GGUF is the file format of the ggml tensor library for storing machine learning models, used for example by llama.cpp for inference. It replaced the older GGML, GGMF and GGJT formats.
A GGUF file consists of a header with metadata as key-value pairs (model architecture, hyperparameters, tokenizer, etc.), a description of each tensor (name, dimensions, type and offset), and finally the tensor data.
This page hosts a formal specification of GGUF (GGML Universal File) using Kaitai Struct. This specification can be automatically translated into a variety of programming languages to get a parsing library.
All Ruby code generated by Kaitai Struct depends on the Kaitai Struct runtime library for Ruby. You must add this dependency to your project before you can parse or serialize any data.
The Ruby runtime library can be installed from RubyGems:
gem install kaitai-structParse a local file and get structure in memory:
data = Gguf.from_file("path/to/local/file.gguf")
Or parse structure from a string of bytes:
bytes = "\x00\x01\x02..."
data = Gguf.new(Kaitai::Struct::Stream.new(bytes))
After that, one can get various attributes from the structure by invoking getter methods like:
data.magic # => GGUF file magic
data.len_file # => Length of the file (computed from the header), including the padding after
the last tensor's data.
This Kaitai Struct implementation stops parsing where the tensor data
starts (which is read lazily, i.e. only on demand). Without reading the
tensor data, you can use this expected file size to check whether the file
has been truncated, to carve the file out of a larger one, etc.
# This is a generated file! Please edit source .ksy file and use kaitai-struct-compiler to rebuild
require 'kaitai/struct/struct'
unless Gem::Version.new(Kaitai::Struct::VERSION) >= Gem::Version.new('0.11')
raise "Incompatible Kaitai Struct Ruby API: 0.11 or later is required, but you have #{Kaitai::Struct::VERSION}"
end
##
# GGUF is the file format of the [ggml](https://github.com/ggml-org/ggml)
# tensor library for storing machine learning models, used for example by
# [llama.cpp](https://github.com/ggml-org/llama.cpp) for inference. It
# replaced the older GGML, GGMF and GGJT formats.
#
# A GGUF file consists of a header with metadata as key-value pairs (model
# architecture, hyperparameters, tokenizer, etc.), a description of each
# tensor (name, dimensions, type and offset), and finally the tensor data.
# @see https://github.com/ggml-org/ggml/blob/a05c9922613c6a2230dbb1d0de679519ecafb4f0/docs/gguf.md Git tag "v0.27.0"
class Gguf < Kaitai::Struct::Struct
GGML_TYPE = {
0 => :ggml_type_f32,
1 => :ggml_type_f16,
2 => :ggml_type_q4_0,
3 => :ggml_type_q4_1,
4 => :ggml_type_q4_2,
5 => :ggml_type_q4_3,
6 => :ggml_type_q5_0,
7 => :ggml_type_q5_1,
8 => :ggml_type_q8_0,
9 => :ggml_type_q8_1,
10 => :ggml_type_q2_k,
11 => :ggml_type_q3_k,
12 => :ggml_type_q4_k,
13 => :ggml_type_q5_k,
14 => :ggml_type_q6_k,
15 => :ggml_type_q8_k,
16 => :ggml_type_iq2_xxs,
17 => :ggml_type_iq2_xs,
18 => :ggml_type_iq3_xxs,
19 => :ggml_type_iq1_s,
20 => :ggml_type_iq4_nl,
21 => :ggml_type_iq3_s,
22 => :ggml_type_iq2_s,
23 => :ggml_type_iq4_xs,
24 => :ggml_type_i8,
