mistralrs-quant 0.8.1

use std::sync::{atomic::AtomicUsize, Arc};

use candle_core::{quantized::GgmlDType, DType, Device, Result, Tensor};
use candle_nn::Linear;

mod ops;
pub use ops::{fp8_vector_dequantize, fp8_vector_quantize};

#[cfg(feature = "cuda")]
pub(crate) mod ffi;

use crate::{
    generate_isq, generate_isq_imatrix,
    hqq::{ISQ_HQQ_DEFAULT_OPT_STEPS, ISQ_HQQ_GROUP_SIZE},
    AfqBits, AfqGroupSize, AfqLayer, DummyLayer, FP8Linear, GgufMatMul, HqqAxis, HqqBits,
    HqqConfig, HqqLayer, IsqType, QuantMethod, QuantMethodConfig, QuantizeOntoGuard,
    QuantizedSerde, Shard, ShardedVarBuilder, UnquantLinear,
};

pub(crate) const VECTOR_SIZE: usize = 128;

#[derive(Debug)]
pub struct VectorFP8Linear {
    weight: Tensor,
    weight_scale_inv: Tensor,
    bias: Option<Tensor>,
    dequant_dtype: DType,
}

impl QuantMethod for VectorFP8Linear {
    fn new(method: QuantMethodConfig) -> candle_core::Result<Self>
    where
        Self: Sized,
    {
        match method {
            QuantMethodConfig::Gguf { .. }
            | QuantMethodConfig::GptqAwq { .. }
            | QuantMethodConfig::Hqq { .. }
            | QuantMethodConfig::Dummy
            | QuantMethodConfig::Unquantized(_)
            | QuantMethodConfig::Bnb { .. }
            | QuantMethodConfig::FP8 { .. }
            | QuantMethodConfig::BlockwiseFP8 { .. }
            | QuantMethodConfig::PerTensorFP8 { .. }
            | QuantMethodConfig::Afq { .. }
            | QuantMethodConfig::MXFP4 { .. } => unreachable!(),
        }
    }

    fn dequantize_w(&self) -> Result<candle_core::Tensor> {
        ops::fp8_vector_dequantize(&self.weight, &self.weight_scale_inv, self.dequant_dtype)
    }

    fn forward(&self, x: &Tensor) -> Result<Tensor> {
        // Dequantize matmul always.
        let weight = self.dequantize_w()?;
        // Dispatch to unquant. This uses some cublaslt for bias & on cuda always, so it is better
        let unquant = UnquantLinear::new(QuantMethodConfig::Unquantized(Linear::new(
            weight,
            self.bias.clone(),
        )))?;
        unquant.forward(x)
    }

    fn quantized_act_type(&self) -> Option<DType> {
        None
    }

    fn add_delta_w(&self, _delta: &Tensor) -> Result<Arc<dyn QuantMethod>> {
        candle_core::bail!("VectorFP8Linear does not support add_delta_w")
    }

    fn dtype_and_device(&self) -> (DType, candle_core::Device) {
        (DType::F8E4M3, self.weight.device().clone())
    }

    fn apply_isq(
        self: Arc<Self>,
        dtype: Option<IsqType>,
        device: Device,
        n_quantized: &AtomicUsize,
        imatrix_weight: Option<Vec<f32>>,
        guard: QuantizeOntoGuard,
    ) -> Result<Arc<dyn QuantMethod>> {
        let weight =
            ops::fp8_vector_dequantize(&self.weight, &self.weight_scale_inv, self.dequant_dtype)?;
        match dtype {
            Some(IsqType::HQQ4 | IsqType::HQQ8) => {
                let _acquired_quantize_guard = guard.acquire(&device);
                if imatrix_weight.is_some() {
                    candle_core::bail!("HQQ does not support imatrix.");
                }

                n_quantized.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
                let bits = match dtype.unwrap() {
                    IsqType::HQQ8 => HqqBits::Eight,
                    IsqType::HQQ4 => HqqBits::Four,
                    _ => unreachable!(),
                };
                let cfg = HqqConfig {
                    bits,
                    group_size: ISQ_HQQ_GROUP_SIZE.try_into()?,
                    axis: HqqAxis::Zero,
                    optimization_steps: ISQ_HQQ_DEFAULT_OPT_STEPS,
                    round_zeros: false,
                    channel_wise: true,
                };
                let res = HqqLayer::quantize(&weight.to_device(&device)?, &device, cfg)?;
                if let Some(bias) = &self.bias {
                    let bias = bias
                        .to_device(&device)?
                        .to_dtype(res.dtype_and_device().0)?;
                    Ok(Arc::new(res.with_bias(bias)))
                } else {
                    Ok(Arc::new(res))
                }
            }
            Some(IsqType::AFQ2 | IsqType::AFQ3 | IsqType::AFQ4 | IsqType::AFQ6 | IsqType::AFQ8) => {
                let _acquired_quantize_guard = guard.acquire(&device);
                if imatrix_weight.is_some() {
                    candle_core::bail!("AFQ does not support imatrix.");
                }

                n_quantized.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
                let bits = match dtype.unwrap() {
                    IsqType::AFQ8 => AfqBits::Eight,
                    IsqType::AFQ6 => AfqBits::Six,
                    IsqType::AFQ4 => AfqBits::Four,
                    IsqType::AFQ3 => AfqBits::Three,
                    IsqType::AFQ2 => AfqBits::Two,
                    _ => unreachable!(),
                };

