burn_tensor/tensor/backend/
base.rs

1use alloc::string::String;
2
3use crate::TensorMetadata;
4use crate::tensor::Element;
5use crate::{ops::*, quantization::QTensorPrimitive};
6
7use super::DeviceOps;
8
9/// This trait defines all types and functions needed for a backend to be used with burn.
10///
11/// ## Design
12///
13/// This trait aims to be as unopinionated as possible and allows implementations to define
14/// their own types and patterns. Therefore, there are few pre-defined abstractions baked
15/// into this trait.
16///
17/// Backends must define their own tensor types for each data type: `float`, `int`, and `bool`.
18/// Since we minimize assumptions, we chose to separate these types, as they are used in
19/// different contexts. However, some backends may have a generic tensor type that is used
20/// for all data types.
21///
22/// ### Eager Mode
23///
24/// Because burn supports dynamic graphs, the backend trait is designed around kernel
25/// implementations that can be called without any mutable context or graph. This may not be
26/// ideal for backends that want to configure their computational graphs and execute them
27/// multiple times.
28///
29/// To implement this kind of backend, channels could be used to communicate with a backend
30/// server thread to build the computation graphs and re-execute the ones that are repeated,
31/// with some form of cache. Once that pattern has matured, a graph mode backend trait could
32/// be extracted from it, allowing other backends of the same kind to be quickly integrated
33/// with burn. This pattern could also be used to create an operation fusion trait, which
34/// allows backends to define what kind of graph structures can be fused into one operation.
35///
36/// ### Multi-Threaded
37///
38/// Backend tensor types are all `Clone` + `Send`, which allows them to be safely
39/// sent between threads. It is recommended to wrap tensors with [Arc](alloc::sync::Arc),
40/// which avoids copying the tensor's buffer. Note that it is still possible to mutate and
41/// reuse tensors' buffer without locking; see the next section on the Mutable API.
42///
43/// ### Mutable API
44///
45/// There is no mutable or inplace operation API to implement, but that does not mean that
46/// backends cannot support them. Using [try_unwrap](alloc::sync::Arc::try_unwrap) and
47/// [get_mut](alloc::sync::Arc::get_mut) allows backends to have access to an owned or mutable
48/// reference to their tensor buffer data structure if the tensor is not shared. In that case,
49/// backends can dispatch to their owned inplace operations for better performance.
50///
51/// ## Documentation
52///
53/// Most of the documentation for each function can be found on the user API [tensor struct](crate::Tensor).
54/// For modules, public functions are often created, which can be used by `burn-core` modules.
55pub trait Backend:
56    FloatTensorOps<Self>
57    + BoolTensorOps<Self>
58    + IntTensorOps<Self>
59    + ModuleOps<Self>
60    + ActivationOps<Self>
61    + QTensorOps<Self>
62    + TransactionOps<Self>
63    + Clone
64    + Default
65    + Sized
66    + Send
67    + Sync
68    + core::fmt::Debug
69    + 'static
70{
71    /// Device type.
72    type Device: DeviceOps;
73
74    /// Tensor primitive to be used for all float operations.
75    type FloatTensorPrimitive: TensorMetadata + 'static;
76    /// Default float element type.
77    type FloatElem: Element;
78
79    /// Tensor primitive to be used for all int operations.
80    type IntTensorPrimitive: TensorMetadata + 'static;
81    /// Int element type.
82    type IntElem: Element;
83
84    /// Tensor primitive to be used for all bool operations.
85    type BoolTensorPrimitive: TensorMetadata + 'static;
86    /// Tensor primitive to be used for all bool operations.
87    type BoolElem: Element;
88
89    /// Tensor primitive to be used for all quantized operations.
90    type QuantizedTensorPrimitive: TensorMetadata + QTensorPrimitive + 'static;
91    /// Quantized tensor encoding type.
92    type QuantizedEncoding: Element;
93
94    /// If autodiff is enabled.
95    fn ad_enabled() -> bool {
96        false
97    }
98
99    /// Name of the backend.
100    fn name(device: &Self::Device) -> String;
101
102    /// Seed the backend.
103    fn seed(seed: u64);
104
105    /// Sync the backend, ensure that all computation are finished.
106    fn sync(_device: &Self::Device) {}
107}
108
109/// Trait that allows a backend to support autodiff.
110pub trait AutodiffBackend: Backend {
111    /// The inner backend type.
112    type InnerBackend: Backend<Device = Self::Device, FloatElem = Self::FloatElem, IntElem = Self::IntElem>;
113
114    /// Gradients type.
115    type Gradients: Send;
116
117    /// Backward pass.
118    ///
119    /// # Arguments
120    ///
121    /// * `tensor` - The tensor is the last node of computational graph where the gradients are computed.