25 => :ggml_type_i16,
26 => :ggml_type_i32,
27 => :ggml_type_i64,
28 => :ggml_type_f64,
29 => :ggml_type_iq1_m,
30 => :ggml_type_bf16,
31 => :ggml_type_q4_0_4_4,
32 => :ggml_type_q4_0_4_8,
33 => :ggml_type_q4_0_8_8,
34 => :ggml_type_tq1_0,
35 => :ggml_type_tq2_0,
36 => :ggml_type_iq4_nl_4_4,
37 => :ggml_type_iq4_nl_4_8,
38 => :ggml_type_iq4_nl_8_8,
39 => :ggml_type_mxfp4,
40 => :ggml_type_nvfp4,
41 => :ggml_type_q1_0,
42 => :ggml_type_q2_0,
}
I__GGML_TYPE = GGML_TYPE.invert
GGUF_TYPE = {
0 => :gguf_type_uint8,
1 => :gguf_type_int8,
2 => :gguf_type_uint16,
3 => :gguf_type_int16,
4 => :gguf_type_uint32,
5 => :gguf_type_int32,
6 => :gguf_type_float32,
7 => :gguf_type_bool,
8 => :gguf_type_string,
9 => :gguf_type_array,
10 => :gguf_type_uint64,
11 => :gguf_type_int64,
12 => :gguf_type_float64,
}
I__GGUF_TYPE = GGUF_TYPE.invert
def initialize(_io, _parent = nil, _root = nil)
super(_io, _parent, _root || self)
_read
end
def _read
@magic = @_io.read_bytes(4)
raise Kaitai::Struct::ValidationNotEqualError.new([71, 71, 85, 70].pack('C*'), @magic, @_io, "/seq/0") if not @magic == [71, 71, 85, 70].pack('C*')
@body = GgufBody.new(@_io, self, @_root)
self
end
class GgufBody < Kaitai::Struct::Struct
def initialize(_io, _parent = nil, _root = nil)
super(_io, _parent, _root)
_read
end
def _read
case _root.is_big_endian
when true
@_is_le = false
when false
@_is_le = true
end
if @_is_le == true
_read_le
elsif @_is_le == false
_read_be
else
raise Kaitai::Struct::UndecidedEndiannessError.new("/types/gguf_body")
end
self
end
def _read_le
@version = @_io.read_u4le
raise Kaitai::Struct::ValidationLessThanError.new(2, @version, @_io, "/types/gguf_body/seq/0") if not @version >= 2
raise Kaitai::Struct::ValidationGreaterThanError.new(3, @version, @_io, "/types/gguf_body/seq/0") if not @version <= 3
@num_tensor_infos = @_io.read_u8le
@num_kv_pairs = @_io.read_u8le
@kv_pairs = []
(num_kv_pairs).times { |i|
@kv_pairs << GgufKv.new(@_io, self, @_root, @_is_le, (i != 0 ? kv_pairs[i - 1].alignment : 32))
}
@tensor_infos = []
(num_tensor_infos).times { |i|
@tensor_infos << GgufTensorInfo.new(@_io, self, @_root, @_is_le, (i != 0 ? tensor_infos[i - 1].expected_next_offset : 0))
}
if num_tensor_infos != 0
@padding = @_io.read_bytes((alignment - _io.pos % alignment) % alignment)
end
if ofs_tensor_data < 0
@_unnamed6 = @_io.read_bytes(0)
end
self
end
def _read_be
@version = @_io.read_u4be
raise Kaitai::Struct::ValidationLessThanError.new(2, @version, @_io, "/types/gguf_body/seq/0") if not @version >= 2
raise Kaitai::Struct::ValidationGreaterThanError.new(3, @version, @_io, "/types/gguf_body/seq/0") if not @version <= 3
@num_tensor_infos = @_io.read_u8be
@num_kv_pairs = @_io.read_u8be
@kv_pairs = []
(num_kv_pairs).times { |i|
@kv_pairs << GgufKv.new(@_io, self, @_root, @_is_le, (i != 0 ? kv_pairs[i - 1].alignment : 32))
}
@tensor_infos = []
(num_tensor_infos).times { |i|
@tensor_infos << GgufTensorInfo.new(@_io, self, @_root, @_is_le, (i != 0 ? tensor_infos[i - 1].expected_next_offset : 0))
}
if num_tensor_infos != 0
@padding = @_io.read_bytes((alignment - _io.pos % alignment) % alignment)
end
if ofs_tensor_data < 0
@_unnamed6 = @_io.read_bytes(0)
