burn_tch/
backend.rs

1use crate::IntoKind;
2
3use super::TchTensor;
4use super::element::TchElement;
5use burn_backend::backend::{Backend, DeviceId, DeviceOps, ExecutionError};
6use burn_backend::ops::IntTensorOps;
7
8#[derive(Clone, Copy, Debug, PartialEq, Eq)]
9/// The device struct when using the `tch` backend.
10///
11/// Note that you need to provide the device index when using Cuda.
12///
13/// # Example
14///
15/// ```no_run
16/// use burn_tch::LibTorchDevice;
17///
18/// let device_gpu_1 = LibTorchDevice::Cuda(0); // First GPU
19/// let device_gpu_2 = LibTorchDevice::Cuda(1); // Second GPU
20/// let device_cpu = LibTorchDevice::Cpu; // CPU
21/// let device_mps = LibTorchDevice::Mps; // Metal Performance Shaders
22/// let device_vulkan = LibTorchDevice::Vulkan; // Vulkan
23/// ```
24#[derive(Default)]
25pub enum LibTorchDevice {
26    /// CPU device.
27    #[default]
28    Cpu,
29
30    /// Cuda device with the given index. The index is the index of the Cuda device in the list of
31    /// all Cuda devices found on the system.
32    Cuda(usize),
33
34    /// Metal Performance Shaders device.
35    Mps,
36
37    /// Vulkan device.
38    Vulkan,
39}
40
41impl From<LibTorchDevice> for tch::Device {
42    #[allow(
43        unreachable_code,
44        reason = "CUDA branch always panics if the library is missing"
45    )]
46    fn from(device: LibTorchDevice) -> Self {
47        match device {
48            LibTorchDevice::Cpu => tch::Device::Cpu,
49            LibTorchDevice::Cuda(_num) => {
50                include!(concat!(env!("OUT_DIR"), "/tch_gpu_check.rs"));
51                tch::Device::Cuda(_num)
52            }
53            LibTorchDevice::Mps => tch::Device::Mps,
54            LibTorchDevice::Vulkan => tch::Device::Vulkan,
55        }
56    }
57}
58
59impl From<tch::Device> for LibTorchDevice {
60    fn from(device: tch::Device) -> Self {
61        match device {
62            tch::Device::Cpu => LibTorchDevice::Cpu,
63            tch::Device::Cuda(num) => LibTorchDevice::Cuda(num),
64            tch::Device::Mps => LibTorchDevice::Mps,
65            tch::Device::Vulkan => LibTorchDevice::Vulkan,
66        }
67    }
68}
69
70impl burn_backend::Device for LibTorchDevice {
71    fn from_id(device_id: DeviceId) -> Self {
72        match device_id.type_id {
73            0 => Self::Cuda(device_id.index_id as usize),
74            1 => Self::Mps,
75            2 => Self::Cpu,
76            3 => Self::Vulkan,
77            _ => LibTorchDevice::Cpu,
78        }
79    }
80
81    fn to_id(&self) -> DeviceId {
82        match self {
83            LibTorchDevice::Cuda(index) => DeviceId::new(0, *index as u32),
84            LibTorchDevice::Mps => DeviceId::new(1, 0),
85            LibTorchDevice::Cpu => DeviceId::new(2, 0),
86            LibTorchDevice::Vulkan => DeviceId::new(3, 0),
87        }
88    }
89
90    fn device_count(_type_id: u16) -> usize {
91        // TODO: Somehow find the info using the tch API.
92        1
93    }
94}
95
96impl DeviceOps for LibTorchDevice {}
97
98/// Tensor backend that uses `LibTorch` with the [tch] crate for executing tensor operations.
99///
100/// This backend is compatible with a wide range of hardwares ranging from CPUs to GPUs, but
101/// requires `LibTorch` to be installed correctly. The CPU version can be downloaded
102/// automatically and the CUDA version as well by setting the `TORCH_CUDA_VERSION` environment
103/// variable. For more complex configurations, check out the manual installation for
104/// [burn-tch](https://github.com/tracel-ai/burn/tree/main/crates/burn-tch).
105///
106/// Refer to the [tch] crate for more information.
107#[derive(Clone, Copy, Default, Debug)]
108pub struct LibTorch<E = f32> {
109    _e: E,
110}
111
112impl<E: TchElement> Backend for LibTorch<E> {
113    type Device = LibTorchDevice;
114
115    type FloatTensorPrimitive = TchTensor;
116    type FloatElem = E;
117
118    type IntTensorPrimitive = TchTensor;
119    type IntElem = i64;
120
121    type BoolTensorPrimitive = TchTensor;
122    type BoolElem = bool;
123
124    type QuantizedTensorPrimitive = TchTensor;
125
126    fn seed(_device: &Self::Device, seed: u64) {
127        tch::manual_seed(seed as i64);
128    }
129
130    fn ad_enabled() -> bool {
131        false
132    }
133
134    fn name(device: &Self::Device) -> String {
135        match device {
136            LibTorchDevice::Cpu => "libtorch<cpu>",
137            LibTorchDevice::Cuda(_) => "libtorch<cuda>",
138            LibTorchDevice::Mps => "libtorch<metal>",
139            LibTorchDevice::Vulkan => "libtorch<vulkan>",
140        }
141        .to_string()
142    }
143
144    fn sync(device: &Self::Device) -> Result<(), ExecutionError> {
145        match device {
146            LibTorchDevice::Cpu => (),
147            LibTorchDevice::Cuda(index) => {
148                tch::Cuda::synchronize(*index as i64);
149            }
150            _ => {
151                // When there is no explicit way to synchronize, we write and read one value to sync
152                burn_backend::read_sync(Self::int_into_data(Self::int_zeros(
153                    [1].into(),
154                    device,
155                    E::dtype().into(),
156                )))
157                .unwrap();
158            }
159        };
160
161        Ok(())
162    }
163
164    fn supports_dtype(_device: &Self::Device, dtype: burn_backend::DType) -> bool {
165        dtype.try_into_kind().is_ok()
166    }
167}
burn_tch/backend.rs

burn_tch/
backend.rs
