libmir-metal 0.3.0

Metal 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
140
141
142

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

use super::{invalid, require};
use crate::engine::{Array, Dtype, ModelTensors, Result, Stream};

#[derive(Debug)]
pub(super) enum DirectFloat8Activation {
    Bfloat16,
    Dynamic,
    Static(Array),
}

pub(super) struct EncodedActivation<'a> {
    input: Array,
    scale: Scale<'a>,
}

enum Scale<'a> {
    Dynamic(Array),
    Static(&'a Array),
}

pub(super) fn valid(format: Float8Quantization, has_input_scale: bool) -> bool {
    match (format.format, format.activation_scale) {
        (Float8Format::E4M3, Float8ActivationScale::None | Float8ActivationScale::DynamicToken) => {
            !has_input_scale
        },
        (Float8Format::E5M2, Float8ActivationScale::None) => !has_input_scale,
        (Float8Format::E4M3, Float8ActivationScale::StaticTensor) => {
            has_input_scale
                && matches!(
                    format.input_scale_dtype,
                    Some(Float8ParameterDType::BF16 | Float8ParameterDType::F32)
                )
                && format.input_scale_dtype == format.scale_dtype
        },
        _ => false,
    }
}

impl DirectFloat8Activation {
    pub(super) fn prepare(
        tensors: &ModelTensors,
        format: Float8Quantization,
        input_scale: Option<&str>,
        binding: &TensorBinding,
        stream: &Stream,
    ) -> Result<Self> {
        match format.activation_scale {
            Float8ActivationScale::DynamicToken => Ok(Self::Dynamic),
            Float8ActivationScale::StaticTensor => {
                let name = input_scale.ok_or_else(|| invalid(binding, "input scale is missing"))?;
                let scale = tensors.get(name)?;
                let dtype = match format.input_scale_dtype {
                    Some(Float8ParameterDType::BF16) => Dtype::Bfloat16,
                    Some(Float8ParameterDType::F32) => Dtype::Float32,
                    None => return Err(invalid(binding, "input scale dtype is missing")),
                };
                if !scale.shape()?.is_empty() && scale.shape()? != [1] {
                    return Err(invalid(binding, "input scale is not scalar"));
                }
                require(&scale, dtype, &[], binding, "input scale")
                    .or_else(|_| require(&scale, dtype, &[1], binding, "input scale"))?;
                let scale = scale.astype(Dtype::Float32, stream)?.reshape(&[1, 1], stream)?;
                scale.async_eval()?;
                Ok(Self::Static(scale))
            },
            Float8ActivationScale::None => Ok(Self::Bfloat16),
        }
    }

    pub(super) fn encode<'a>(
        &'a self,
        input: &Array,
        stream: &Stream,
    ) -> Result<EncodedActivation<'a>> {
        match self {
            Self::Bfloat16 => Err(crate::engine::Error::InvalidQuantization(
                "BF16 direct FP8 activation does not require encoding".into(),
            )),
            Self::Dynamic => {
                let (input, scale) = dynamic(input, stream)?;
                Ok(EncodedActivation { input, scale: Scale::Dynamic(scale) })
            },
            Self::Static(scale) => Ok(EncodedActivation {
                input: encode(input, scale, stream)?,
                scale: Scale::Static(scale),
            }),
        }
    }

    pub(super) const fn is_bfloat16(&self) -> bool {
        matches!(self, Self::Bfloat16)
    }
}

impl EncodedActivation<'_> {
    pub(super) const fn input(&self) -> &Array {
        &self.input
    }

    pub(super) const fn scale(&self) -> &Array {
        match &self.scale {
            Scale::Dynamic(scale) => scale,
            Scale::Static(scale) => scale,
        }
    }

    pub(super) const fn scale_stride(&self) -> usize {
        match self.scale {
            Scale::Dynamic(_) => 1,
            Scale::Static(_) => 0,
        }
    }
}

fn dynamic(input: &Array, stream: &Stream) -> Result<(Array, Array)> {
    const FP8_MAX: f32 = 448.0;
    const MINIMUM_SCALE: f32 = 1.0 / (FP8_MAX * 512.0);

    let graph = stream.native().graph();
    let input_f32 = graph.astype(input.native(), mirtal::DType::Float32)?;
    let absolute = graph.maximum(&input_f32, &graph.negative(&input_f32)?)?;
    let maximum = graph.reduce_max(&absolute, -1, true)?;
    let scale = graph.multiply_scalar(&maximum, 1.0 / FP8_MAX)?;
    let minimum = graph.full(&mirtal::Shape::new([])?, MINIMUM_SCALE, mirtal::DType::Float32)?;
    let scale = graph.maximum(&scale, &minimum)?;
    Ok((encode_native(&input_f32, &scale, stream)?, Array::from_native(scale)?))
}

fn encode(input: &Array, scale: &Array, stream: &Stream) -> Result<Array> {
    let input = stream.native().graph().astype(input.native(), mirtal::DType::Float32)?;
    encode_native(&input, scale.native(), stream)
}

fn encode_native(input: &mirtal::Array, scale: &mirtal::Array, stream: &Stream) -> Result<Array> {
    let graph = stream.native().graph();
    let normalized = graph.divide(input, scale)?;
    Array::from_native(graph.to_fp8(&normalized)?)
}