mistralrs-quant 0.8.1

Fast, flexible LLM inference.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153

use candle_core::{DType, Result, Tensor};
use candle_nn::Linear;
use float8::F8E4M3;

use super::FP8Linear;

pub(super) struct QuantizationResult {
    /// Quantized tensor (f8)
    pub(super) qw: Tensor,
    /// Scalar, f32 tensor.
    ///
    /// Convert unquantized to quantized tensor as follows:
    /// `q = x * qs`
    pub(super) quantize_scale: Tensor,
    /// Scalar, f32 tensor. Reciprocal of `quantize_scale`.
    ///
    /// Convert unquantized to quantized tensor as follows:
    /// `x = q * dqs`
    pub(super) dequantize_scale: Tensor,
}

impl FP8Linear {
    pub(super) fn quantize(data: &Tensor, dtype: DType) -> Result<QuantizationResult> {
        let data = data.to_dtype(DType::BF16)?;
        let mut absmax = data.abs()?;
        while !absmax.dims().is_empty() {
            absmax = absmax.max(0)?;
        }

        let max_v = F8E4M3::MAX.to_f64();
        let scale = (max_v / absmax)?
            .clamp(F8E4M3::MIN.to_f32(), F8E4M3::MAX.to_f32())?
            .to_dtype(DType::F32)?;
        let to_cast = data.broadcast_mul(&scale.to_dtype(data.dtype())?)?;
        let qw = if dtype == DType::F8E4M3 {
            crate::scalar_fp8::ops::dtype_to_fp8(&to_cast)?
        } else {
            to_cast.to_dtype(dtype)?
        };
        Ok(QuantizationResult {
            qw,
            quantize_scale: scale.clone(),
            dequantize_scale: scale.recip()?,
        })
    }

    pub(super) fn dequantize(&self, dtype: DType) -> Result<Linear> {
        let dequant_w = self
            .lin
            .weight()
            .to_dtype(dtype)?
            .broadcast_mul(&self.dequant_w_scale.to_dtype(dtype)?)?;
        Ok(Linear::new(dequant_w, self.lin.bias().cloned()))
    }
}

#[cfg(test)]
mod tests {
    #[cfg(not(feature = "metal"))]
    use candle_core::{
        quantized::{GgmlDType, QTensor},
        DType, Device, Result, Tensor,
    };

    #[cfg(not(feature = "metal"))]
    use crate::fp8::FP8Linear;

    #[cfg(not(feature = "metal"))]
    use super::QuantizationResult;

    #[test]
    #[cfg(not(feature = "metal"))]
    fn test_roundtrip_f8e4m3() -> Result<()> {
        let dev = Device::cuda_if_available(0)?;

        let data = Tensor::rand(0f32, 1f32, (32, 32), &dev)?;

        let QuantizationResult {
            qw,
            quantize_scale: _,
            dequantize_scale,
        } = FP8Linear::quantize(&data, DType::F8E4M3)?;

        let dequant = crate::scalar_fp8::ops::fp8_to_dtype(&qw, DType::F32)?
            .broadcast_mul(&dequantize_scale)?;

        let diff1 = (&data - dequant)?.abs()?.mean_all()?;

        println!("{diff1}");

        let q8_0 = QTensor::quantize(&data, GgmlDType::Q8_0)?.dequantize(&dev)?;
        let diff2 = (&data - q8_0)?.abs()?.mean_all()?;

        println!("{diff2}");
        Ok(())
    }

    #[test]
    #[cfg(feature = "cuda")]
    fn test_cublaslt_matmul() -> Result<()> {
        use crate::cublaslt::{maybe_init_cublas_lt_wrapper, CUBLASLT_CONTROLLER};
        let dev = Device::new_cuda(0)?;

        // Use 128x128 matrices for FP8 tensor core compatibility across GPU architectures
        let w = Tensor::rand(0., 1., (1, 128, 128), &dev)?
            .to_dtype(DType::F32)?
            .contiguous()?;
        let mut x = Tensor::rand(0., 1., (1, 128, 128), &dev)?
            .to_dtype(DType::F32)?
            .contiguous()?;

        // Batch matrix multiplication
        maybe_init_cublas_lt_wrapper(x.device().clone());

        let handle = CUBLASLT_CONTROLLER.get_for_device(x.device()).unwrap();

        let QuantizationResult {
            qw,
            quantize_scale: quant_scale,
            dequantize_scale: dequant_a_scale,
        } = FP8Linear::quantize(&w, DType::F8E4M3)?;

        let mut dequant_b_scale = dequant_a_scale.clone();
        if !matches!(x.dtype(), DType::F8E4M3) {
            let QuantizationResult {
                qw,
                quantize_scale: _,
                dequantize_scale,
            } = FP8Linear::quantize(&x, DType::F8E4M3)?;
            x = qw;
            dequant_b_scale = dequantize_scale;
        }

        let a = qw;
        let b = x;

        // FP8 quantized matmul
        let _res = handle.batch_matmul_f8(
            &a,
            &b,
            &dequant_a_scale,
            &dequant_b_scale,
            &quant_scale,
            None,
            None,
            None,
            None,
            None,
        )?;

        Ok(())
    }
}