end
self
end
##
# Nested arrays are supported. Although ggml (v0.27.0) doesn't support them,
# they are explicitly allowed by `gguf.md` - see
# <https://github.com/ggml-org/ggml/blob/a05c9922613c6a2230dbb1d0de679519ecafb4f0/docs/gguf.md?plain=1#L221>.
class GgufArray < Kaitai::Struct::Struct
def initialize(_io, _parent = nil, _root = nil, _is_le = nil)
super(_io, _parent, _root)
@_is_le = _is_le
_read
end
def _read
if @_is_le == true
_read_le
elsif @_is_le == false
_read_be
else
raise Kaitai::Struct::UndecidedEndiannessError.new("/types/gguf_body/types/gguf_array")
end
self
end
def _read_le
@type = Kaitai::Struct::Stream::resolve_enum(Gguf::GGUF_TYPE, @_io.read_u4le)
raise Kaitai::Struct::ValidationNotInEnumError.new(@type, @_io, "/types/gguf_body/types/gguf_array/seq/0") if not Gguf::I__GGUF_TYPE.key?(@type)
@num_elems = @_io.read_u8le
@elems = []
(num_elems).times { |i|
@elems << GgufValue.new(@_io, self, @_root, @_is_le, type)
}
self
end
def _read_be
@type = Kaitai::Struct::Stream::resolve_enum(Gguf::GGUF_TYPE, @_io.read_u4be)
raise Kaitai::Struct::ValidationNotInEnumError.new(@type, @_io, "/types/gguf_body/types/gguf_array/seq/0") if not Gguf::I__GGUF_TYPE.key?(@type)
@num_elems = @_io.read_u8be
@elems = []
(num_elems).times { |i|
@elems << GgufValue.new(@_io, self, @_root, @_is_le, type)
}
self
end
attr_reader :type
attr_reader :num_elems
attr_reader :elems
end
class GgufBool < Kaitai::Struct::Struct
def initialize(_io, _parent = nil, _root = nil, _is_le = nil)
super(_io, _parent, _root)
@_is_le = _is_le
_read
end
def _read
if @_is_le == true
_read_le
elsif @_is_le == false
_read_be
else
raise Kaitai::Struct::UndecidedEndiannessError.new("/types/gguf_body/types/gguf_bool")
end
self
end
def _read_le
@value_raw = @_io.read_u1
raise Kaitai::Struct::ValidationGreaterThanError.new(1, @value_raw, @_io, "/types/gguf_body/types/gguf_bool/seq/0") if not @value_raw <= 1
self
end
def _read_be
@value_raw = @_io.read_u1
raise Kaitai::Struct::ValidationGreaterThanError.new(1, @value_raw, @_io, "/types/gguf_body/types/gguf_bool/seq/0") if not @value_raw <= 1
self
end
##
# @see https://github.com/ggml-org/ggml/blob/a05c9922613c6a2230dbb1d0de679519ecafb4f0/src/gguf.cpp#L321 Git tag "v0.27.0"
def value
return @value unless @value.nil?
@value = value_raw != 0
@value
end
attr_reader :value_raw
end
class GgufFloat32 < Kaitai::Struct::Struct
def initialize(_io, _parent = nil, _root = nil, _is_le = nil)
super(_io, _parent, _root)
@_is_le = _is_le
_read
end
def _read
if @_is_le == true
_read_le
elsif @_is_le == false
_read_be
else
raise Kaitai::Struct::UndecidedEndiannessError.new("/types/gguf_body/types/gguf_float32")
end
self
end
def _read_le
@value = @_io.read_f4le
self
end
def _read_be
@value = @_io.read_f4be
self
end
attr_reader :value
end
class GgufFloat64 < Kaitai::Struct::Struct
def initialize(_io, _parent = nil, _root = nil, _is_le = nil)
super(_io, _parent, _root)
@_is_le = _is_le
_read
end
def _read
if @_is_le == true
_read_le
elsif @_is_le == false
_read_be
else
raise Kaitai::Struct::UndecidedEndiannessError.new("/types/gguf_body/types/gguf_float64")
end
self
end
def _read_le
@value = @_io.read_f8le
self
end
def _read_be
@value = @_io.read_f8be
self
end
attr_reader :value
end
class GgufInt16 < Kaitai::Struct::Struct
def initialize(_io, _parent = nil, _root = nil, _is_le = nil)
super(_io, _parent, _root)
@_is_le = _is_le
_read
end
def _read
if @_is_le == true
_read_le
elsif @_is_le == false
_read_be
else
raise Kaitai::Struct::UndecidedEndiannessError.new("/types/gguf_body/types/gguf_int16")
