialog.accept) + buttons.rejected.connect(dialog.reject) + layout.addWidget(buttons) + + if dialog.exec() == QDialog.DialogCode.Accepted: + # Get the selected value + new_value = combo.currentData() + enum_val = self.enum_type(new_value) + display_text = f"{enum_val.name} ({new_value})" + + # Update all selected rows + for row in selected_rows: + orig_item = self.items_table.item(row, 0) + new_item = self.items_table.item(row, 1) + if orig_item and new_item: + orig_idx = orig_item.data(Qt.ItemDataRole.UserRole) + self.array_values[orig_idx] = new_value + + # Update the display + new_item.setText(display_text) + new_item.setData(Qt.ItemDataRole.UserRole, new_value) + + def add_item(self): + # Add to the end of the array + orig_idx = len(self.array_values) + + # Add default value based on type + if self.enum_type is not None: + # Default to first enum value + default_value = list(self.enum_type)[0].value + self.array_values.append(default_value) + else: + if self.element_type == GGUFValueType.STRING: + self.array_values.append("") + else: + self.array_values.append(0) + + # Add to filtered indices if it matches the current filter + self.filtered_indices.append(orig_idx) + + # Update pagination + self.total_pages = max(1, (len(self.filtered_indices) + self.page_size - 1) // self.page_size) + + # Go to the last page to show the new item + self.current_page = self.total_pages - 1 + self.page_label.setText(f"Page {self.current_page + 1} of {self.total_pages}") + + # Reload the page + self.load_page() + + def remove_selected(self): + selected_rows = [] + for item in self.items_table.selectedItems(): + row = item.row() + if row not in selected_rows: + selected_rows.append(row) + + if not selected_rows: + return + + # Get original indices in descending order to avoid index shifting + orig_indices = list() + for row in selected_rows: + orig_item = self.items_table.item(row, 0) + if orig_item: + orig_indices.append(orig_item.data(Qt.ItemDataRole.UserRole)) + orig_indices.sort(reverse=True) + + # Remove from array_values + for idx in orig_indices: + del self.array_values[idx] + + # Rebuild filtered_indices + self.filtered_indices = [] + filter_text = self.filter_edit.text().lower() + + for i, value in enumerate(self.array_values): + if not filter_text: + self.filtered_indices.append(i) + else: + # Apply filter + if self.enum_type is not None and isinstance(value, int): + try: + enum_val = self.enum_type(value) + display_text = f"{enum_val.name} ({value})".lower() + if filter_text in display_text: + self.filtered_indices.append(i) + except (ValueError, KeyError): + if filter_text in str(value).lower(): + self.filtered_indices.append(i) + else: + if filter_text in str(value).lower(): + self.filtered_indices.append(i) + + # Update pagination + self.total_pages = max(1, (len(self.filtered_indices) + self.page_size - 1) // self.page_size) + self.current_page = min(self.current_page, self.total_pages - 1) + self.page_label.setText(f"Page {self.current_page + 1} of {self.total_pages}") + + # Reload the page + self.load_page() + + def edit_enum_value(self, row: int, enum_type: Type[enum.Enum]): + """Edit an enum value using a dialog with a dropdown of all enum options.""" + # Get the original index from the table item + orig_item = self.items_table.item(row, 0) + if orig_item: + orig_idx = orig_item.data(Qt.ItemDataRole.UserRole) + else: + return + current_value = self.array_values[orig_idx] + + # Create a dialog with enum options + dialog = QDialog(self) + dialog.setWindowTitle(f"Select {enum_type.__name__} Value") + layout = QVBoxLayout(dialog) + + # Add description + description = QLabel(f"Select a {enum_type.__name__} value:") + layout.addWidget(description) + + # Use a combo box for quick selection + combo = QComboBox() + for enum_val in enum_type: + combo.addItem(f"{enum_val.name} ({enum_val.value})", enum_val.value) + + # Set current value + try: + if isinstance(current_value, int): + enum_val = enum_type(current_value) + combo.setCurrentText(f"{enum_val.name} ({current_value})") + except (ValueError, KeyError): + pass + + layout.addWidget(combo) + + buttons = QDialogButtonBox(QDialogButtonBox.StandardButton.Ok | QDialogButtonBox.StandardButton.Cancel) + buttons.accepted.connect(dialog.accept) + buttons.rejected.connect(dialog.reject) + layout.addWidget(buttons) + + if dialog.exec() == QDialog.DialogCode.Accepted: + # Update the value display and stored data + new_value = combo.currentData() + enum_val = enum_type(new_value) + display_text = f"{enum_val.name} ({new_value})" + + new_item = self.items_table.item(row, 1) + if new_item: + new_item.setText(display_text) + new_item.setData(Qt.ItemDataRole.UserRole, new_value) + + # Update the actual array value + self.array_values[orig_idx] = new_value + return True + return False + + def get_array_values(self): + # The array_values list is kept up-to-date as edits are made + return self.array_values + + +class AddMetadataDialog(QDialog): + def __init__(self, parent=None): + super().