candle/candle-core/src/quantized/mod.rs

use crate::{Device, Result, Shape, Tensor};

#[cfg(target_feature = "avx")]
pub mod avx;
pub mod ggml_file;
pub mod gguf_file;
pub mod k_quants;
#[cfg(target_feature = "neon")]
pub mod neon;
pub mod utils;

pub use k_quants::GgmlType;

pub struct QTensor {
    data: Box<dyn QuantizedType>,
    shape: Shape,
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum GgmlDType {
    F32,
    F16,
    Q4_0,
    Q4_1,
    Q5_0,
    Q5_1,
    Q8_0,
    Q8_1,
    Q2K,
    Q3K,
    Q4K,
    Q5K,
    Q6K,
    Q8K,
}

impl GgmlDType {
    pub(crate) fn from_u32(u: u32) -> Result<Self> {
        let dtype = match u {
            0 => Self::F32,
            1 => Self::F16,
            2 => Self::Q4_0,
            3 => Self::Q4_1,
            6 => Self::Q5_0,
            7 => Self::Q5_1,
            8 => Self::Q8_0,
            9 => Self::Q8_1,
            10 => Self::Q2K,
            11 => Self::Q3K,
            12 => Self::Q4K,
            13 => Self::Q5K,
            14 => Self::Q6K,
            15 => Self::Q8K,
            _ => crate::bail!("unknown dtype for tensor {u}"),
        };
        Ok(dtype)
    }

    /// The type size for blocks in bytes.
    pub fn type_size(&self) -> usize {
        use k_quants::*;
        match self {
            Self::F32 => 4,
            Self::F16 => 2,
            Self::Q4_0 => std::mem::size_of::<BlockQ4_0>(),
            Self::Q4_1 => std::mem::size_of::<BlockQ4_1>(),
            Self::Q5_0 => std::mem::size_of::<BlockQ5_0>(),
            Self::Q5_1 => std::mem::size_of::<BlockQ5_1>(),
            // https://github.com/ggerganov/llama.cpp/blob/468ea24fb4633a0d681f7ac84089566c1c6190cb/ggml.c#L932
            Self::Q8_0 => std::mem::size_of::<BlockQ8_0>(),
            Self::Q8_1 => std::mem::size_of::<BlockQ8_1>(),
            Self::Q2K => std::mem::size_of::<BlockQ2K>(),
            Self::Q3K => std::mem::size_of::<BlockQ3K>(),
            Self::Q4K => std::mem::size_of::<BlockQ4K>(),
            Self::Q5K => std::mem::size_of::<BlockQ5K>(),
            Self::Q6K => std::mem::size_of::<BlockQ6K>(),
            Self::Q8K => std::mem::size_of::<BlockQ8K>(),
        }
    }

    /// The block size, i.e. the number of elements stored in each block.
    pub fn blck_size(&self) -> usize {
        match self {
            Self::F32 => 1,
            Self::F16 => 1,
            Self::Q4_0 => k_quants::QK4_0,
            Self::Q4_1 => k_quants::QK4_1,
            Self::Q5_0 => k_quants::QK5_0,
            Self::Q5_1 => k_quants::QK5_1,
            Self::Q8_0 => k_quants::QK8_0,
            Self::Q8_1 => k_quants::QK8_1,
            Self::Q2K | Self::Q3K | Self::Q4K | Self::Q5K | Self::Q6K | Self::Q8K => k_quants::QK_K,
        }
    }
}

// A version of GgmlType without `vec_dot` so that it can be dyn boxed.
pub trait QuantizedType: Send + Sync {
    fn dtype(&self) -> GgmlDType;
    fn matmul_t(&self, mkn: (usize, usize, usize), lhs: &[f32], dst: &mut [f32]) -> Result<()>;
    fn to_float(&self, ys: &mut [f32]) -> Result<()>;
}

impl<T: k_quants::GgmlType + Send + Sync> QuantizedType for Vec<T> {
    fn matmul_t(&self, mkn: (usize, usize, usize), lhs: &[f32], dst: &mut [f32]) -> Result<()> {
        k_quants::matmul(mkn, lhs, self.as_slice(), dst)
    }

    fn dtype(&self) -> GgmlDType {
        T::DTYPE
    }

    fn to_float(&self, ys: &mut [f32]) -> Result<()> {
        T::to_float(self.as_slice(), ys)
    }
}

