burn-dispatch 0.21.0-pre.4

Backend dispatch for the Burn framework
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
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161

use alloc::vec::Vec;
use core::mem::discriminant;

use crate::backends::*;

use burn_backend::{
    DeviceId,
    distributed::{
        CollectiveTensor, DistributedBackend, DistributedConfig, DistributedParams,
        ReduceOperation, TensorRef,
    },
    tensor::FloatTensor,
};

use crate::{Dispatch, DispatchDevice};

macro_rules! dispatch_devices_arms {
    (
        $device:expr,
        $devices:expr,
        |$inner_devices:ident| $body:expr;
        $([$Backend:ident, $cfg:meta]),*
    ) => {
        match $device {
            // Autodiff arm first
            #[cfg(feature = "autodiff")]
            $crate::DispatchDevice::Autodiff(inner) => {
                // Recursively dispatch on inner
                dispatch_devices_arms!(
                    @autodiff
                    &**inner,
                    $devices,
                    |$inner_devices| $body;
                    $([$Backend, $cfg]),*
                )
            },
            $(
                #[cfg($cfg)]
                $crate::DispatchDevice::$Backend(_) => {
                    assert!(
                        $devices
                            .iter()
                            .all(|d| discriminant(d) == discriminant($device)),
                        "All devices are expected to be of the same variant."
                    );
                    type B = $Backend<f32>;
                    let $inner_devices = $devices
                        .iter()
                        .map(|d| {
                            let DispatchDevice::$Backend(dev) = d else {
                                unreachable!()
                            };
                            dev.clone()
                        })
                        .collect::<Vec<_>>();
                    $body
                }
            )*
        }
    };
    (
        @autodiff
        $device:expr,
        $devices:expr,
        |$inner_devices:ident| $body:expr;
        $([$Backend:ident, $cfg:meta]),*
    ) => {
        match $device {
            $(
                #[cfg($cfg)]
                $crate::DispatchDevice::$Backend(_) => {
                    assert!(
                        $devices
                            .iter()
                            .all(|d| discriminant(d) == discriminant($device)),
                        "All devices are expected to be of the same variant."
                    );
                    type B = Autodiff<$Backend<f32>>;
                    let $inner_devices = $devices
                        .iter()
                        .map(|d| {
                            let DispatchDevice::$Backend(dev) = d else {
                                unreachable!()
                            };
                            dev.clone()
                        })
                        .collect::<Vec<_>>();
                    $body
                }
            )*
            $crate::DispatchDevice::Autodiff(_) => panic!("Autodiff should not wrap an autodiff device.")
        }
    };
}

/// Dispatches an operation body based on the provided devices.
macro_rules! dispatch_devices {
    ($device:expr, $devices:expr, |$inner_devices:ident| $body:expr) => {
        backend_list!(dispatch_devices_arms, $device, $devices, |$inner_devices| {
            $body
        })
    };
}

impl DistributedBackend for Dispatch {
    fn start_communication_server(devices: &[DispatchDevice], config: DistributedConfig) {
        if !devices.is_empty() {
            let first = &devices[0];
            dispatch_devices!(first, devices, |inner_devices| {
                B::start_communication_server(&inner_devices, config)
            });
        }
    }

    fn close_communication_server(device: &DispatchDevice) {
        dispatch_device!(device, |device| B::close_communication_server(device))
    }

    fn register_sync_parameters(
        device: &DispatchDevice,
        sharded_param_ids: Vec<DistributedParams>,
    ) {
        dispatch_device!(device, |device| B::register_sync_parameters(
            device,
            sharded_param_ids,
        ))
    }

    fn submit_sync_collective(device: &DispatchDevice) {
        dispatch_device!(device, |device| B::submit_sync_collective(device))
    }

    fn submit_gradient_sync(_tensor: TensorRef<Self>, _distributed_params: DistributedParams) {
        unimplemented!()
    }

    fn all_reduce(
        tensor: FloatTensor<Self>,
        op: ReduceOperation,
        device_ids: Vec<DeviceId>,
    ) -> CollectiveTensor<Self> {
        // Safety: we call `assume_resolved` only to wrap it in a new `CollectiveTensor`.
        let tensor = unary_float!(tensor, float, |tensor| {
            let collective_tensor = B::all_reduce(tensor, op, device_ids);
            unsafe { collective_tensor.assume_resolved() }
        } => Float);
        CollectiveTensor::new(tensor)
    }

    fn sync_collective(device: &DispatchDevice) {
        dispatch_device!(device, |device| B::sync_collective(device))
    }

    unsafe fn comm_device(_tensor: &TensorRef<Self>) -> DispatchDevice {
        unimplemented!()
    }

    unsafe fn float_from_ref(_tensor: &TensorRef<Self>) -> FloatTensor<Self> {
        unimplemented!()
    }
}