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

use models::weights::{
    Float8ActivationScale, Float8Format, Float8ParameterDType, Float8Quantization,
    Float8ScaleGranularity, Float8ScaleMode,
};

use crate::{
    Error, Result,
    kernels::{DirectFp8Activation, DirectFp8Format, DirectFp8Scale},
};

pub(super) fn execution_contract(
    name: &str,
    format: Float8Quantization,
    input: usize,
    output: usize,
) -> Result<(DirectFp8Format, DirectFp8Scale, bool, DirectFp8Activation)> {
    let value_format = match format.format {
        Float8Format::E4M3 => DirectFp8Format::E4M3,
        Float8Format::E5M2 => DirectFp8Format::E5M2,
    };
    let identity = format.scale_mode == Float8ScaleMode::None
        && format.scale_granularity == Float8ScaleGranularity::None
        && format.scale_dtype.is_none();
    if identity {
        return activation_contract(name, value_format, DirectFp8Scale::Tensor, false, format);
    }
    if format.scale_mode == Float8ScaleMode::None
        || format.scale_granularity == Float8ScaleGranularity::None
        || !matches!(
            format.scale_dtype,
            Some(Float8ParameterDType::BF16 | Float8ParameterDType::F32)
        )
    {
        return Err(unsupported(name, "has an incomplete scale contract"));
    }
    let inverse = format.scale_mode == Float8ScaleMode::InverseMultiplier;
    let scale = match format.scale_granularity {
        Float8ScaleGranularity::Tensor => DirectFp8Scale::Tensor,
        Float8ScaleGranularity::OutputChannel => DirectFp8Scale::OutputChannel,
        Float8ScaleGranularity::BlockGrid {
            output_groups,
            input_groups,
            output_block_size,
            input_block_size,
        } => block_scale(
            name,
            input,
            output,
            output_groups,
            input_groups,
            output_block_size,
            input_block_size,
        )?,
        Float8ScaleGranularity::None => unreachable!("validated above"),
    };
    activation_contract(name, value_format, scale, inverse, format)
}

#[allow(clippy::too_many_arguments)]
fn block_scale(
    name: &str,
    input: usize,
    output: usize,
    output_groups: usize,
    input_groups: usize,
    declared_output_block: Option<usize>,
    declared_input_block: Option<usize>,
) -> Result<DirectFp8Scale> {
    let (output_block_size, input_block_size) = match (declared_output_block, declared_input_block)
    {
        (Some(rows), Some(columns))
            if rows > 0
                && columns.is_multiple_of(4)
                && output_groups == output.div_ceil(rows)
                && input_groups == input.div_ceil(columns) =>
        {
            (rows, columns)
        },
        (None, None)
            if output_groups > 0
                && input_groups > 0
                && output.is_multiple_of(output_groups)
                && input.is_multiple_of(input_groups)
                && (input / input_groups).is_multiple_of(4) =>
        {
            (output / output_groups, input / input_groups)
        },
        _ => return Err(unsupported(name, "has invalid or ambiguous block-grid geometry")),
    };
    Ok(DirectFp8Scale::BlockGrid {
        output_groups,
        input_groups,
        output_block_size,
        input_block_size,
    })
}

fn activation_contract(
    name: &str,
    value_format: DirectFp8Format,
    scale: DirectFp8Scale,
    inverse: bool,
    format: Float8Quantization,
) -> Result<(DirectFp8Format, DirectFp8Scale, bool, DirectFp8Activation)> {
    let activation = match format.activation_scale {
        Float8ActivationScale::None => DirectFp8Activation::Bf16,
        Float8ActivationScale::DynamicToken if value_format == DirectFp8Format::E4M3 => {
            DirectFp8Activation::DynamicE4M3Token
        },
        Float8ActivationScale::DynamicToken => {
            return Err(unsupported(name, "does not support dynamic E5M2 activations"));
        },
        Float8ActivationScale::StaticTensor
            if value_format == DirectFp8Format::E4M3
                && matches!(
                    format.input_scale_dtype,
                    Some(Float8ParameterDType::BF16 | Float8ParameterDType::F32)
                )
                && format.input_scale_dtype == format.scale_dtype =>
        {
            DirectFp8Activation::StaticE4M3Tensor
        },
        Float8ActivationScale::StaticTensor => {
            return Err(unsupported(
                name,
                "requires a static E4M3 activation scale matching the weight scale dtype",
            ));
        },
    };
    Ok((value_format, scale, inverse, activation))
}

pub(super) fn unsupported(name: &str, requirement: &str) -> Error {
    Error::UnsupportedDecoderLayer(format!("direct FP8 projection {name} {requirement}"))
}