end
self
end
def _read_le
@value = @_io.read_s2le
self
end
def _read_be
@value = @_io.read_s2be
self
end
attr_reader :value
end
class GgufInt32 < Kaitai::Struct::Struct
def initialize(_io, _parent = nil, _root = nil, _is_le = nil)
super(_io, _parent, _root)
@_is_le = _is_le
_read
end
def _read
if @_is_le == true
_read_le
elsif @_is_le == false
_read_be
else
raise Kaitai::Struct::UndecidedEndiannessError.new("/types/gguf_body/types/gguf_int32")
end
self
end
def _read_le
@value = @_io.read_s4le
self
end
def _read_be
@value = @_io.read_s4be
self
end
attr_reader :value
end
class GgufInt64 < Kaitai::Struct::Struct
def initialize(_io, _parent = nil, _root = nil, _is_le = nil)
super(_io, _parent, _root)
@_is_le = _is_le
_read
end
def _read
if @_is_le == true
_read_le
elsif @_is_le == false
_read_be
else
raise Kaitai::Struct::UndecidedEndiannessError.new("/types/gguf_body/types/gguf_int64")
end
self
end
def _read_le
@value = @_io.read_s8le
self
end
def _read_be
@value = @_io.read_s8be
self
end
attr_reader :value
end
class GgufInt8 < Kaitai::Struct::Struct
def initialize(_io, _parent = nil, _root = nil, _is_le = nil)
super(_io, _parent, _root)
@_is_le = _is_le
_read
end
def _read
if @_is_le == true
_read_le
elsif @_is_le == false
_read_be
else
raise Kaitai::Struct::UndecidedEndiannessError.new("/types/gguf_body/types/gguf_int8")
end
self
end
def _read_le
@value = @_io.read_s1
self
end
def _read_be
@value = @_io.read_s1
self
end
attr_reader :value
end
##
# @see https://github.com/ggml-org/ggml/blob/a05c9922613c6a2230dbb1d0de679519ecafb4f0/docs/gguf.md?plain=1#L262-L275 Git tag "v0.27.0"
class GgufKv < Kaitai::Struct::Struct
def initialize(_io, _parent = nil, _root = nil, _is_le = nil, prev_alignment)
super(_io, _parent, _root)
@_is_le = _is_le
@prev_alignment = prev_alignment
_read
end
def _read
if @_is_le == true
_read_le
elsif @_is_le == false
_read_be
else
raise Kaitai::Struct::UndecidedEndiannessError.new("/types/gguf_body/types/gguf_kv")
end
self
end
def _read_le
@key = GgufStr.new(@_io, self, @_root, @_is_le)
_ = @key
raise Kaitai::Struct::ValidationExprError.new(@key, @_io, "/types/gguf_body/types/gguf_kv/seq/0") if not _.len_value != 0
@type = Kaitai::Struct::Stream::resolve_enum(Gguf::GGUF_TYPE, @_io.read_u4le)
raise Kaitai::Struct::ValidationNotInEnumError.new(@type, @_io, "/types/gguf_body/types/gguf_kv/seq/1") if not Gguf::I__GGUF_TYPE.key?(@type)
@value = GgufValue.new(@_io, self, @_root, @_is_le, type)
_ = @value
raise Kaitai::Struct::ValidationExprError.new(@value, @_io, "/types/gguf_body/types/gguf_kv/seq/2") if not (key.value == "general.alignment" ? ((type == :gguf_type_uint32) && (_.value.value != 0) && (_.value.value & _.value.value - 1 == 0)) : true)
self
end
def _read_be
@key = GgufStr.new(@_io, self, @_root, @_is_le)
_ = @key
raise Kaitai::Struct::ValidationExprError.new(@key, @_io, "/types/gguf_body/types/gguf_kv/seq/0") if not _.len_value != 0
@type = Kaitai::Struct::Stream::resolve_enum(Gguf::GGUF_TYPE, @_io.read_u4be)
raise Kaitai::Struct::ValidationNotInEnumError.new(@type, @_io, "/types/gguf_body/types/gguf_kv/seq/1") if not Gguf::I__GGUF_TYPE.key?(@type)
@value = GgufValue.new(@_io, self, @_root, @_is_le, type)
_ = @value
raise Kaitai::Struct::ValidationExprError.new(@value, @_io, "/types/gguf_body/types/gguf_kv/seq/2") if not (key.value == "general.alignment" ? ((type == :gguf_type_uint32) && (_.value.value != 0) && (_.value.value & _.value.value - 1 == 0)) : true)
self
end
def alignment
return @alignment unless @alignment.nil?