                Ok(Arc::new(AfqLayer::new(QuantMethodConfig::Afq {
                    weight: weight.to_device(&device)?,
                    bias: self.bias.as_ref().map(|b| b.to_device(&device).unwrap()),
                    bits,
                    group_size: AfqGroupSize::default(),
                })?))
            }
            Some(
                IsqType::Q2K
                | IsqType::Q3K
                | IsqType::Q4K
                | IsqType::Q4_0
                | IsqType::Q4_1
                | IsqType::Q5K
                | IsqType::Q5_0
                | IsqType::Q5_1
                | IsqType::Q6K
                | IsqType::Q8K
                | IsqType::Q8_0
                | IsqType::Q8_1,
            ) => {
                let dtype: GgmlDType = dtype.unwrap().try_into()?;
                let res = if let Some(imatrix_weight) = imatrix_weight {
                    generate_isq_imatrix!(weight, imatrix_weight, device, dtype, n_quantized, guard)
                } else {
                    generate_isq!(weight, device, dtype, n_quantized, guard)
                };
                Ok(Arc::new(GgufMatMul::new(QuantMethodConfig::Gguf {
                    q_weight: res,
                    b: self
                        .bias
                        .as_ref()
                        .map(|b| b.to_dtype(DType::F32).unwrap().to_device(&device).unwrap()),
                })?))
            }
            Some(IsqType::F8E4M3) => {
                let _acquired_quantize_guard = guard.acquire(&device);
                if imatrix_weight.is_some() {
                    candle_core::bail!("F8E4M3 does not support imatrix.");
                }

                let w = weight.to_device(&device)?;
                let b = if let Some(b) = &self.bias {
                    Some(b.to_device(&device)?)
                } else {
                    None
                };
                Ok(Arc::new(FP8Linear::new(QuantMethodConfig::FP8 {
                    lin: Linear::new(w, b),
                    dtype: DType::F8E4M3,
                })?))
            }
            Some(IsqType::F8Q8) => {
                let _acquired_quantize_guard = guard.acquire(&device);
                if imatrix_weight.is_some() {
                    candle_core::bail!("F8Q8 does not support imatrix.");
                }

                let w = weight.to_device(&device)?;
                let b = if let Some(b) = &self.bias {
                    Some(b.to_device(&device)?)
                } else {
                    None
                };
                Ok(Arc::new(crate::F8Q8Linear::from_weight(&w, b)?))
            }
            Some(IsqType::MXFP4) => {
                let _acquired_quantize_guard = guard.acquire(&device);
                if imatrix_weight.is_some() {
                    candle_core::bail!("MXFP4 does not support imatrix.");
                }

                n_quantized.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
                let w = weight.to_device(&device)?;
                let b = self
                    .bias
                    .as_ref()
                    .map(|b| b.to_device(&device))
                    .transpose()?;
                crate::MXFP4Layer::quantize(&w, b, &device)
            }
            None => {
                let _acquired_quantize_guard = guard.acquire(&device);
                // Ignore imatrix altogether

                let w = weight.to_device(&device)?;
                let b = if let Some(b) = &self.bias {
                    Some(b.to_device(&device)?)
                } else {
                    None
                };
                Ok(Arc::new(UnquantLinear::new(
                    QuantMethodConfig::Unquantized(Linear::new(w, b)),
                )?))
            }
        }
    }
}

impl QuantizedSerde for VectorFP8Linear {
    fn isq_serde_supported(&self) -> bool {
        false
    }
    fn name(&self) -> &'static str {
        "vector-fp8-linear"
    }
}

#[allow(dead_code)]
pub fn vector_fp8_linear_b(
    in_dim: usize,
    out_dim: usize,
    bias: bool,
    hints: Shard,
    vb: ShardedVarBuilder,
) -> Result<Arc<dyn QuantMethod>> {
    // Check that dimensions are divisible by VECTOR_SIZE
    if !in_dim.is_multiple_of(VECTOR_SIZE) {
        candle_core::bail!(
            "Input dimension {} must be divisible by {} for vector FP8 quantization",
            in_dim,
            VECTOR_SIZE
        );
    }
    if !out_dim.is_multiple_of(VECTOR_SIZE) {
        candle_core::bail!(
            "Output dimension {} must be divisible by {} for vector FP8 quantization",
            out_dim,
            VECTOR_SIZE
        );
    }

    // Handle the case where we actually have an unquantized
    if vb.contains_tensor("weight") && !vb.contains_tensor("weight_scale_inv") {
        return crate::linear_b(in_dim, out_dim, bias, &None, vb);
    }

    // Handle the case where the layer is dummy (no tensors)
    if !(vb.contains_tensor("weight") && vb.contains_tensor("weight_scale_inv")) {
        let layer = <DummyLayer as QuantMethod>::new(QuantMethodConfig::Dummy)?;
        return Ok(Arc::new(layer) as Arc<dyn QuantMethod>);
    }

    let weight = vb.get_with_hints_dtype((out_dim, in_dim), "weight", hints, DType::F8E4M3)?;

    // For vector quantization, we have one scale per vector of 128 elements
    // Since weight is (out_dim, in_dim), we'll treat it as out_dim * in_dim / 128 vectors
    let total_elements = out_dim * in_dim;
    let num_vectors = total_elements.div_ceil(VECTOR_SIZE);

    let weight_scale_inv =
        vb.get_with_hints_dtype(num_vectors, "weight_scale_inv", hints, DType::F32)?;

    let bias = if bias {
        Some(vb.get((out_dim,), "bias")?)
    } else {
        None
    };

    Ok(Arc::new(VectorFP8Linear {
        weight,
        weight_scale_inv,
        bias,
        dequant_dtype: vb.dtype(),
    }))
}