impl std::fmt::Debug for QTensor {
    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
        write!(f, "QTensor[{:?}; {:?}]", self.shape, self.dtype())
    }
}

fn check_shape<T: k_quants::GgmlType>(shape: &Shape) -> Result<()> {
    let dims = shape.dims();
    if dims.is_empty() {
        crate::bail!("scalar tensor cannot be quantized {shape:?}")
    }
    if dims[dims.len() - 1] % T::BLCK_SIZE != 0 {
        crate::bail!(
            "quantized tensor must have their last dim divisible by block size {shape:?} {}",
            T::BLCK_SIZE
        )
    }
    Ok(())
}

impl QTensor {
    pub fn new<S: Into<Shape>, T: k_quants::GgmlType + Send + Sync + 'static>(
        data: Vec<T>,
        shape: S,
    ) -> Result<Self> {
        let shape = shape.into();
        check_shape::<T>(&shape)?;
        Ok(Self {
            data: Box::new(data),
            shape,
        })
    }

    pub fn quantize<T: k_quants::GgmlType + Send + Sync + 'static>(src: &Tensor) -> Result<Self> {
        let shape = src.shape();
        check_shape::<T>(shape)?;
        let src = src
            .to_dtype(crate::DType::F32)?
            .flatten_all()?
            .to_vec1::<f32>()?;
        if src.len() % T::BLCK_SIZE != 0 {
            crate::bail!(
                "tensor size ({shape:?}) is not divisible by block size {}",
                T::BLCK_SIZE
            )
        }
        let mut data = vec![T::zeros(); src.len() / T::BLCK_SIZE];
        T::from_float(&src, &mut data)?;
        Ok(Self {
            data: Box::new(data),
            shape: shape.clone(),
        })
    }

    pub fn dtype(&self) -> GgmlDType {
        self.data.dtype()
    }

    pub fn shape(&self) -> &Shape {
        &self.shape
    }

    pub fn dequantize(&self, device: &Device) -> Result<Tensor> {
        let mut f32_data = vec![0f32; self.shape.elem_count()];
        self.data.to_float(&mut f32_data)?;
        Tensor::from_vec(f32_data, &self.shape, device)
    }

    pub fn matmul_t(&self, mkn: (usize, usize, usize), lhs: &[f32], dst: &mut [f32]) -> Result<()> {
        self.data.matmul_t(mkn, lhs, dst)
    }
}

#[derive(Debug)]
pub struct QMatMul(QTensor);

impl QMatMul {
    pub fn from_qtensor(qtensor: QTensor) -> Self {
        Self(qtensor)
    }
}

impl crate::CustomOp1 for QTensor {
    fn name(&self) -> &'static str {
        "qmatmul"
    }

    fn cpu_fwd(
        &self,
        storage: &crate::CpuStorage,
        layout: &crate::Layout,
    ) -> Result<(crate::CpuStorage, Shape)> {
        if !layout.is_contiguous() {
            crate::bail!("input tensor is not contiguous {layout:?}")
        }
        let src_shape = layout.shape();
        // self is transposed so n is first then k.
        let (n, k) = self.shape.dims2()?;
        if src_shape.rank() < 2 {
            crate::bail!("input tensor has only one dimension {layout:?}")
        }
        let mut dst_shape = src_shape.dims().to_vec();
        let last_k = dst_shape.pop().unwrap();
        if last_k != k {
            crate::bail!("input tensor {layout:?} incompatible with {:?}", self.shape)
        }
        dst_shape.push(n);
        let dst_shape = Shape::from(dst_shape);
        let storage = storage.as_slice::<f32>()?;
        let storage =
            &storage[layout.start_offset()..layout.start_offset() + src_shape.elem_count()];
        let mut dst_storage = vec![0f32; dst_shape.elem_count()];
        self.matmul_t(
            (dst_shape.elem_count() / n, k, n),
            storage,
            &mut dst_storage,
        )?;
        Ok((crate::CpuStorage::F32(dst_storage), dst_shape))
    }
}

impl QMatMul {
    pub fn forward(&self, xs: &Tensor) -> Result<Tensor> {
        xs.apply_op1_no_bwd(&self.0)
    }
}