gguf

ggus 库

gguf 实用工具

帮助信息

cargo xtask --help

gguf-utils is a command-line tool for working with gguf files

Usage: xtask.exe <COMMAND>

Commands:
  show      Show the contents of gguf files
  split     Split gguf files into shards
  merge     Merge shards into a single gguf file
  filter    Filter gguf files based on wildcard patterns
  convert   Convert gguf files to different format
  set-meta  Set metadata of gguf files
  help      Print this message or the help of the given subcommand(s)

Options:
  -h, --help     Print help
  -V, --version  Print version

阅读内容

cargo show --help

Show the contents of gguf files

Usage: xtask.exe show [OPTIONS] <FILE>

Arguments:
  <FILE>  The file to show

Options:
      --shards                         If set, show all shards in the directory
  -n, --array-detail <ARRAY_DETAIL>    How many elements to show in arrays, `all` for all elements [default: 8]
  -m, --filter-meta <FILTER_META>      Meta to show [default: *]
  -t, --filter-tensor <FILTER_TENSOR>  Tensors to show [default: *]
      --log <LOG>                      Log level, may be "off", "trace", "debug", "info" or "error"
  -h, --help                           Print help

内容过滤

cargo filter --help

Filter gguf files based on wildcard patterns

Usage: xtask.exe filter [OPTIONS] <FILE>

Arguments:
  <FILE>  The file to filter

Options:
  -o, --output-dir <OUTPUT_DIR>        Output directory for filtered file
  -m, --filter-meta <FILTER_META>      Meta to keep [default: *]
  -t, --filter-tensor <FILTER_TENSOR>  Tensors to keep [default: *]
      --log <LOG>                      Log level, may be "off", "trace", "debug", "info" or "error"
  -h, --help                           Print help

分片

cargo split --help

Split gguf files into shards

Usage: xtask.exe split [OPTIONS] <FILE>

Arguments:
  <FILE>  File to split

Options:
  -o, --output-dir <OUTPUT_DIR>    Output directory for splited shards
  -t, --max-tensors <MAX_TENSORS>  Max count of tensors per shard
  -s, --max-bytes <MAX_BYTES>      Max size in bytes per shard
  -n, --no-tensor-first            If set, the first shard will not contain any tensor
      --log <LOG>                  Log level, may be "off", "trace", "debug", "info" or "error"
  -h, --help                       Print help

合并

cargo merge --help

Merge shards into a single gguf file

Usage: xtask.exe merge [OPTIONS] <FILE>

Arguments:
  <FILE>  One of the shards to merge

Options:
  -o, --output-dir <OUTPUT_DIR>  Output directory for merged file
      --log <LOG>                Log level, may be "off", "trace", "debug", "info" or "error"
  -h, --help                     Print help

转换格式

cargo convert --help

Convert gguf files to different format

Usage: xtask.exe convert [OPTIONS] --steps <STEPS> <FILE>

Arguments:
  <FILE>  File to convert

Options:
  -o, --output-dir <OUTPUT_DIR>    Output directory for converted files
  -x, --steps <STEPS>              Steps to apply, separated by "->", maybe "sort", "merge-linear", "split-linear", "filter-meta:<key>", "filter-tensor:<name>", "cast:<dtype>" or "distribute:<n>"
  -t, --max-tensors <MAX_TENSORS>  Max count of tensors per shard
  -s, --max-bytes <MAX_BYTES>      Max size in bytes per shard
  -n, --no-tensor-first            If set, the first shard will not contain any tensor
      --log <LOG>                  Log level, may be "off", "trace", "debug", "info" or "error"
  -h, --help                       Print help

修改元信息

cargo set-meta --help

Set metadata of gguf files

Usage: xtask.exe set-meta [OPTIONS] <FILE> <META_KVS>

Arguments:
  <FILE>      File to set metadata
  <META_KVS>  Meta data to set for the file

Options:
  -o, --output-dir <OUTPUT_DIR>  Output directory for changed file
      --log <LOG>                Log level, may be "off", "trace", "debug", "info" or "error"
  -h, --help                     Print help

<META_KVS> 是具有类似如下格式的文本文件：

1. 配置代数类型元信息

   代数类型包括整型、无符号整型、浮点型和布尔。

   `<KEY>`<Ty> <VAL>
   

2. 配置字符串元信息

   单行字符串：

   `<KEY>`str "<VAL>"
   

   多行字符串：

   `<KEY>`str<Sep>
   <Sep> [Content]
   <Sep> [Content]
   <Sep> [Content]
   
   

   其中 Sep 是表示字符串继续的分隔符。必须紧邻 str，之间不能包含空白字符，且分隔符中也不能包含空白字符。 连续的多行字符串，每行必须以分隔符+空格起始，此行后续所有字符（包括换行符）都被视作多行字符串的内容，不转义。 任何不以分隔符开始的行（包括空行）都将结束多行字符串。

3. 配置数组元信息

   TODO: 当前此功能未实现。

这是一个配置元信息的示例文件内容：

`llama.block_count`             u64 22
`llama.context_length`          u64 2048
`llama.embedding_length`        u64 2048
`llama.feed_forward_length`     u64 5632
`llama.attention.head_count`    u64 32
`llama.attention.head_count_kv` u64 4
`llama.rope.dimension_count`    u64 64

`tokenizer.chat_template` str|
| {%- for message in messages -%}
| {%- if message['role'] == 'user' -%}
| {{ '<|user|>
| ' + message['content'] + eos_token }}
| {%- elif message['role'] == 'system' -%}
| {{ '<|system|>
| ' + message['content'] + eos_token }}
| {%- elif message['role'] == 'assistant' -%}
| {{ '<|assistant|>
| ' + message['content'] + eos_token }}
| {%- endif -%}
| {%- if loop.last and add_generation_prompt -%}
| {{ '<|assistant|>
| ' }}
| {%- endif -%}
| {%- endfor -%}