burn-backend 0.21.0

Core backend interfaces and data structures for executing tensor operations in Burn.
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

//! Types and helpers for inter-device operations.

pub(crate) mod api;
pub(crate) mod client;
mod ops;
pub(crate) mod server;

pub use api::*;
pub use ops::*;

// TODO: should be removed I think!
// instead of having a new `DistributedParamId`, we can keep `ParamId` only since it maps to the param id.
// (in `ModuleSharder` the map_float impl simply calls `DistributedParamId::from(param_id)`).
// and we don't need to query the tensor to get the param id. Simply, check *if* `tensor.is_distributed()`
// and then afterwards , if t was distributed we can recover the state here since we have the param id.
use serde::{Deserialize, Serialize};

use crate::tensor::FloatTensor;

/// A unique identifier for a parameter distributed across multiple devices.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize)]
pub struct DistributedParamId(u64);

impl From<u64> for DistributedParamId {
    fn from(value: u64) -> Self {
        Self(value)
    }
}

/// Parameters for a tensor that is sharded across multiple devices.
#[derive(Debug, Clone)]
pub struct DistributedParams {
    /// The tensor's [DistributedParamId].
    pub param_id: DistributedParamId,
}

/// The different ways to execute the reduce operation.
#[derive(Debug, PartialEq, Clone, Copy, Serialize, Deserialize)]
pub enum ReduceOperation {
    /// The sum of the values.
    Sum,
    /// The mean of the values.
    Mean,
}

/// Parameter struct for setting up and getting parameters for distributed operations.
#[derive(Clone)]
pub struct DistributedConfig {
    /// How to execute the all_reduce operation.
    pub all_reduce_op: ReduceOperation,
}

/// A tensor handle used for a collective operation, that is not yet valid for use.
/// We must ensure collective operations are completed before accessing the underlying data.
#[derive(new, Clone)]
pub struct CollectiveTensor<B: DistributedBackend> {
    handle: FloatTensor<B>,
}

impl<B: DistributedBackend> CollectiveTensor<B> {
    /// Synchronizes the collective operation and returns a valid tensor handle.
    pub fn resolve(self) -> FloatTensor<B> {
        B::sync_collective(&B::float_device(&self.handle));
        self.handle
    }

    /// Returns the tensor handle without synchronizing.
    ///
    /// # Safety
    ///
    /// The caller must ensure that `sync_collective()` is called before
    /// the returned handle is used in any computation.
    pub unsafe fn assume_resolved(self) -> FloatTensor<B> {
        self.handle
    }
}