@alignment = (key.value == "general.alignment" ? value.value.value : prev_alignment)
@alignment
end
attr_reader :key
attr_reader :type
attr_reader :value
attr_reader :prev_alignment
end
##
# @see https://github.com/ggml-org/ggml/blob/a05c9922613c6a2230dbb1d0de679519ecafb4f0/docs/gguf.md?plain=1#L232-L237 Git tag "v0.27.0"
class GgufStr < Kaitai::Struct::Struct
def initialize(_io, _parent = nil, _root = nil, _is_le = nil)
super(_io, _parent, _root)
@_is_le = _is_le
_read
end
def _read
if @_is_le == true
_read_le
elsif @_is_le == false
_read_be
else
raise Kaitai::Struct::UndecidedEndiannessError.new("/types/gguf_body/types/gguf_str")
end
self
end
def _read_le
@len_value = @_io.read_u8le
@value = (@_io.read_bytes(len_value)).force_encoding("UTF-8")
self
end
def _read_be
@len_value = @_io.read_u8be
@value = (@_io.read_bytes(len_value)).force_encoding("UTF-8")
self
end
attr_reader :len_value
attr_reader :value
end
##
# @see https://github.com/ggml-org/ggml/blob/a05c9922613c6a2230dbb1d0de679519ecafb4f0/docs/gguf.md?plain=1#L305-L325 Git tag "v0.27.0"
class GgufTensorInfo < Kaitai::Struct::Struct
def initialize(_io, _parent = nil, _root = nil, _is_le = nil, expected_offset)
super(_io, _parent, _root)
@_is_le = _is_le
@expected_offset = expected_offset
_read
end
def _read
if @_is_le == true
_read_le
elsif @_is_le == false
_read_be
else
raise Kaitai::Struct::UndecidedEndiannessError.new("/types/gguf_body/types/gguf_tensor_info")
end
self
end
def _read_le
@name = GgufStr.new(@_io, self, @_root, @_is_le)
_ = @name
raise Kaitai::Struct::ValidationExprError.new(@name, @_io, "/types/gguf_body/types/gguf_tensor_info/seq/0") if not _.len_value < 64
@num_dims = @_io.read_u4le
raise Kaitai::Struct::ValidationGreaterThanError.new(4, @num_dims, @_io, "/types/gguf_body/types/gguf_tensor_info/seq/1") if not @num_dims <= 4
@dims = []
(num_dims).times { |i|
@dims << @_io.read_s8le
raise Kaitai::Struct::ValidationLessThanError.new(0, @dims[i], @_io, "/types/gguf_body/types/gguf_tensor_info/seq/2") if not @dims[i] >= 0
}
@type = Kaitai::Struct::Stream::resolve_enum(Gguf::GGML_TYPE, @_io.read_u4le)
_ = @type
raise Kaitai::Struct::ValidationExprError.new(@type, @_io, "/types/gguf_body/types/gguf_tensor_info/seq/3") if not (((Gguf::I__GGML_TYPE[_] || _) < 43) && (num_elements_per_block != 0) && ((num_dims > 0 ? dims[0] : 1) % num_elements_per_block == 0))
@offset = @_io.read_u8le
raise Kaitai::Struct::ValidationNotEqualError.new(expected_offset, @offset, @_io, "/types/gguf_body/types/gguf_tensor_info/seq/4") if not @offset == expected_offset
self
end
def _read_be
@name = GgufStr.new(@_io, self, @_root, @_is_le)
_ = @name
raise Kaitai::Struct::ValidationExprError.new(@name, @_io, "/types/gguf_body/types/gguf_tensor_info/seq/0") if not _.len_value < 64
@num_dims = @_io.read_u4be
raise Kaitai::Struct::ValidationGreaterThanError.new(4, @num_dims, @_io, "/types/gguf_body/types/gguf_tensor_info/seq/1") if not @num_dims <= 4
@dims = []
(num_dims).times { |i|
@dims << @_io.read_s8be
raise Kaitai::Struct::ValidationLessThanError.new(0, @dims[i], @_io, "/types/gguf_body/types/gguf_tensor_info/seq/2") if not @dims[i] >= 0
}
@type = Kaitai::Struct::Stream::resolve_enum(Gguf::GGML_TYPE, @_io.read_u4be)
_ = @type
raise Kaitai::Struct::ValidationExprError.new(@type, @_io, "/types/gguf_body/types/gguf_tensor_info/seq/3") if not (((Gguf::I__GGML_TYPE[_] || _) < 43) && (num_elements_per_block != 0) && ((num_dims > 0 ? dims[0] : 1) % num_elements_per_block == 0))
@offset = @_io.read_u8be
raise Kaitai::Struct::ValidationNotEqualError.new(expected_offset, @offset, @_io, "/types/gguf_body/types/gguf_tensor_info/seq/4") if not @offset == expected_offset
self
end
def data
return @data unless @data.nil?
