ruCCL 0.21.1

Ruda collective communication algorithms and orchestration.
Documentation 
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

use super::{Primitive, TensorBuffer, TensorDevice, TensorDeviceError, TensorElement};
use ruda_tensor::{Backend, DType, FloatDType, IntDType, Slice, TensorData, read_sync};
use std::ops::Range;

pub(super) fn checked_length<T: TensorElement>(length: usize) -> Result<(), TensorDeviceError> {
    let bytes = length.checked_mul(std::mem::size_of::<T>())
        .ok_or(TensorDeviceError::InvalidBuffer("collective buffer byte length overflow"))?;
    if bytes > isize::MAX as usize {
        return Err(TensorDeviceError::InvalidBuffer("collective buffer exceeds addressable size"));
    }
    Ok(())
}

pub(super) fn checked_range(
    total: usize,
    offset: usize,
    length: usize,
) -> Result<Range<usize>, TensorDeviceError> {
    let end = offset.checked_add(length)
        .ok_or(TensorDeviceError::InvalidBuffer("collective element range overflow"))?;
    if end > total {
        return Err(TensorDeviceError::InvalidBuffer("collective element range out of bounds"));
    }
    Ok(offset..end)
}

impl<B: Backend> Primitive<B> {
    pub(super) fn slice(self, range: Range<usize>) -> Self {
        let slice = Slice::from(range);
        match self {
            Self::Float(value) => Self::Float(B::float_slice(value, &[slice])),
            Self::Int(value) => Self::Int(B::int_slice(value, &[slice])),
        }
    }

    pub(super) fn assign(self, range: Range<usize>, value: Self) -> Self {
        let slice = Slice::from(range);
        match (self, value) {
            (Self::Float(tensor), Self::Float(value)) => {
                Self::Float(B::float_slice_assign(tensor, &[slice], value))
            }
            (Self::Int(tensor), Self::Int(value)) => {
                Self::Int(B::int_slice_assign(tensor, &[slice], value))
            }
            _ => unreachable!("typed collective buffer storage kind"),
        }
    }
}

impl<B: Backend> TensorDevice<B> {
    pub fn allocate<T: TensorElement>(
        &self,
        length: usize,
    ) -> Result<TensorBuffer<B, T>, TensorDeviceError> {
        self.validate_type::<T>()?;
        checked_length::<T>(length)?;
        let shape = [length].into();
        let value = match T::dtype() {
            DType::F32 => Primitive::Float(B::float_empty(shape, &self.device, FloatDType::F32)),
            DType::F16 => Primitive::Float(B::float_empty(shape, &self.device, FloatDType::F16)),
            DType::BF16 => Primitive::Float(B::float_empty(shape, &self.device, FloatDType::BF16)),
            DType::I32 => Primitive::Int(B::int_empty(shape, &self.device, IntDType::I32)),
            _ => unreachable!("sealed collective element type"),
        };
        Ok(self.wrap(value, length))
    }

    pub(super) fn from_values<T: TensorElement>(&self, values: &[T]) -> Primitive<B> {
        let data = TensorData::new(values.to_vec(), [values.len()]);
        if T::dtype() == DType::I32 {
            Primitive::Int(B::int_from_data(data, &self.device))
        } else {
            Primitive::Float(B::float_from_data(data, &self.device))
        }
    }

    pub fn write<T: TensorElement>(
        &self,
        buffer: &TensorBuffer<B, T>,
        values: &[T],
    ) -> Result<(), TensorDeviceError> {
        if values.len() != buffer.length {
            return Err(TensorDeviceError::InvalidBuffer("collective full write length mismatch"));
        }
        self.write_at(buffer, 0, values)
    }

    pub fn write_at<T: TensorElement>(
        &self,
        buffer: &TensorBuffer<B, T>,
        offset: usize,
        values: &[T],
    ) -> Result<(), TensorDeviceError> {
        self.validate_buffer(buffer)?;
        let range = checked_range(buffer.length, offset, values.len())?;
        if range.is_empty() {
            return Ok(());
        }
        let incoming = self.from_values(values);
        buffer.update(|current| {
            if range.start == 0 && range.end == buffer.length {
                incoming
            } else {
                current.assign(range, incoming)
            }
        })
    }

    pub fn download<T: TensorElement>(
        &self,
        buffer: &TensorBuffer<B, T>,
    ) -> Result<Vec<T>, TensorDeviceError> {
        self.download_at(buffer, 0, buffer.length)
    }

    pub fn download_at<T: TensorElement>(
        &self,
        buffer: &TensorBuffer<B, T>,
        offset: usize,
        length: usize,
    ) -> Result<Vec<T>, TensorDeviceError> {
        self.validate_buffer(buffer)?;
        let range = checked_range(buffer.length, offset, length)?;
        if range.is_empty() {
            return Ok(Vec::new());
        }
        let mut value = buffer.snapshot()?;
        if range.start != 0 || range.end != buffer.length {
            value = value.slice(range);
        }
        let data = match value {
            Primitive::Float(value) => read_sync(B::float_into_data(value))?,
            Primitive::Int(value) => read_sync(B::int_into_data(value))?,
        };
        data.into_vec::<T>().map_err(|error| TensorDeviceError::Data(format!("{error:?}")))
    }
}