__init__(parent) + self.setWindowTitle("Add Metadata") + self.resize(400, 200) + + layout = QVBoxLayout(self) + + form_layout = QFormLayout() + + self.key_edit = QLineEdit() + form_layout.addRow("Key:", self.key_edit) + + self.type_combo = QComboBox() + for value_type in GGUFValueType: + if value_type != GGUFValueType.ARRAY: # Skip array type for simplicity + self.type_combo.addItem(value_type.name, value_type) + form_layout.addRow("Type:", self.type_combo) + + self.value_edit = QTextEdit() + form_layout.addRow("Value:", self.value_edit) + + layout.addLayout(form_layout) + + buttons = QDialogButtonBox(QDialogButtonBox.StandardButton.Ok | QDialogButtonBox.StandardButton.Cancel) + buttons.accepted.connect(self.accept) + buttons.rejected.connect(self.reject) + layout.addWidget(buttons) + + def get_data(self) -> Tuple[str, GGUFValueType, Any]: + key = self.key_edit.text() + value_type = self.type_combo.currentData() + value_text = self.value_edit.toPlainText() + + # Convert value based on type + if value_type == GGUFValueType.UINT8: + value = np.uint8(int(value_text)) + elif value_type == GGUFValueType.INT8: + value = np.int8(int(value_text)) + elif value_type == GGUFValueType.UINT16: + value = np.uint16(int(value_text)) + elif value_type == GGUFValueType.INT16: + value = np.int16(int(value_text)) + elif value_type == GGUFValueType.UINT32: + value = np.uint32(int(value_text)) + elif value_type == GGUFValueType.INT32: + value = np.int32(int(value_text)) + elif value_type == GGUFValueType.FLOAT32: + value = np.float32(float(value_text)) + elif value_type == GGUFValueType.BOOL: + value = value_text.lower() in ('true', 'yes', '1') + elif value_type == GGUFValueType.STRING: + value = value_text + else: + value = value_text + + return key, value_type, value + + +class GGUFEditorWindow(QMainWindow): + def __init__(self): + super().__init__() + + self.setWindowTitle("GGUF Editor") + self.resize(1000, 800) + + self.current_file = None + self.reader = None + self.modified = False + self.metadata_changes = {} # Store changes to apply when saving + self.metadata_to_remove = set() # Store keys to remove when saving + self.on_metadata_changed_is_connected = False + + self.setup_ui() + + def setup_ui(self): + central_widget = QWidget() + self.setCentralWidget(central_widget) + + main_layout = QVBoxLayout(central_widget) + + # File controls + file_layout = QHBoxLayout() + + self.file_path_edit = QLineEdit() + self.file_path_edit.setReadOnly(True) + file_layout.addWidget(self.file_path_edit) + + open_button = QPushButton("Open GGUF") + open_button.clicked.connect(self.open_file) + file_layout.addWidget(open_button) + + save_button = QPushButton("Save As...") + save_button.clicked.connect(self.save_file) + file_layout.addWidget(save_button) + + main_layout.addLayout(file_layout) + + # Tabs for different views + self.tabs = QTabWidget() + + # Metadata tab + self.metadata_tab = QWidget() + metadata_layout = QVBoxLayout(self.metadata_tab) + + # Metadata table + self.metadata_table = QTableWidget() + self.metadata_table.setColumnCount(4) + self.metadata_table.setHorizontalHeaderLabels(["Key", "Type", "Value", "Actions"]) + self.metadata_table.horizontalHeader().setSectionResizeMode(0, QHeaderView.ResizeMode.Stretch) + self.metadata_table.horizontalHeader().setSectionResizeMode(1, QHeaderView.ResizeMode.ResizeToContents) + self.metadata_table.horizontalHeader().setSectionResizeMode(2, QHeaderView.ResizeMode.Stretch) + self.metadata_table.horizontalHeader().setSectionResizeMode(3, QHeaderView.ResizeMode.ResizeToContents) + metadata_layout.addWidget(self.metadata_table) + + # Metadata controls + metadata_controls = QHBoxLayout() + + add_metadata_button = QPushButton("Add Metadata") + add_metadata_button.clicked.connect(self.add_metadata) + metadata_controls.addWidget(add_metadata_button) + + metadata_controls.addStretch() + + metadata_layout.addLayout(metadata_controls) + + # Tensors tab + self.tensors_tab = QWidget() + tensors_layout = QVBoxLayout(self.tensors_tab) + + self.tensors_table = QTableWidget() + self.tensors_table.setColumnCount(5) + self.tensors_table.setHorizontalHeaderLabels(["Name", "Type", "Shape", "Elements", "Size (bytes)"]) + self.tensors_table.horizontalHeader().setSectionResizeMode(0, QHeaderView.ResizeMode.Stretch) + self.tensors_table.horizontalHeader().setSectionResizeMode(1, QHeaderView.ResizeMode.ResizeToContents) + self.tensors_table.horizontalHeader().setSectionResizeMode(2, QHeaderView.ResizeMode.ResizeToContents) + self.tensors_table.horizontalHeader().setSectionResizeMode(3, QHeaderView.ResizeMode.ResizeToContents) + self.tensors_table.horizontalHeader().setSectionResizeMode(4, QHeaderView.ResizeMode.ResizeToContents) + tensors_layout.addWidget(self.tensors_table) + + # Add tabs to tab widget + self.tabs.addTab(self.metadata_tab, "Metadata") + self.tabs.addTab(self.tensors_tab, "Tensors") + + main_layout.addWidget(self.tabs) + + # Status bar + self.statusBar().showMessage("Ready") + + def load_file(self, file_path): + """Load a GGUF file by path""" + try: + self.statusBar().showMessage(f"Loading {file_path}...") + QApplication.processEvents() + + self.reader = GGUFReader(file_path, 'r') + self.current_file = file_path + self.file_path_edit.setText(file_path) + + self.load_metadata() + self.load_tensors() + + self.metadata_changes = {} + self.metadata_to_remove = set() + self.modified = False + + self.statusBar().showMessage(f"Loaded {file_path}") + return True + except Exception as e: + QMessageBox.critical(self, "Error", f"Failed to open file: {str(e)}") + self.statusBar().showMessage("Error loading file") + return False + + def open_file(self): + file_path, _ = QFileDialog.getOpenFileName( + self, "Open GGUF File", "", "GGUF Files (*.gguf);;All Files (*)" + ) + + if not file_path: + return + + self.load_file(file_path) + + def load_metadata(self): + self.metadata_table.setRowCount(0) + + if not self.reader: + return + + # Disconnect to prevent triggering during loading + if self.on_metadata_changed_is_connected: + with warnings.catch_warnings(): + warnings.filterwarnings('ignore') + self.metadata_table.itemChanged.disconnect(self.on_metadata_changed) + self.on_metadata_changed_is_connected = False + + for i, (key, field) in enumerate(self.reader.fields.items()): + self.metadata_table.insertRow(i) + + # Key + key_item = QTableWidgetItem(key) + key_item.setFlags(key_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.metadata_table.setItem(i, 0, key_item) + + # Type + if not field.types: + type_str = "N/A" + elif field.types[0] == GGUFValueType.ARRAY: + nest_count = len(field.types) - 1 + element_type = field.types[-1].name + # Check if this is an enum array + enum_type = self.get_enum_for_key(key) + if enum_type is not None and field.types[-1] == GGUFValueType.INT32: + element_type = enum_type.