_pos = @_io.pos
@_io.seek(_parent.ofs_tensor_data + offset)
if @_is_le
@data = @_io.read_bytes(len_data)
else
@data = @_io.read_bytes(len_data)
end
@_io.seek(_pos)
@data
end
##
# `dims` joined with the string `" x "`. This is necessary for
# `-webide-representation`, which cannot format `dims` directly -
# see <https://github.com/kaitai-io/kaitai_struct_webide/issues/160>.
def dump_dims
return @dump_dims unless @dump_dims.nil?
@dump_dims = (((num_dims > 0 ? dims[0].to_s : "") + (num_dims > 1 ? " x " + dims[1].to_s : "")) + (num_dims > 2 ? " x " + dims[2].to_s : "")) + (num_dims > 3 ? " x " + dims[3].to_s : "")
@dump_dims
end
##
# `offset` that the next tensor must have
def expected_next_offset
return @expected_next_offset unless @expected_next_offset.nil?
@expected_next_offset = ((offset + len_data) + len_padding)
@expected_next_offset
end
##
# Size of a block in bytes
def len_block
return @len_block unless @len_block.nil?
@len_block = [4, 2, 18, 20, 0, 0, 22, 24, 34, 36, 84, 110, 144, 176, 210, 292, 66, 74, 98, 50, 18, 110, 82, 136, 1, 2, 4, 8, 8, 56, 2, 0, 0, 0, 54, 66, 0, 0, 0, 17, 36, 18, 18][(Gguf::I__GGML_TYPE[type] || type)]
@len_block
end
def len_data
return @len_data unless @len_data.nil?
@len_data = (len_block * num_blocks)
@len_data
end
##
# Number of padding bytes after the data
def len_padding
return @len_padding unless @len_padding.nil?
@len_padding = (_parent.alignment - len_data % _parent.alignment) % _parent.alignment
@len_padding
end
def num_blocks
return @num_blocks unless @num_blocks.nil?
@num_blocks = (num_elements / num_elements_per_block)
@num_blocks
end
def num_elements
return @num_elements unless @num_elements.nil?
@num_elements = ((((num_dims > 0 ? dims[0] : 1) * (num_dims > 1 ? dims[1] : 1)) * (num_dims > 2 ? dims[2] : 1)) * (num_dims > 3 ? dims[3] : 1))
@num_elements
end
##
# Number of elements in a block
def num_elements_per_block
return @num_elements_per_block unless @num_elements_per_block.nil?
@num_elements_per_block = [1, 1, 32, 32, 0, 0, 32, 32, 32, 32, 256, 256, 256, 256, 256, 256, 256, 256, 256, 256, 32, 256, 256, 256, 1, 1, 1, 1, 1, 256, 1, 0, 0, 0, 256, 256, 0, 0, 0, 32, 64, 128, 64][(Gguf::I__GGML_TYPE[type] || type)]
@num_elements_per_block
end
##
# Padding after the data. ggml requires it even after the last tensor
# when reading the tensor data.
# @see https://github.com/ggml-org/ggml/blob/a05c9922613c6a2230dbb1d0de679519ecafb4f0/src/gguf.cpp#L884-L893 Git tag "v0.27.0"
def padding
return @padding unless @padding.nil?
