libmir-cuda 0.3.0

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, cuda_export};

use super::{ClampedRoutedKernels, linear_launch, narrow};
use crate::Result;

cuda_export!(pub(super) QkvKernel = "libmir_cuda_clamped_routed_qkv_bf16"(
    packed: &DeviceBuffer<bf16>, q_bias: &DeviceBuffer<bf16>,
    k_bias: &DeviceBuffer<bf16>, v_bias: &DeviceBuffer<bf16>,
    query: &mut DeviceBuffer<bf16>, key: &mut DeviceBuffer<bf16>,
    value: &mut DeviceBuffer<bf16>, tokens: u32, query_heads: u32,
    kv_heads: u32, head_dim: u32, rope_sines: &DeviceBuffer<f32>,
    rope_cosines: &DeviceBuffer<f32>, concentration: f32,
));
cuda_export!(pub(super) QkvSplitKernel = "libmir_cuda_clamped_routed_qkv_split_bf16"(
    q_input: &DeviceBuffer<bf16>, k_input: &DeviceBuffer<bf16>,
    v_input: &DeviceBuffer<bf16>, q_bias: &DeviceBuffer<bf16>,
    k_bias: &DeviceBuffer<bf16>, v_bias: &DeviceBuffer<bf16>,
    query: &mut DeviceBuffer<bf16>, key: &mut DeviceBuffer<bf16>,
    value: &mut DeviceBuffer<bf16>, tokens: u32, query_heads: u32,
    kv_heads: u32, head_dim: u32, rope_sines: &DeviceBuffer<f32>,
    rope_cosines: &DeviceBuffer<f32>, concentration: f32,
));
cuda_export!(pub(super) RopeKernel = "libmir_cuda_clamped_routed_rope_angles"(
    positions: &DeviceBuffer<u32>, inverse_frequencies: &DeviceBuffer<f32>,
    sines: &mut DeviceBuffer<f32>, cosines: &mut DeviceBuffer<f32>,
    tokens: u32, half_head_dim: u32,
));

impl ClampedRoutedKernels {
    #[allow(clippy::too_many_arguments)]
    pub(crate) fn qkv_native(
        &self,
        stream: &Stream,
        packed: &DeviceBuffer<bf16>,
        biases: &[crate::CudaTensor; 3],
        query: &mut DeviceBuffer<bf16>,
        key: &mut DeviceBuffer<bf16>,
        value: &mut DeviceBuffer<bf16>,
        rope_sines: &DeviceBuffer<f32>,
        rope_cosines: &DeviceBuffer<f32>,
        concentration: f32,
    ) -> Result<()> {
        let biases = bias_buffers(biases)?;
        let elements = qkv_elements(self.spec);
        Ok(self.qkv.launch(
            stream,
            linear_launch(elements)?,
            (
                packed,
                biases[0],
                biases[1],
                biases[2],
                query,
                key,
                value,
                narrow(self.spec.tokens)?,
                narrow(self.spec.query_heads)?,
                narrow(self.spec.kv_heads)?,
                narrow(self.spec.head_dim)?,
                rope_sines,
                rope_cosines,
                concentration,
            ),
        )?)
    }

    #[allow(clippy::too_many_arguments)]
    pub(crate) fn qkv_mlx(
        &self,
        stream: &Stream,
        inputs: [&DeviceBuffer<bf16>; 3],
        biases: &[crate::CudaTensor; 3],
        query: &mut DeviceBuffer<bf16>,
        key: &mut DeviceBuffer<bf16>,
        value: &mut DeviceBuffer<bf16>,
        rope_sines: &DeviceBuffer<f32>,
        rope_cosines: &DeviceBuffer<f32>,
        concentration: f32,
    ) -> Result<()> {
        let biases = bias_buffers(biases)?;
        let elements = qkv_elements(self.spec);
        Ok(self.qkv_split.launch(
            stream,
            linear_launch(elements)?,
            (
                inputs[0],
                inputs[1],
                inputs[2],
                biases[0],
                biases[1],
                biases[2],
                query,
                key,
                value,
                narrow(self.spec.tokens)?,
                narrow(self.spec.query_heads)?,
                narrow(self.spec.kv_heads)?,
                narrow(self.spec.head_dim)?,
                rope_sines,
                rope_cosines,
                concentration,
            ),
        )?)
    }

    pub(crate) fn prepare_rope(
        &self,
        stream: &Stream,
        positions: &DeviceBuffer<u32>,
        inverse: &DeviceBuffer<f32>,
        sines: &mut DeviceBuffer<f32>,
        cosines: &mut DeviceBuffer<f32>,
    ) -> Result<()> {
        let half = self.spec.head_dim / 2;
        let elements = self.spec.tokens * half;
        Ok(self.rope.launch(
            stream,
            linear_launch(elements)?,
            (positions, inverse, sines, cosines, narrow(self.spec.tokens)?, narrow(half)?),
        )?)
    }
}

const fn qkv_elements(spec: super::ClampedRoutedSpec) -> usize {
    let rotary = (spec.query_heads + spec.kv_heads) * (spec.head_dim / 2);
    let values = spec.kv_heads * spec.head_dim;
    spec.tokens * (rotary + values)
}

fn bias_buffers(biases: &[crate::CudaTensor; 3]) -> Result<[&DeviceBuffer<bf16>; 3]> {
    Ok([bf16_tensor(&biases[0])?, bf16_tensor(&biases[1])?, bf16_tensor(&biases[2])?])
}

fn bf16_tensor(tensor: &crate::CudaTensor) -> Result<&DeviceBuffer<bf16>> {
    tensor.as_bf16().ok_or_else(|| crate::Error::DTypeMismatch {
        name: tensor.name().into(),
        expected: "BF16",
    })
}