libmir-cuda 0.1.0

CUDA inference backend for libmir
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

use mircuda::{DeviceBuffer, Stream, bf16};

use super::CudaBackend;
use crate::{
    CudaTensor, Error, Result,
    kernels::{
        BlockFp8LinearKernels, BlockFp8LinearSpec, Fp8OutputKernels, Fp8OutputSpec,
        Fp8ResidualWeightBuffers,
    },
};

/// Model-owned block-scaled E4M3 weight for BF16-input decode projection.
#[derive(Clone, Debug)]
pub struct BlockFp8LinearWeight {
    kernels: BlockFp8LinearKernels,
    weight: DeviceBuffer<u8>,
    scales: DeviceBuffer<f32>,
    spec: BlockFp8LinearSpec,
}

/// Model-owned E4M3 base plus block-scaled INT4 correction.
#[derive(Clone, Debug)]
pub struct Fp8ResidualLinearWeight {
    kernels: Fp8OutputKernels,
    weight: DeviceBuffer<u8>,
    row_scales: DeviceBuffer<f32>,
    residual: DeviceBuffer<u8>,
    residual_scales: DeviceBuffer<f32>,
}

impl CudaBackend {
    pub(in crate::backend) fn prepare_block_fp8_linear_weight(
        &self,
        source: &CudaTensor,
    ) -> Result<BlockFp8LinearWeight> {
        let [output_features, input_features] = source.shape() else {
            return Err(Error::InvalidLinearWeight {
                name: source.name().into(),
                expected: [0, 0],
                actual: source.shape().to_vec(),
            });
        };
        let source_weight = source.as_bf16().ok_or_else(|| Error::DTypeMismatch {
            name: source.name().into(),
            expected: "BF16",
        })?;
        let spec = BlockFp8LinearSpec::new(*input_features, *output_features)?;
        let kernels = BlockFp8LinearKernels::compile(&self.inner.compiler, spec)?;
        let mut weight =
            self.inner.pool.allocate::<u8>(&self.inner.stream, spec.weight_elements()?)?;
        let mut scales =
            self.inner.pool.allocate::<f32>(&self.inner.stream, spec.scale_elements()?)?;
        kernels.quantize(&self.inner.stream, source_weight, &mut weight, &mut scales)?;
        Ok(BlockFp8LinearWeight { kernels, weight, scales, spec })
    }

    pub(in crate::backend) fn prepare_fp8_residual_linear_weight(
        &self,
        source: &CudaTensor,
    ) -> Result<Fp8ResidualLinearWeight> {
        let [output_features, input_features] = source.shape() else {
            return Err(Error::InvalidLinearWeight {
                name: source.name().into(),
                expected: [0, 0],
                actual: source.shape().to_vec(),
            });
        };
        let source_weight = source.as_bf16().ok_or_else(|| Error::DTypeMismatch {
            name: source.name().into(),
            expected: "BF16",
        })?;
        let spec = Fp8OutputSpec::new(*input_features, *output_features)?;
        let kernels = Fp8OutputKernels::compile(&self.inner.compiler, spec)?;
        let mut weight =
            self.inner.pool.allocate::<u8>(&self.inner.stream, spec.weight_elements()?)?;
        let mut block_scales = self
            .inner
            .pool
            .allocate::<f32>(&self.inner.stream, spec.weight_scale_elements()?)?;
        let mut row_scales =
            self.inner.pool.allocate::<f32>(&self.inner.stream, spec.output_features)?;
        let mut residual =
            self.inner.pool.allocate::<u8>(&self.inner.stream, spec.residual_elements()?)?;
        let mut residual_scales = self
            .inner
            .pool
            .allocate::<f32>(&self.inner.stream, spec.residual_scale_elements()?)?;
        let mut buffers = Fp8ResidualWeightBuffers::new(
            &mut weight,
            &mut block_scales,
            &mut row_scales,
            &mut residual,
            &mut residual_scales,
        );
        kernels.quantize_weight_residual(&self.inner.stream, source_weight, &mut buffers)?;
        Ok(Fp8ResidualLinearWeight {
            kernels,
            weight,
            row_scales,
            residual,
            residual_scales,
        })
    }
}

impl BlockFp8LinearWeight {
    pub(in crate::backend) fn execute(
        &self,
        stream: &Stream,
        input: &DeviceBuffer<bf16>,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        self.kernels.project(stream, input, &self.weight, &self.scales, output)
    }

    #[must_use]
    pub const fn spec(&self) -> BlockFp8LinearSpec {
        self.spec
    }
}

impl Fp8ResidualLinearWeight {
    pub(in crate::backend) fn execute(
        &self,
        stream: &Stream,
        input: &DeviceBuffer<bf16>,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        self.kernels.project_residual(
            stream,
            input,
            &self.weight,
            &self.row_scales,
            &self.residual,
            &self.residual_scales,
            output,
        )
    }
}