__name__ + type_str = '[' * nest_count + element_type + ']' * nest_count + else: + type_str = str(field.types[0].name) + # Check if this is an enum field + enum_type = self.get_enum_for_key(key) + if enum_type is not None and field.types[0] == GGUFValueType.INT32: + type_str = enum_type.__name__ + + type_item = QTableWidgetItem(type_str) + type_item.setFlags(type_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.metadata_table.setItem(i, 1, type_item) + + # Value + value_str = self.format_field_value(field) + value_item = QTableWidgetItem(value_str) + + # Make only simple values editable + if len(field.types) == 1 and field.types[0] != GGUFValueType.ARRAY: + value_item.setFlags(value_item.flags() | Qt.ItemFlag.ItemIsEditable) + else: + value_item.setFlags(value_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + + self.metadata_table.setItem(i, 2, value_item) + + # Actions + actions_widget = QWidget() + actions_layout = QHBoxLayout(actions_widget) + actions_layout.setContentsMargins(2, 2, 2, 2) + + # Add Edit button for arrays and enum fields + if field.types and field.types[0] == GGUFValueType.ARRAY: + edit_button = QPushButton("Edit") + edit_button.setProperty("row", i) + edit_button.setProperty("key", key) + edit_button.clicked.connect(self.edit_array_metadata) + actions_layout.addWidget(edit_button) + + # Add special label for tokenizer linked fields + if key in TOKENIZER_LINKED_KEYS: + edit_button.setText("Edit Tokenizer") + edit_button.setToolTip("Edit all tokenizer data together") + elif len(field.types) == 1 and self.get_enum_for_key(key) is not None: + edit_button = QPushButton("Edit") + edit_button.setProperty("row", i) + edit_button.setProperty("key", key) + edit_button.clicked.connect(self.edit_metadata_enum) + actions_layout.addWidget(edit_button) + + remove_button = QPushButton("Remove") + remove_button.setProperty("row", i) + remove_button.setProperty("key", key) + remove_button.clicked.connect(self.remove_metadata) + actions_layout.addWidget(remove_button) + + self.metadata_table.setCellWidget(i, 3, actions_widget) + + # Reconnect after loading + self.metadata_table.itemChanged.connect(self.on_metadata_changed) + self.on_metadata_changed_is_connected = True + + def extract_array_values(self, field: ReaderField) -> list: + """Extract all values from an array field.""" + if not field.types or field.types[0] != GGUFValueType.ARRAY: + return [] + + curr_type = field.types[1] + array_values = [] + total_elements = len(field.data) + + if curr_type == GGUFValueType.STRING: + for element_pos in range(total_elements): + value_string = str(bytes(field.parts[-1 - (total_elements - element_pos - 1) * 2]), encoding='utf-8') + array_values.append(value_string) + elif self.reader and curr_type in self.reader.gguf_scalar_to_np: + for element_pos in range(total_elements): + array_values.append(field.parts[-1 - (total_elements - element_pos - 1)][0]) + + return array_values + + def get_enum_for_key(self, key: str) -> Optional[Type[enum.Enum]]: + """Get the enum type for a given key if it exists.""" + return KEY_TO_ENUM_TYPE.get(key) + + def format_enum_value(self, value: Any, enum_type: Type[enum.Enum]) -> str: + """Format a value as an enum if possible.""" + try: + if isinstance(value, (int, str)): + enum_value = enum_type(value) + return f"{enum_value.name} ({value})" + except (ValueError, KeyError): + pass + return str(value) + + def format_field_value(self, field: ReaderField) -> str: + if not field.types: + return "N/A" + + if len(field.types) == 1: + curr_type = field.types[0] + if curr_type == GGUFValueType.STRING: + return str(bytes(field.parts[-1]), encoding='utf-8') + elif self.reader and curr_type in self.reader.gguf_scalar_to_np: + value = field.parts[-1][0] + # Check if this field has an enum type + enum_type = self.get_enum_for_key(field.name) + if enum_type is not None: + return self.format_enum_value(value, enum_type) + return str(value) + + if field.types[0] == GGUFValueType.ARRAY: + array_values = self.extract_array_values(field) + render_element = min(5, len(array_values)) + + # Get enum type for this array if applicable + enum_type = self.get_enum_for_key(field.name) + + if enum_type is not None: + array_elements = [] + for i in range(render_element): + array_elements.append(self.format_enum_value(array_values[i], enum_type)) + else: + array_elements = [str(array_values[i]) for i in range(render_element)] + + return f"[ {', '.join(array_elements).strip()}{', ...' if len(array_values) > len(array_elements) else ''} ]" + + return "Complex value" + + def load_tensors(self): + self.tensors_table.setRowCount(0) + + if not self.reader: + return + + for i, tensor in enumerate(self.reader.tensors): + self.tensors_table.insertRow(i) + + # Name + name_item = QTableWidgetItem(tensor.name) + name_item.setFlags(name_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.tensors_table.setItem(i, 0, name_item) + + # Type + type_item = QTableWidgetItem(tensor.tensor_type.name) + type_item.setFlags(type_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.tensors_table.setItem(i, 1, type_item) + + # Shape + shape_str = " × ".join(str(d) for d in tensor.shape) + shape_item = QTableWidgetItem(shape_str) + shape_item.setFlags(shape_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.tensors_table.setItem(i, 2, shape_item) + + # Elements + elements_item = QTableWidgetItem(str(tensor.n_elements)) + elements_item.setFlags(elements_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.tensors_table.setItem(i, 3, elements_item) + + # Size + size_item = QTableWidgetItem(f"{tensor.n_bytes:,}") + size_item.setFlags(size_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.tensors_table.setItem(i, 4, size_item) + + def on_metadata_changed(self, item): + if