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

use super::geometry::product;
use crate::{Error, Result};

mod embedding;
mod gathered;
mod portable;
pub use embedding::{MxFp4Embedding, MxFp4EmbeddingOperands, MxFp4EmbeddingSpec};
pub use gathered::{MxFp4GatheredLinear, MxFp4GatheredOperands};
pub use portable::MxFp4Linear;

#[derive(Clone, Copy, Debug, Eq, Hash, PartialEq)]
/// Fixed BF16-by-OCP-MXFP4 projection geometry.
pub struct MxFp4Spec {
    pub tokens: usize,
    pub input_features: usize,
    pub output_features: usize,
}

#[derive(Clone, Copy, Debug, Eq, Hash, PartialEq)]
/// Fixed gathered BF16-by-OCP-MXFP4 matrix-bank geometry.
pub struct MxFp4GatheredSpec {
    pub input_rows: usize,
    pub selections_per_input: usize,
    pub assignments: usize,
    pub matrices: usize,
    pub input_features: usize,
    pub output_features: usize,
}

impl MxFp4GatheredSpec {
    pub fn new(
        assignments: usize,
        matrices: usize,
        input_features: usize,
        output_features: usize,
    ) -> Result<Self> {
        Self::new_routed(assignments, 1, matrices, input_features, output_features)
    }

    pub fn new_routed(
        input_rows: usize,
        selections_per_input: usize,
        matrices: usize,
        input_features: usize,
        output_features: usize,
    ) -> Result<Self> {
        let assignments = product(input_rows, selections_per_input)?;
        let projection = MxFp4Spec::new(1, input_features, output_features)?;
        if matrices == 0 || selections_per_input == 0 {
            return Err(Error::InvalidDecoderKernel("invalid gathered MXFP4 matrix count"));
        }
        let spec = Self {
            input_rows,
            selections_per_input,
            assignments,
            matrices,
            input_features,
            output_features,
        };
        let _ = product(matrices, projection.weight_elements()?)?;
        let _ = product(matrices, projection.scale_elements()?)?;
        let _ = product(matrices, output_features)?;
        Ok(spec)
    }

    pub fn projection(self) -> Result<MxFp4Spec> {
        MxFp4Spec::new(self.input_rows, self.input_features, self.output_features)
    }

    pub fn output_elements(self) -> Result<usize> {
        product(self.assignments, self.output_features)
    }

    pub fn weight_elements(self) -> Result<usize> {
        product(self.matrices, self.projection()?.weight_elements()?)
    }

    pub fn scale_elements(self) -> Result<usize> {
        product(self.matrices, self.projection()?.scale_elements()?)
    }

    pub fn bias_elements(self) -> Result<usize> {
        product(self.matrices, self.output_features)
    }
}

impl MxFp4Spec {
    pub fn new(tokens: usize, input_features: usize, output_features: usize) -> Result<Self> {
        if tokens == 0
            || output_features == 0
            || input_features == 0
            || !input_features.is_multiple_of(32)
        {
            return Err(Error::InvalidDecoderKernel("invalid MXFP4 linear geometry"));
        }
        let spec = Self { tokens, input_features, output_features };
        let _ = spec.input_elements()?;
        let _ = spec.weight_elements()?;
        let _ = spec.scale_elements()?;
        let _ = spec.output_elements()?;
        Ok(spec)
    }

    pub fn input_elements(self) -> Result<usize> {
        product(self.tokens, self.input_features)
    }

    pub fn weight_elements(self) -> Result<usize> {
        product(self.output_features, self.input_features / 2)
    }

    pub fn scale_elements(self) -> Result<usize> {
        product(self.output_features, self.input_features / 32)
    }

    pub fn output_elements(self) -> Result<usize> {
        product(self.tokens, self.output_features)
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn requires_complete_mxfp4_blocks() -> Result<()> {
        assert!(MxFp4Spec::new(2, 64, 3).is_ok());
        assert!(MxFp4Spec::new(2, 48, 3).is_err());
        assert!(MxFp4GatheredSpec::new(2, 4, 64, 3).is_ok());
        assert!(MxFp4GatheredSpec::new(2, 0, 64, 3).is_err());
        assert_eq!(MxFp4GatheredSpec::new_routed(2, 3, 4, 64, 3)?.assignments, 6);
        Ok(())
    }
}