llama-gguf 0.14.0

//! GGUF format constants

/// GGUF magic number: "GGUF" in little-endian
pub const GGUF_MAGIC: u32 = 0x46554747;

/// Supported GGUF versions
pub const GGUF_VERSION_V1: u32 = 1;
pub const GGUF_VERSION_V2: u32 = 2;
pub const GGUF_VERSION_V3: u32 = 3;

/// Default alignment for tensor data
pub const GGUF_DEFAULT_ALIGNMENT: usize = 32;

/// Metadata value types
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
#[repr(u32)]
pub enum GgufMetadataValueType {
    Uint8 = 0,
    Int8 = 1,
    Uint16 = 2,
    Int16 = 3,
    Uint32 = 4,
    Int32 = 5,
    Float32 = 6,
    Bool = 7,
    String = 8,
    Array = 9,
    Uint64 = 10,
    Int64 = 11,
    Float64 = 12,
}

impl TryFrom<u32> for GgufMetadataValueType {
    type Error = u32;
    fn try_from(value: u32) -> Result<Self, Self::Error> {
        match value {
            0 => Ok(Self::Uint8),
            1 => Ok(Self::Int8),
            2 => Ok(Self::Uint16),
            3 => Ok(Self::Int16),
            4 => Ok(Self::Uint32),
            5 => Ok(Self::Int32),
            6 => Ok(Self::Float32),
            7 => Ok(Self::Bool),
            8 => Ok(Self::String),
            9 => Ok(Self::Array),
            10 => Ok(Self::Uint64),
            11 => Ok(Self::Int64),
            12 => Ok(Self::Float64),
            _ => Err(value),
        }
    }
}

/// GGML tensor types (used in GGUF tensor info)
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
#[repr(u32)]
pub enum GgmlType {
    F32 = 0,
    F16 = 1,
    Q4_0 = 2,
    Q4_1 = 3,
    Q5_0 = 6,
    Q5_1 = 7,
    Q8_0 = 8,
    Q8_1 = 9,
    Q2K = 10,
    Q3K = 11,
    Q4K = 12,
    Q5K = 13,
    Q6K = 14,
    Q8K = 15,
    IQ2XXS = 16,
    IQ2XS = 17,
    IQ3XXS = 18,
    IQ1S = 19,
    IQ4NL = 20,
    IQ3S = 21,
    IQ2S = 22,
    IQ4XS = 23,
    I8 = 24,
    I16 = 25,
    I32 = 26,
    I64 = 27,
    F64 = 28,
    IQ1M = 29,
    BF16 = 30,
    TQ1_0 = 34,
    TQ2_0 = 35,
}

impl TryFrom<u32> for GgmlType {
    type Error = u32;
    fn try_from(value: u32) -> Result<Self, Self::Error> {
        match value {
            0 => Ok(Self::F32),
            1 => Ok(Self::F16),
            2 => Ok(Self::Q4_0),
            3 => Ok(Self::Q4_1),
            6 => Ok(Self::Q5_0),
            7 => Ok(Self::Q5_1),
            8 => Ok(Self::Q8_0),
            9 => Ok(Self::Q8_1),
            10 => Ok(Self::Q2K),
            11 => Ok(Self::Q3K),
            12 => Ok(Self::Q4K),
            13 => Ok(Self::Q5K),
            14 => Ok(Self::Q6K),
            15 => Ok(Self::Q8K),
            16 => Ok(Self::IQ2XXS),
            17 => Ok(Self::IQ2XS),
            18 => Ok(Self::IQ3XXS),
            19 => Ok(Self::IQ1S),
            20 => Ok(Self::IQ4NL),
            21 => Ok(Self::IQ3S),
            22 => Ok(Self::IQ2S),
            23 => Ok(Self::IQ4XS),
            24 => Ok(Self::I8),
            25 => Ok(Self::I16),
            26 => Ok(Self::I32),
            27 => Ok(Self::I64),
            28 => Ok(Self::F64),
            29 => Ok(Self::IQ1M),
            30 => Ok(Self::BF16),
            34 => Ok(Self::TQ1_0),
            35 => Ok(Self::TQ2_0),
            _ => Err(value),
        }
    }
}

impl GgmlType {
    /// Block size for this type (number of elements per block)
    pub const fn block_size(&self) -> usize {
        match self {
            Self::F32 | Self::F16 | Self::BF16 | Self::F64 => 1,
            Self::I8 | Self::I16 | Self::I32 | Self::I64 => 1,
            Self::Q4_0 | Self::Q4_1 | Self::Q5_0 | Self::Q5_1 | Self::Q8_0 | Self::Q8_1 => 32,
            Self::Q2K | Self::Q3K | Self::Q4K | Self::Q5K | Self::Q6K | Self::Q8K => 256,
            Self::IQ2XXS
            | Self::IQ2XS
            | Self::IQ2S
            | Self::IQ3XXS
            | Self::IQ3S
            | Self::IQ4XS
            | Self::IQ4NL
            | Self::IQ1S
            | Self::IQ1M => 256,
            Self::TQ1_0 => 256,
            Self::TQ2_0 => 256,
        }
    }

    /// Bytes per block for this type
    pub const fn type_size(&self) -> usize {
        match self {
            Self::F32 => 4,
            Self::F16 | Self::BF16 => 2,
            Self::F64 => 8,
            Self::I8 => 1,
            Self::I16 => 2,
            Self::I32 => 4,
            Self::I64 => 8,
            Self::Q4_0 => 18,
            Self::Q4_1 => 20,
            Self::Q5_0 => 22,
            Self::Q5_1 => 24,
            Self::Q8_0 => 34,
            Self::Q8_1 => 36,
            Self::Q2K => 84,
            Self::Q3K => 110,
            Self::Q4K => 144,
            Self::Q5K => 176,
            Self::Q6K => 210,
            Self::Q8K => 292,
            Self::IQ2XXS => 66,
            Self::IQ2XS => 74,
            Self::IQ2S => 82,
            Self::IQ3XXS => 98,
            Self::IQ3S => 110,
            Self::IQ4XS => 136,
            Self::IQ4NL => 132,
            Self::IQ1S => 50,
            Self::IQ1M => 56,
            Self::TQ1_0 => 54,
            Self::TQ2_0 => 66,
        }
    }
}