item.column() != 2: # Only handle value column changes + return + + row = item.row() + orig_item = self.metadata_table.item(row, 0) + key = None + if orig_item: + key = orig_item.text() + new_value = item.text() + + field = None + if self.reader and key: + field = self.reader.get_field(key) + if not field or not field.types or not key: + return + + value_type = field.types[0] + + # Check if this is an enum field + enum_type = self.get_enum_for_key(key) + if enum_type is not None and value_type == GGUFValueType.INT32: + # Try to parse the enum value from the text + try: + # Check if it's a name + try: + enum_val = enum_type[new_value] + converted_value = enum_val.value + except (KeyError, AttributeError): + # Check if it's a number or "NAME (value)" format + if '(' in new_value and ')' in new_value: + # Extract the value from "NAME (value)" format + value_part = new_value.split('(')[1].split(')')[0].strip() + converted_value = int(value_part) + else: + # Try to convert directly to int + converted_value = int(new_value) + + # Validate that it's a valid enum value + enum_type(converted_value) + + # Store the change + self.metadata_changes[key] = (value_type, converted_value) + self.modified = True + + # Update display with formatted enum value + formatted_value = self.format_enum_value(converted_value, enum_type) + item.setText(formatted_value) + + self.statusBar().showMessage(f"Changed {key} to {formatted_value}") + return + except (ValueError, KeyError) as e: + QMessageBox.warning( + self, + f"Invalid Enum Value ({e})", + f"'{new_value}' is not a valid {enum_type.__name__} value.\n" + f"Valid values are: {', '.join(v.name for v in enum_type)}") + + # Revert to original value + original_value = self.format_field_value(field) + item.setText(original_value) + return + + try: + # Convert the string value to the appropriate type + if value_type == GGUFValueType.UINT8: + converted_value = np.uint8(int(new_value)) + elif value_type == GGUFValueType.INT8: + converted_value = np.int8(int(new_value)) + elif value_type == GGUFValueType.UINT16: + converted_value = np.uint16(int(new_value)) + elif value_type == GGUFValueType.INT16: + converted_value = np.int16(int(new_value)) + elif value_type == GGUFValueType.UINT32: + converted_value = np.uint32(int(new_value)) + elif value_type == GGUFValueType.INT32: + converted_value = np.int32(int(new_value)) + elif value_type == GGUFValueType.FLOAT32: + converted_value = np.float32(float(new_value)) + elif value_type == GGUFValueType.BOOL: + converted_value = new_value.lower() in ('true', 'yes', '1') + elif value_type == GGUFValueType.STRING: + converted_value = new_value + else: + # Unsupported type for editing + return + + # Store the change + self.metadata_changes[key] = (value_type, converted_value) + self.modified = True + + self.statusBar().showMessage(f"Changed {key} to {new_value}") + except ValueError: + QMessageBox.warning(self, "Invalid Value", f"The value '{new_value}' is not valid for type {value_type.name}") + + # Revert to original value + original_value = self.format_field_value(field) + item.setText(original_value) + + def remove_metadata(self): + button = self.sender() + key = button.property("key") + row = button.property("row") + + reply = QMessageBox.question( + self, "Confirm Removal", + f"Are you sure you want to remove the metadata key '{key}'?", + QMessageBox.StandardButton.Yes | QMessageBox.StandardButton.No, QMessageBox.StandardButton.No + ) + + if reply == QMessageBox.StandardButton.Yes: + self.metadata_table.removeRow(row) + self.metadata_to_remove.add(key) + + # If we previously had changes for this key, remove them + if key in self.metadata_changes: + del self.metadata_changes[key] + + self.modified = True + self.statusBar().showMessage(f"Marked {key} for removal") + + def edit_metadata_enum(self): + """Edit an enum metadata field.""" + button = self.sender() + key = button.property("key") + row = button.property("row") + + field = None + if self.reader: + field = self.reader.get_field(key) + if not field or not field.types: + return + + enum_type = self.get_enum_for_key(key) + if enum_type is None: + return + + # Get current value + current_value = field.contents() + + # Create a dialog with enum options + dialog = QDialog(self) + dialog.setWindowTitle(f"Select {enum_type.__name__} Value") + layout = QVBoxLayout(dialog) + + combo = QComboBox() + for enum_val in enum_type: + combo.addItem(f"{enum_val.name} ({enum_val.value})", enum_val.value) + + # Set current value + try: + if isinstance(current_value, (int, str)): + enum_val = enum_type(current_value) + combo.setCurrentText(f"{enum_val.name} ({current_value})") + except (ValueError, KeyError): + pass + + layout.addWidget(combo) + + buttons = QDialogButtonBox(QDialogButtonBox.StandardButton.Ok | QDialogButtonBox.StandardButton.Cancel) + buttons.accepted.connect(dialog.accept) + buttons.rejected.connect(dialog.reject) + layout.addWidget(buttons) + + if dialog.exec() == QDialog.DialogCode.Accepted: + # Get the selected value + new_value = combo.currentData() + enum_val = enum_type(new_value) + + # Store the change + self.metadata_changes[key] = (field.types[0], new_value) + self.modified = True + + # Update display + display_text = f"{enum_val.name} ({new_value})" + target_item = self.metadata_table.item(row, 2) + if target_item: + target_item.setText(display_text) + + self.statusBar().showMessage(f"Changed {key} to {display_text}") + + def edit_array_metadata(self): + button = self.sender() + key = button.property("key") + row = button.property("row") + + # Check if this is one of the linked tokenizer keys + if key in TOKENIZER_LINKED_KEYS: + self.edit_tokenizer_metadata(key) + return + + field = None + if self.reader: + field = self.reader.get_field(key) + if not field or not field.types or field.types[0] != GGUFValueType.ARRAY: + return + + # Get