122    ///
123    /// # Returns
124    ///
125    /// The gradients.
126    fn backward(tensor: FloatTensor<Self>) -> Self::Gradients;
127
128    /// Returns the gradients of a tensor.
129    ///
130    /// # Arguments
131    ///
132    /// * `tensor` - The tensor to extract the gradients from.
133    ///
134    /// # Returns
135    ///
136    /// An optional tensor containing the gradient.
137    fn grad(
138        tensor: &FloatTensor<Self>,
139        grads: &Self::Gradients,
140    ) -> Option<FloatTensor<Self::InnerBackend>>;
141
142    /// Pops the gradients of a tensor and returns them.
143    ///
144    /// # Arguments
145    ///
146    /// * `tensor` - The tensor to pop the gradients from.
147    /// * `grads` - The gradients.
148    ///
149    /// # Returns
150    ///
151    /// An optional tensor containing the given gradients.
152    fn grad_remove(
153        tensor: &FloatTensor<Self>,
154        grads: &mut Self::Gradients,
155    ) -> Option<FloatTensor<Self::InnerBackend>>;
156
157    /// Replace the gradients of a tensor with the one provided.
158    ///
159    /// If no gradient existed for the provided tensor, register it.
160    ///
161    /// # Arguments
162    ///
163    /// * `tensor` - The tensor to pop the gradients from.
164    /// * `grads` - The gradients.
165    /// * `grad` - The updated grad tensor.
166    fn grad_replace(
167        tensor: &FloatTensor<Self>,
168        grads: &mut Self::Gradients,
169        grad: FloatTensor<Self::InnerBackend>,
170    );
171
172    /// Returns the tensor with inner backend type.
173    ///
174    /// # Arguments
175    ///
176    /// * `tensor` - The tensor to get the inner backend tensor for.
177    ///
178    /// # Returns
179    ///
180    /// The inner backend tensor.
181    fn inner(tensor: FloatTensor<Self>) -> FloatTensor<Self::InnerBackend>;
182
183    /// Returns the tensor with inner backend type.
184    ///
185    /// # Arguments
186    ///
187    /// * `tensor` - The tensor to get the inner backend tensor for.
188    ///
189    /// # Returns
190    ///
191    /// The inner backend tensor.
192    fn int_inner(tensor: IntTensor<Self>) -> IntTensor<Self::InnerBackend>;
193
194    /// Returns the tensor with inner backend type.
195    ///
196    /// # Arguments
197    ///
198    /// * `tensor` - The tensor to get the inner backend tensor for.
199    ///
200    /// # Returns
201    ///
202    /// The inner backend tensor.
203    fn bool_inner(tensor: BoolTensor<Self>) -> BoolTensor<Self::InnerBackend>;
204
205    /// Returns the tensor with inner backend type.
206    ///
207    /// # Arguments
208    ///
209    /// * `tensor` - The tensor to get the inner backend tensor for.
210    ///
211    /// # Returns
212    ///
213    /// The inner backend tensor.
214    fn q_inner(tensor: QuantizedTensor<Self>) -> QuantizedTensor<Self::InnerBackend>;
215
216    /// Converts the inner backend tensor to the autodiff backend tensor.
217    ///
218    /// # Arguments
219    ///
220    /// * `tensor` - The inner backend tensor to convert.
221    ///
222    ///
223    /// # Returns
224    ///
225    /// The autodiff backend tensor.
226    fn from_inner(tensor: FloatTensor<Self::InnerBackend>) -> FloatTensor<Self>;
227
228    /// Converts the inner backend tensor to the autodiff backend tensor.
229    ///
230    /// # Arguments
231    ///
232    /// * `tensor` - The inner backend tensor to convert.
233    ///
234    ///
235    /// # Returns
236    ///
237    /// The autodiff backend tensor.
238    fn int_from_inner(tensor: IntTensor<Self::InnerBackend>) -> IntTensor<Self>;
239
240    /// Converts the inner backend tensor to the autodiff backend tensor.
241    ///
242    /// # Arguments
243    ///
244    /// * `tensor` - The inner backend tensor to convert.
245    ///
246    ///
247    /// # Returns
248    ///
249    /// The autodiff backend tensor.
250    fn bool_from_inner(tensor: BoolTensor<Self::InnerBackend>) -> BoolTensor<Self>;
251
252    /// Converts the inner backend tensor to the autodiff backend tensor.
253    ///
254    /// # Arguments
255    ///
256    /// * `tensor` - The inner backend tensor to convert.
257    ///
258    ///
259    /// # Returns
260    ///
261    /// The autodiff backend tensor.
262    fn q_from_inner(tensor: QuantizedTensor<Self::InnerBackend>) -> QuantizedTensor<Self>;
263}