candle_core/
lib.rs

1//! ML framework for Rust
2//!
3//! ```rust
4//! use candle_core::{Tensor, DType, Device};
5//! # use candle_core::Error;
6//! # fn main() -> Result<(), Error>{
7//!
8//! let a = Tensor::arange(0f32, 6f32, &Device::Cpu)?.reshape((2, 3))?;
9//! let b = Tensor::arange(0f32, 12f32, &Device::Cpu)?.reshape((3, 4))?;
10//! let c = a.matmul(&b)?;
11//!
12//! # Ok(())}
13//! ```
14//!
15//! ## Features
16//!
17//! - Simple syntax (looks and feels like PyTorch)
18//! - CPU and Cuda backends (and M1 support)
19//! - Enable serverless (CPU) small and fast deployments
20//! - Model training
21//! - Distributed computing (NCCL).
22//! - Models out of the box (Llama, Whisper, Falcon, ...)
23//!
24//! ## FAQ
25//!
26//! - Why Candle?
27//!
28//! Candle stems from the need to reduce binary size in order to *enable serverless*
29//! possible by making the whole engine smaller than PyTorch very large library volume
30//!
31//! And simply *removing Python* from production workloads.
32//! Python can really add overhead in more complex workflows and the [GIL](https://www.backblaze.com/blog/the-python-gil-past-present-and-future/) is a notorious source of headaches.
33//!
34//! Rust is cool, and a lot of the HF ecosystem already has Rust crates [safetensors](https://github.com/huggingface/safetensors) and [tokenizers](https://github.com/huggingface/tokenizers)
35//!
36//! ## Other Crates
37//!
38//! Candle consists of a number of crates. This crate holds core the common data structures but you may wish
39//! to look at the docs for the other crates which can be found here:
40//!
41//! - [candle-core](https://docs.rs/candle-core/). Core Datastructures and DataTypes.
42//! - [candle-nn](https://docs.rs/candle-nn/). Building blocks for Neural Nets.
43//! - [candle-datasets](https://docs.rs/candle-datasets/). Rust access to commonly used Datasets like MNIST.
44//! - [candle-examples](https://docs.rs/candle-examples/). Examples of Candle in Use.
45//! - [candle-onnx](https://docs.rs/candle-onnx/). Loading and using ONNX models.
46//! - [candle-pyo3](https://docs.rs/candle-pyo3/). Access to Candle from Python.
47//! - [candle-transformers](https://docs.rs/candle-transformers/). Candle implementation of many published transformer models.
48//!
49
50#[cfg(feature = "accelerate")]
51mod accelerate;
52pub mod backend;
53pub mod backprop;
54pub mod conv;
55mod convert;
56pub mod cpu;
57pub mod cpu_backend;
58#[cfg(feature = "cuda")]
59pub mod cuda_backend;
60mod custom_op;
61mod device;
62pub mod display;
63mod dtype;
64pub mod dummy_cuda_backend;
65pub mod dummy_dtype;
66mod dummy_metal_backend;
67pub mod error;
68mod indexer;
69pub mod layout;
70#[cfg(feature = "metal")]
71pub mod metal_backend;
72#[cfg(feature = "mkl")]
73mod mkl;
74pub mod npy;
75pub mod op;
76pub mod pickle;
77pub mod quantized;
78pub mod safetensors;
79pub mod scalar;
80pub mod shape;
81mod sort;
82mod storage;
83pub mod streaming;
84mod strided_index;
85mod tensor;
86mod tensor_cat;
87pub mod test_utils;
88pub mod utils;
89mod variable;
90
91#[cfg(feature = "cudnn")]
92pub use cuda_backend::cudnn;
93
94pub use cpu_backend::{CpuStorage, CpuStorageRef};
95#[cfg(feature = "ug")]
96pub use custom_op::UgIOp1;
97pub use custom_op::{CustomOp1, CustomOp2, CustomOp3, InplaceOp1, InplaceOp2, InplaceOp3};
98pub use device::{Device, DeviceLocation, NdArray};
99pub use dtype::{DType, DTypeParseError, FloatDType, IntDType, WithDType};
100pub use dummy_dtype::{F4, F6E2M3, F6E3M2, F8E8M0};
101pub use error::{Context, Error, Result};
102pub use indexer::{IndexOp, TensorIndexer};
103pub use layout::Layout;
104pub use shape::{Shape, D};
105pub use storage::Storage;
106pub use streaming::{StreamTensor, StreamingBinOp, StreamingModule};
107pub use strided_index::{StridedBlocks, StridedIndex};
108pub use tensor::{Tensor, TensorId};
109pub use variable::Var;
110
111#[cfg(feature = "cuda")]
112pub use cuda_backend as cuda;
113
114#[cfg(not(feature = "cuda"))]
115pub use dummy_cuda_backend as cuda;
116
117pub use cuda::{CudaDevice, CudaStorage};
118
119#[cfg(feature = "metal")]
120pub use metal_backend::{MetalDevice, MetalError, MetalStorage};
121
122#[cfg(not(feature = "metal"))]
123pub use dummy_metal_backend::{MetalDevice, MetalError, MetalStorage};
124
125#[cfg(feature = "mkl")]
126extern crate intel_mkl_src;
127
128#[cfg(feature = "accelerate")]
129extern crate accelerate_src;
130
131pub trait ToUsize2 {
132    fn to_usize2(self) -> (usize, usize);
133}
134
135impl ToUsize2 for usize {
136    fn to_usize2(self) -> (usize, usize) {
137        (self, self)
138    }
139}
140
141impl ToUsize2 for (usize, usize) {
142    fn to_usize2(self) -> (usize, usize) {
143        self
144    }
145}
146
147/// Defining a module with forward method using a single argument.
148pub trait Module {
149    fn forward(&self, xs: &Tensor) -> Result<Tensor>;
150}
151
152impl<T: Fn(&Tensor) -> Result<Tensor>> Module for T {
153    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
154        self(xs)
155    }
156}
157
158impl<M: Module> Module for Option<&M> {
159    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
160        match self {
161            None => Ok(xs.clone()),
162            Some(m) => m.forward(xs),
163        }
164    }
165}
166
167/// A single forward method using a single single tensor argument and a flag to
168/// separate the training and evaluation behaviors.
169pub trait ModuleT {
170    fn forward_t(&self, xs: &Tensor, train: bool) -> Result<Tensor>;
171}
172
173impl<M: Module> ModuleT for M {
174    fn forward_t(&self, xs: &Tensor, _train: bool) -> Result<Tensor> {
175        self.forward(xs)
176    }
177}
candle_core/lib.rs

candle_core/
lib.rs