array element type + element_type = field.types[1] + + # Extract array values + array_values = self.extract_array_values(field) + + # Open array editor dialog + dialog = ArrayEditorDialog(array_values, element_type, key, self) + if dialog.exec() == QDialog.DialogCode.Accepted: + new_values = dialog.get_array_values() + + # Store the change + self.metadata_changes[key] = (GGUFValueType.ARRAY, (element_type, new_values)) + self.modified = True + + # Update display + enum_type = self.get_enum_for_key(key) + if enum_type is not None and element_type == GGUFValueType.INT32: + value_str = f"[ {', '.join(self.format_enum_value(v, enum_type) for v in new_values[:5])}{', ...' if len(new_values) > 5 else ''} ]" + else: + value_str = f"[ {', '.join(str(v) for v in new_values[:5])}{', ...' if len(new_values) > 5 else ''} ]" + target_item = self.metadata_table.item(row, 2) + if target_item: + target_item.setText(value_str) + + self.statusBar().showMessage(f"Updated array values for {key}") + + def edit_tokenizer_metadata(self, trigger_key): + """Edit the linked tokenizer metadata arrays together.""" + if not self.reader: + return + + # Get all three fields + tokens_field = self.reader.get_field(gguf.Keys.Tokenizer.LIST) + token_types_field = self.reader.get_field(gguf.Keys.Tokenizer.TOKEN_TYPE) + scores_field = self.reader.get_field(gguf.Keys.Tokenizer.SCORES) + + # Extract values from each field + tokens = self.extract_array_values(tokens_field) if tokens_field else [] + token_types = self.extract_array_values(token_types_field) if token_types_field else [] + scores = self.extract_array_values(scores_field) if scores_field else [] + + # Apply any pending changes + if gguf.Keys.Tokenizer.LIST in self.metadata_changes: + _, (_, tokens) = self.metadata_changes[gguf.Keys.Tokenizer.LIST] + if gguf.Keys.Tokenizer.TOKEN_TYPE in self.metadata_changes: + _, (_, token_types) = self.metadata_changes[gguf.Keys.Tokenizer.TOKEN_TYPE] + if gguf.Keys.Tokenizer.SCORES in self.metadata_changes: + _, (_, scores) = self.metadata_changes[gguf.Keys.Tokenizer.SCORES] + + # Open the tokenizer editor dialog + dialog = TokenizerEditorDialog(tokens, token_types, scores, self) + if dialog.exec() == QDialog.DialogCode.Accepted: + new_tokens, new_token_types, new_scores = dialog.get_data() + + # Store changes for all three arrays + if tokens_field: + self.metadata_changes[gguf.Keys.Tokenizer.LIST] = ( + GGUFValueType.ARRAY, + (tokens_field.types[1], new_tokens) + ) + + if token_types_field: + self.metadata_changes[gguf.Keys.Tokenizer.TOKEN_TYPE] = ( + GGUFValueType.ARRAY, + (token_types_field.types[1], new_token_types) + ) + + if scores_field: + self.metadata_changes[gguf.Keys.Tokenizer.SCORES] = ( + GGUFValueType.ARRAY, + (scores_field.types[1], new_scores) + ) + + self.modified = True + + # Update display for all three fields + self.update_tokenizer_display(gguf.Keys.Tokenizer.LIST, new_tokens) + self.update_tokenizer_display(gguf.Keys.Tokenizer.TOKEN_TYPE, new_token_types) + self.update_tokenizer_display(gguf.Keys.Tokenizer.SCORES, new_scores) + + self.statusBar().showMessage("Updated tokenizer data") + + def update_tokenizer_display(self, key, values): + """Update the display of a tokenizer field in the metadata table.""" + for row in range(self.metadata_table.rowCount()): + key_item = self.metadata_table.item(row, 0) + if key_item and key_item.text() == key: + value_str = f"[ {', '.join(str(v) for v in values[:5])}{', ...' if len(values) > 5 else ''} ]" + value_item = self.metadata_table.item(row, 2) + if value_item: + value_item.setText(value_str) + break + + def add_metadata(self): + dialog = AddMetadataDialog(self) + if dialog.exec() == QDialog.DialogCode.Accepted: + key, value_type, value = dialog.get_data() + + if not key: + QMessageBox.warning(self, "Invalid Key", "Key cannot be empty") + return + + # Check if key already exists + for row in range(self.metadata_table.rowCount()): + orig_item = self.metadata_table.item(row, 0) + if orig_item and orig_item.text() == key: + QMessageBox.warning(self, "Duplicate Key", f"Key '{key}' already exists") + return + + # Add to table + row = self.metadata_table.rowCount() + self.metadata_table.insertRow(row) + + # Key + key_item = QTableWidgetItem(key) + key_item.setFlags(key_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.metadata_table.setItem(row, 0, key_item) + + # Type + type_item = QTableWidgetItem(value_type.name) + type_item.setFlags(type_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.metadata_table.setItem(row, 1, type_item) + + # Value + value_item = QTableWidgetItem(str(value)) + value_item.setFlags(value_item.flags() | Qt.ItemFlag.ItemIsEditable) + self.metadata_table.setItem(row, 2, value_item) + + # Actions + actions_widget = QWidget() + actions_layout = QHBoxLayout(actions_widget) + actions_layout.setContentsMargins(2, 2, 2, 2) + + remove_button = QPushButton("Remove") + remove_button.setProperty("row", row) + remove_button.setProperty("key", key) + remove_button.clicked.connect(self.remove_metadata) + actions_layout.addWidget(remove_button) + + self.metadata_table.setCellWidget(row, 3, actions_widget) + + # Store the change + self.metadata_changes[key] = (value_type, value) + self.modified = True + + self.statusBar().showMessage(f"Added new metadata key {key}") + + def save_file(self): + if not self.reader: + QMessageBox.warning(self, "No File Open", "Please open a GGUF file first") + return + + if not self.modified and not self.metadata_changes and not self.metadata_to_remove: + QMessageBox.information(self, "No Changes", "No changes to save") + return + + file_path, _ = QFileDialog.getSaveFileName( + self, "Save GGUF File As", "", "GGUF Files (*.gguf);;All Files (*)" + ) + + if not file_path: + return + + try: + self.statusBar().showMessage(f"Saving to {file_path}...") + QApplication.processEvents() + + # Get architecture