_pos = @_io.pos
@_io.seek((_parent.ofs_tensor_data + offset) + len_data)
if @_is_le
@padding = @_io.read_bytes(len_padding)
else
@padding = @_io.read_bytes(len_padding)
end
@_io.seek(_pos)
@padding
end
attr_reader :name
attr_reader :num_dims
attr_reader :dims
attr_reader :type
##
# Offset of the tensor's data, relative to `_parent.ofs_tensor_data`
attr_reader :offset
attr_reader :expected_offset
end
class GgufUint16 < Kaitai::Struct::Struct
def initialize(_io, _parent = nil, _root = nil, _is_le = nil)
super(_io, _parent, _root)
@_is_le = _is_le
_read
end
def _read
if @_is_le == true
_read_le
elsif @_is_le == false
_read_be
else
raise Kaitai::Struct::UndecidedEndiannessError.new("/types/gguf_body/types/gguf_uint16")
end
self
end
def _read_le
@value = @_io.read_u2le
self
end
def _read_be
@value = @_io.read_u2be
self
end
attr_reader :value
end
class GgufUint32 < Kaitai::Struct::Struct
def initialize(_io, _parent = nil, _root = nil, _is_le = nil)
super(_io, _parent, _root)
@_is_le = _is_le
_read
end
def _read
if @_is_le == true
_read_le
elsif @_is_le == false
_read_be
else
raise Kaitai::Struct::UndecidedEndiannessError.new("/types/gguf_body/types/gguf_uint32")
end
self
end
def _read_le
@value = @_io.read_u4le
self
end
def _read_be
@value = @_io.read_u4be
self
end
attr_reader :value
end
class GgufUint64 < Kaitai::Struct::Struct
def initialize(_io, _parent = nil, _root = nil, _is_le = nil)
super(_io, _parent, _root)
@_is_le = _is_le
_read
end
def _read
if @_is_le == true
_read_le
elsif @_is_le == false
_read_be
else
raise Kaitai::Struct::UndecidedEndiannessError.new("/types/gguf_body/types/gguf_uint64")
end
self
end
def _read_le
@value = @_io.read_u8le
self
end
def _read_be
@value = @_io.read_u8be
self
end
attr_reader :value
end
class GgufUint8 < Kaitai::Struct::Struct
def initialize(_io, _parent = nil, _root = nil, _is_le = nil)
super(_io, _parent, _root)
@_is_le = _is_le
_read
end
def _read
if @_is_le == true
_read_le
elsif @_is_le == false
_read_be
else
raise Kaitai::Struct::UndecidedEndiannessError.new("/types/gguf_body/types/gguf_uint8")
end
self
end
def _read_le
@value = @_io.read_u1
self
end
def _read_be
@value = @_io.read_u1
self
end
attr_reader :value
end
##
# @see https://github.com/ggml-org/ggml/blob/a05c9922613c6a2230dbb1d0de679519ecafb4f0/docs/gguf.md?plain=1#L239-L260 Git tag "v0.27.0"
class GgufValue < Kaitai::Struct::Struct
def initialize(_io, _parent = nil, _root = nil, _is_le = nil, type)
super(_io, _parent, _root)
@_is_le = _is_le
@type = type
_read
end
def _read
if @_is_le == true
_read_le
elsif @_is_le == false
_read_be
else
raise Kaitai::Struct::UndecidedEndiannessError.new("/types/gguf_body/types/gguf_value")
end
self
end
def _read_le
case type
when :gguf_type_array
@value = GgufArray.new(@_io, self, @_root, @_is_le)
when :gguf_type_bool
@value = GgufBool.new(@_io, self, @_root, @_is_le)
when :gguf_type_float32
@value = GgufFloat32.new(@_io, self, @_root, @_is_le)
when :gguf_type_float64
@value = GgufFloat64.new(@_io, self, @_root, @_is_le)
when :gguf_type_int16
@value = GgufInt16.new(@_io, self, @_root, @_is_le)
when :gguf_type_int32
@value = GgufInt32.new(@_io, self, @_root, @_is_le)
when :gguf_type_int64
@value = GgufInt64.new(@_io, self, @_root, @_is_le)
when :gguf_type_int8
@value = GgufInt8.new(@_io, self, @_root, @_is_le)
when :gguf_type_string
@value = GgufStr.new(@_io, self, @_root, @_is_le)
when :gguf_type_uint16
@value = GgufUint16.new(@_io, self, @_root, @_is_le)
when :gguf_type_uint32
@value = GgufUint32.new(@_io, self, @_root, @_is_le)
when :gguf_type_uint64
@value = GgufUint64.new(@_io, self, @_root, @_is_le)
when :gguf_type_uint8
@value = GgufUint8.new(@_io, self, @_root, @_is_le)
end
self
end
def _read_be
case type
when :gguf_type_array
@value = GgufArray.new(@_io, self, @_root, @_is_le)
when :gguf_type_bool
@value = GgufBool.new(@_io, self, @_root, @_is_le)
when :gguf_type_float32
@value = GgufFloat32.new(@_io, self, @_root, @_is_le)
when :gguf_type_float64
@value = GgufFloat64.new(@_io, self, @_root, @_is_le)
when :gguf_type_int16
@value = GgufInt16.new(@_io, self, @_root, @_is_le)
when :gguf_type_int32
@value = GgufInt32.new(@_io, self, @_root, @_is_le)
when :gguf_type_int64
@value = GgufInt64.new(@_io, self, @_root, @_is_le)
when :gguf_type_int8
@value = GgufInt8.new(@_io, self, @_root, @_is_le)
when :gguf_type_string
@value = GgufStr.new(@_io, self, @_root, @_is_le)
when :gguf_type_uint16
@value = GgufUint16.new(@_io, self, @_root, @_is_le)
when :gguf_type_uint32
@value = GgufUint32.new(@_io, self, @_root, @_is_le)
when :gguf_type_uint64
@value = GgufUint64.new(@_io, self, @_root, @_is_le)
when :gguf_type_uint8
@value = GgufUint8.new(@_io, self, @_root, @_is_le)
end
self
end
attr_reader :value
attr_reader :type
end
##
# Global alignment - from the `general.alignment` key, or 32 if there's no
# such key.
# @see https://github.com/ggml-org/ggml/blob/a05c9922613c6a2230dbb1d0de679519ecafb4f0/src/gguf.cpp#L622-L630 Git tag "v0.27.0"
def alignment
return @alignment unless @alignment.nil?
@alignment = (kv_pairs.length != 0 ? kv_pairs.last.alignment : 32)
@alignment
end
##
# Offset of the tensor data. Each tensor's `offset` is relative to this
# position.
def ofs_tensor_data
return @ofs_tensor_data unless @ofs_tensor_data.nil?
@ofs_tensor_data = _io.pos
@ofs_tensor_data
end
##
# File format version. Version 1 used 32-bit counts and lengths, which
# this Kaitai Struct implementation doesn't support, and neither does
# ggml.
attr_reader :version
##
# The number of tensors in the file
attr_reader :num_tensor_infos
##
# The number of key-value pairs in the file header
attr_reader :num_kv_pairs
##
# Key-value pairs. ggml (v0.27.0)
# [rejects](https://github.com/ggml-org/ggml/blob/a05c9922613c6a2230dbb1d0de679519ecafb4f0/src/gguf.cpp#L575-L578)
# duplicate keys, but this Kaitai Struct implementation can't detect
# them. If there are multiple `general.alignment` keys, the last one is
# used for the `alignment` value instance.
attr_reader :kv_pairs
##
# Tensor metadata. ggml (v0.27.0)
# [rejects](https://github.com/ggml-org/ggml/blob/a05c9922613c6a2230dbb1d0de679519ecafb4f0/src/gguf.cpp#L664-L669)
# duplicate tensor names, but this Kaitai Struct implementation can't
# detect them.
attr_reader :tensor_infos
##
# @see https://github.com/ggml-org/ggml/blob/a05c9922613c6a2230dbb1d0de679519ecafb4f0/src/gguf.cpp#L780-L785 Git tag "v0.27.0"
attr_reader :padding
attr_reader :_unnamed6
end
def is_big_endian
return @is_big_endian unless @is_big_endian.nil?
@is_big_endian = version_le & 65535 == 0
@is_big_endian
end
##
# Length of the file (computed from the header), including the padding after
# the last tensor's data.
#
# This Kaitai Struct implementation stops parsing where the tensor data
# starts (which is read lazily, i.e. only on demand). Without reading the
# tensor data, you can use this expected file size to check whether the file
# has been truncated, to carve the file out of a larger one, etc.
def len_file
return @len_file unless @len_file.nil?
@len_file = (body.ofs_tensor_data + (body.num_tensor_infos != 0 ? body.tensor_infos.last.expected_next_offset : 0))
@len_file
end
def version_le
return @version_le unless @version_le.nil?
_pos = @_io.pos
@_io.seek(4)
@version_le = @_io.read_u4le
@_io.seek(_pos)
@version_le
end
##
# GGUF file magic
attr_reader :magic
attr_reader :body
end