and endianness from the original file + arch = 'unknown' + field = self.reader.get_field(gguf.Keys.General.ARCHITECTURE) + if field: + arch = field.contents() + + # Create writer + writer = GGUFWriter(file_path, arch=arch, endianess=self.reader.endianess) + + # Get alignment if present + alignment = None + field = self.reader.get_field(gguf.Keys.General.ALIGNMENT) + if field: + alignment = field.contents() + if alignment is not None: + writer.data_alignment = alignment + + # Copy metadata with changes + for field in self.reader.fields.values(): + # Skip virtual fields and fields written by GGUFWriter + if field.name == gguf.Keys.General.ARCHITECTURE or field.name.startswith('GGUF.'): + continue + + # Skip fields marked for removal + if field.name in self.metadata_to_remove: + continue + + # Apply changes if any + sub_type = None + if field.name in self.metadata_changes: + value_type, value = self.metadata_changes[field.name] + if value_type == GGUFValueType.ARRAY: + # Handle array values + sub_type, value = value + else: + # Copy original value + value = field.contents() + value_type = field.types[0] + if value_type == GGUFValueType.ARRAY: + sub_type = field.types[-1] + + if value is not None: + writer.add_key_value(field.name, value, value_type, sub_type=sub_type) + + # Add new metadata + for key, (value_type, value) in self.metadata_changes.items(): + # Skip if the key already existed (we handled it above) + if self.reader.get_field(key) is not None: + continue + + sub_type = None + if value_type == GGUFValueType.ARRAY: + # Handle array values + sub_type, value = value + + writer.add_key_value(key, value, value_type, sub_type=sub_type) + + # Add tensors (including data) + for tensor in self.reader.tensors: + writer.add_tensor(tensor.name, tensor.data, raw_shape=tensor.data.shape, raw_dtype=tensor.tensor_type, tensor_endianess=self.reader.endianess) + + # Write header and metadata + writer.open_output_file(Path(file_path)) + writer.write_header_to_file() + writer.write_kv_data_to_file() + + # Write tensor data using the optimized method + writer.write_tensors_to_file(progress=False) + + writer.close() + + self.statusBar().showMessage(f"Saved to {file_path}") + + # Ask if user wants to open the new file + reply = QMessageBox.question( + self, "Open Saved File", + "Would you like to open the newly saved file?", + QMessageBox.StandardButton.Yes | QMessageBox.StandardButton.No, QMessageBox.StandardButton.Yes + ) + + if reply == QMessageBox.StandardButton.Yes: + self.reader = GGUFReader(file_path, 'r') + self.current_file = file_path + self.file_path_edit.setText(file_path) + + self.load_metadata() + self.load_tensors() + + self.metadata_changes = {} + self.metadata_to_remove = set() + self.modified = False + + except Exception as e: + QMessageBox.critical(self, "Error", f"Failed to save file: {str(e)}") + self.statusBar().showMessage("Error saving file") + + +def main() -> None: + parser = argparse.ArgumentParser(description="GUI GGUF Editor") + parser.add_argument("model_path", nargs="?", help="path to GGUF model file to load at startup") + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") + + args = parser.parse_args() + + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + + app = QApplication(sys.argv) + window = GGUFEditorWindow() + window.show() + + # Load model if specified + if args.model_path: + if os.path.isfile(args.model_path) and args.model_path.endswith('.gguf'): + window.load_file(args.model_path) + else: + logger.error(f"Invalid model path: {args.model_path}") + QMessageBox.warning( + window, + "Invalid Model Path", + f"The specified file does not exist or is not a GGUF file: {args.model_path}") + + sys.exit(app.exec()) + + +if __name__ == '__main__': + main() diff --git a/llama.cpp/gguf-py/gguf/scripts/gguf_hash.py b/llama.cpp/gguf-py/gguf/scripts/gguf_hash.py new file mode 100755 index 0000000..3ef9899 --- /dev/null +++ b/llama.cpp/gguf-py/gguf/scripts/gguf_hash.py @@ -0,0 +1,102 @@ +#!/usr/bin/env python3 +from __future__ import annotations + +import uuid +import hashlib + +import logging +import argparse +import os +import sys +from pathlib import Path + +from tqdm import tqdm + +# Necessary to load the local gguf package +if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent.parent / 'gguf-py').exists(): + sys.path.insert(0, str(Path(__file__).parent.parent.parent)) + +from gguf import GGUFReader # noqa: E402 + + +logger = logging.getLogger("gguf-hash") + +# UUID_NAMESPACE_LLAMA_CPP = uuid.uuid5(uuid.NAMESPACE_URL, 'en.wikipedia.org/wiki/Llama.cpp') +UUID_NAMESPACE_LLAMA_CPP = uuid.UUID('ef001206-dadc-5f6d-a15f-3359e577d4e5') + + +# For more information about what field.parts and field.data represent, +# please see the comments in the modify_gguf.py example. +def gguf_hash(reader: GGUFReader, filename: str, disable_progress_bar: bool, no_layer: bool) -> None: + sha1 = hashlib.sha1() + sha256 = hashlib.sha256() + uuidv5_sha1 = hashlib.sha1() + uuidv5_sha1.update(UUID_NAMESPACE_LLAMA_CPP.bytes) + + # Total Weight Calculation For Progress Bar + total_weights = 0 + for n, tensor in enumerate(reader.tensors, 1): + + # We don't need these + if tensor.name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")): + continue + + # Calculate Tensor Volume + sum_weights_in_tensor = 1 + for dim in tensor.shape: + sum_weights_in_tensor *= dim + total_weights += sum_weights_in_tensor + + # Hash Progress Bar + bar = tqdm(desc="Hashing", total=total_weights, unit="weights", unit_scale=True, disable=disable_progress_bar) + + # Hashing Process + for tensor in reader.tensors: + + # We don't need these + if tensor.name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")): + continue + + # Progressbar + sum_weights_in_tensor = 1 + for dim in tensor.shape: + sum_weights_in_tensor *= dim + bar.update(sum_weights_in_tensor) + + if not no_layer: + + sha1_layer = hashlib.sha1() + sha1_layer.update(tensor.data.data) + print("sha1 {0} {1}:{2}".format(sha1_layer.hexdigest(), filename, tensor.name)) # noqa: NP100 + + sha256_layer = hashlib.sha256() + sha256_layer.update(tensor.data.data) + print("sha256 {0} {1}:{2}".format(sha256_layer.hexdigest(), filename, tensor.name)) # noqa: NP100 + + sha1.update(tensor.data.data) + sha256.update(tensor.data.data) + uuidv5_sha1.update(tensor.data.data) + + # Flush Hash Progress Bar + bar.close() + + # Display Hash Output + print("sha1 {0} {1}".format(sha1.hexdigest(), filename)) # noqa: NP100 + print("sha256 {0} {1}".format(sha256.hexdigest(), filename)) # noqa: NP100 + print("uuid {0} {1}".format(uuid.UUID(bytes=uuidv5_sha1.digest()[:16], version=5), filename)) # noqa: NP100 + + +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-layer", action="store_true", help="exclude per layer hash") + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") + parser.add_argument("--progressbar", action="store_true", help="enable progressbar") + args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"]) + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + reader = GGUFReader(args.model, 'r') + gguf_hash(reader, args.model, not args.progressbar, args.no_layer) + + +if __name__ == '__main__': + main() diff --git a/llama.cpp/gguf-py/gguf/scripts/gguf_new_metadata.py b/llama.cpp/gguf-py/gguf/scripts/gguf_new_metadata.py new file mode 100755 index 0000000..c67436b --- /dev/null +++ b/llama.cpp/gguf-py/gguf/scripts/gguf_new_metadata.py @@ -0,0 +1,216 @@ +#!/usr/bin/env python3 +from __future__ import annotations + +import logging +import argparse +import os +import sys +import json +from pathlib import Path + +from tqdm import tqdm +from typing import Any, Sequence, NamedTuple + +# Necessary to load the local gguf package +if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent.parent / 'gguf-py').exists(): + sys.path.insert(0, str(Path(__file__).parent.parent.parent)) + +import gguf + +logger = logging.getLogger("gguf-new-metadata") + + +class MetadataDetails(NamedTuple): + type: gguf.GGUFValueType + value: Any + description: str = '' + sub_type: gguf.GGUFValueType | None = None + + +def get_field_data(reader: gguf.GGUFReader, key: str) -> Any: + field = reader.get_field(key) + + return field.contents() if field else None + + +def find_token(token_list: Sequence[int], token: str) -> Sequence[int]: + token_ids = [index for index, value in enumerate(token_list) if value == token] + + if len(token_ids) == 0: + raise LookupError(f'Unable to find "{token}" in token list!') + + return token_ids + + +def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new_metadata: dict[str, MetadataDetails], remove_metadata: Sequence[str]) -> None: + for field in reader.fields.values(): + # Suppress virtual fields and fields written by GGUFWriter + if field.name == gguf.Keys.General.ARCHITECTURE or field.name.startswith('GGUF.'): + logger.debug(f'Suppressing {field.name}') + continue + + # 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: + logger.debug(f'Skipping {field.name}') + continue + + if field.name in remove_metadata: + logger.debug(f'Removing {field.name}') + continue + + val_type = field.types[0] + sub_type = field.types[-1] if val_type == gguf.GGUFValueType.ARRAY else None + old_val = MetadataDetails(val_type, field.contents(), sub_type=sub_type) + val = new_metadata.get(field.name, old_val) + + if field.name in new_metadata: + logger.debug(f'Modifying {field.name}: "{old_val.value}" -> "{val.value}" {val.description}') + del new_metadata[field.name] + elif val.value is not None: + logger.debug(f'Copying {field.name}') + + if val.value is not None: + writer.add_key_value(field.name, val.value, val.type, sub_type=sub_type if val.sub_type is None else val.sub_type) + + if gguf.Keys.Tokenizer.CHAT_TEMPLATE in new_metadata: + logger.debug('Adding chat template(s)') + writer.add_chat_template(new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE].value) + del new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] + + for key, val in new_metadata.items(): + logger.debug(f'Adding {key}: "{val.value}" {val.description}') + writer.add_key_value(key, val.value, val.type) + + total_bytes = 0 + + for tensor in reader.tensors: + total_bytes += tensor.n_bytes + writer.add_tensor_info(tensor.name, tensor.data.shape, tensor.data.dtype, tensor.data.nbytes, tensor.tensor_type) + + bar = tqdm(desc="Writing", total=total_bytes, unit="byte", unit_scale=True) + + writer.write_header_to_file() + writer.write_kv_data_to_file() + writer.write_ti_data_to_file() + + for tensor in reader.tensors: + writer.write_tensor_data(tensor.data, tensor_endianess=reader.endianess) + bar.update(tensor.n_bytes) + + writer.close() + + +def main() -> None: + tokenizer_metadata = (getattr(gguf.Keys.Tokenizer, n) for n in gguf.Keys.Tokenizer.__dict__.keys() if not n.startswith('_')) + token_names = dict((n.split('.')[-1][:-len('_token_id')], n) for n in tokenizer_metadata if n.endswith('_token_id')) + + parser = argparse.ArgumentParser(description="Make a copy of a GGUF file with new metadata") + parser.add_argument("input", type=Path, help="GGUF format model input filename") + parser.add_argument("output", type=Path, help="GGUF format model output filename") + 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("--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("--chat-template-file", type=Path, help="Jinja file containing chat template", metavar='chat_template.jinja') + 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("--special-token", action="append", type=str, help="Special token by value", nargs=2, metavar=(' | '.join(token_names.keys()), '"<token>"')) + parser.add_argument("--special-token-by-id", action="append", type=str, help="Special token by id", nargs=2, metavar=(' | '.join(token_names.keys()), '0')) + parser.add_argument("--force", action="store_true", help="Bypass warnings without confirmation") + parser.add_argument("--verbose", action="store_true", help="Increase output verbosity") + args = parser.parse_args(None if len(sys.argv) > 2 else ["--help"]) + + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + + new_metadata = {} + remove_metadata = args.remove_metadata or [] + + if args.general_name: + new_metadata[gguf.Keys.General.NAME] = MetadataDetails(gguf.GGUFValueType.STRING, args.general_name) + + if args.general_description: + new_metadata[gguf.Keys.General.DESCRIPTION] = MetadataDetails(gguf.GGUFValueType.STRING, args.general_description) + + 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) + + if args.chat_template_config: + with open(args.chat_template_config, 'r', encoding='utf-8') as fp: + config = json.load(fp) + template = config.get('chat_template') + if template: + new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, template) + + if args.chat_template_file: + with open(args.chat_template_file, 'r', encoding='utf-8') as fp: + template = fp.read() + new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, template) + + if args.pre_tokenizer: + new_metadata[gguf.Keys.Tokenizer.PRE] = MetadataDetails(gguf.GGUFValueType.STRING, args.pre_tokenizer) + + if remove_metadata: + logger.warning('*** Warning *** Warning *** Warning **') + logger.warning('* Most metadata is required for a fully functional GGUF file,') + logger.warning('* removing crucial metadata may result in a corrupt output file!') + + if not args.force: + logger.warning('* Enter exactly YES if you are positive you want to proceed:') + response = input('YES, I am sure> ') + if response != 'YES': + logger.info("You didn't enter YES. Okay then, see ya!") + sys.exit(0) + + logger.info(f'* Loading: {args.input}') + reader = gguf.GGUFReader(args.input, 'r') + + arch = get_field_data(reader, gguf.Keys.General.ARCHITECTURE) + + token_list = get_field_data(reader, gguf.Keys.Tokenizer.LIST) or [] + + for name, token in args.special_token or []: + if name not in token_names: + logger.warning(f'Unknown special token "{name}", ignoring...') + else: + ids = find_token(token_list, token) + new_metadata[token_names[name]] = MetadataDetails(gguf.GGUFValueType.UINT32, ids[0], f'= {token}') + + if len(ids) > 1: + logger.warning(f'Multiple "{token}" tokens found, choosing ID {ids[0]}, use --special-token-by-id if you want another:') + logger.warning(', '.join(str(i) for i in ids)) + + for name, id_string in args.special_token_by_id or []: + if name not in token_names: + logger.warning(f'Unknown special token "{name}", ignoring...') + elif not id_string.isdecimal(): + raise LookupError(f'Token ID "{id_string}" is not a valid ID!') + else: + id_int = int(id_string) + + if id_int >= 0 and id_int < len(token_list): + new_metadata[token_names[name]] = MetadataDetails(gguf.GGUFValueType.UINT32, id_int, f'= {token_list[id_int]}') + else: + raise LookupError(f'Token ID {id_int} is not within token list!') + + if os.path.isfile(args.output) and not args.force: + logger.warning('*** Warning *** Warning *** Warning **') + logger.warning(f'* The "{args.output}" GGUF file already exists, it will be overwritten!') + logger.warning('* Enter exactly YES if you are positive you want to proceed:') + response = input('YES, I am sure> ') + if response != 'YES': + logger.info("You didn't enter YES. Okay then, see ya!") + sys.exit(0) + + logger.info(f'* Writing: {args.output}') + writer = gguf.GGUFWriter(args.output, arch=arch, endianess=reader.endianess) + + alignment = get_field_data(reader, gguf.Keys.General.ALIGNMENT) + if alignment is not None: + logger.debug(f'Setting custom alignment: {alignment}') + writer.data_alignment = alignment + + copy_with_new_metadata(reader, writer, new_metadata, remove_metadata) + + +if __name__ == '__main__': + main() diff --git a/llama.cpp/gguf-py/gguf/scripts/gguf_set_metadata.py b/llama.cpp/gguf-py/gguf/scripts/gguf_set_metadata.py new file mode 100755 index 0000000..f5809c3 --- /dev/null +++ b/llama.cpp/gguf-py/gguf/scripts/gguf_set_metadata.py @@ -0,0 +1,95 @@ +#!/usr/bin/env python3 +import logging +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.parent / 'gguf-py').exists(): + sys.path.insert(0, str(Path(__file__).parent.parent.parent)) + +from gguf import GGUFReader # noqa: E402 + +logger = logging.getLogger("gguf-set-metadata") + + +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: + logger.error(f'! Field {repr(args.key)} not found') + 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: + logger.error(f'! This tool only supports changing simple values, {repr(args.key)} has unsupported type {field.types}') + sys.exit(1) + current_value = field.parts[field.data[0]][0] + new_value = handler(args.value) + logger.info(f'* Preparing to change field {repr(args.key)} from {current_value} to {new_value}') + if current_value == new_value: + logger.info(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: + logger.warning('*** Warning *** Warning *** Warning **') + logger.warning('* Changing fields in a GGUF file can make it unusable. Proceed at your own risk.') + logger.warning('* Enter exactly YES if you are positive you want to proceed:') + response = input('YES, I am sure> ') + if response != 'YES': + logger.info("You didn't enter YES. Okay then, see ya!") + sys.exit(0) + field.parts[field.data[0]][0] = new_value + logger.info('* 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") + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") + + args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"]) + + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + + logger.info(f'* Loading: {args.model}') + reader = GGUFReader(args.model, 'r' if args.dry_run else 'r+') + set_metadata(reader, args) + + +if __name__ == '__main__': + main() |