Trait Backend 

pub trait Backend:
    Debug
    + Send
    + Sync {
    // Required methods
    fn zeros(&self, shape: &Shape) -> Result<Storage>;
    fn ones(&self, shape: &Shape) -> Result<Storage>;
    fn from_slice(&self, data: &[f32], shape: &Shape) -> Result<Storage>;
    fn add(&self, lhs: &Storage, rhs: &Storage) -> Result<Storage>;
    fn sub(&self, lhs: &Storage, rhs: &Storage) -> Result<Storage>;
    fn mul(&self, lhs: &Storage, rhs: &Storage) -> Result<Storage>;
    fn div(&self, lhs: &Storage, rhs: &Storage) -> Result<Storage>;
    fn sum(
        &self,
        storage: &Storage,
        shape: &Shape,
        axis: Option<usize>,
    ) -> Result<Storage>;
    fn mean(
        &self,
        storage: &Storage,
        shape: &Shape,
        axis: Option<usize>,
    ) -> Result<Storage>;
    fn transpose(&self, storage: &Storage, shape: &Shape) -> Result<Storage>;
    fn to_vec_f32(&self, storage: &Storage) -> Result<Vec<f32>>;
    fn matmul(
        &self,
        lhs: &Storage,
        rhs: &Storage,
        lhs_shape: &Shape,
        rhs_shape: &Shape,
    ) -> Result<Storage>;
    fn bmm(
        &self,
        lhs: &Storage,
        rhs: &Storage,
        lhs_shape: &Shape,
        rhs_shape: &Shape,
    ) -> Result<Storage>;
    fn exp(&self, storage: &Storage) -> Result<Storage>;
    fn log(&self, storage: &Storage) -> Result<Storage>;
    fn sqrt(&self, storage: &Storage) -> Result<Storage>;
    fn pow(&self, storage: &Storage, power: f32) -> Result<Storage>;
    fn sin(&self, storage: &Storage) -> Result<Storage>;
    fn cos(&self, storage: &Storage) -> Result<Storage>;
    fn relu(&self, storage: &Storage) -> Result<Storage>;
    fn sigmoid(&self, storage: &Storage) -> Result<Storage>;
    fn tanh(&self, storage: &Storage) -> Result<Storage>;

    // Provided method
    fn is_available(&self) -> bool { ... }
}

The backend trait for tensor operations The main backend trait that all compute backends must implement.

This trait defines the interface for tensor operations across different hardware backends. Each backend is responsible for:

• Creating tensors (zeros, ones, from data)
• Performing arithmetic operations
• Performing reductions and transformations
• Converting between storage formats

Implementation Notes

Backends should be thread-safe (Send + Sync) and implement Debug for diagnostic purposes. The is_available method allows backends to report whether they can be used on the current system.

Required Methods

fn zeros(&self, shape: &Shape) -> Result<Storage>

Creates a tensor filled with zeros.

fn ones(&self, shape: &Shape) -> Result<Storage>

Creates a tensor filled with ones.

fn from_slice(&self, data: &[f32], shape: &Shape) -> Result<Storage>

Creates a tensor from a slice of f32 values.

fn add(&self, lhs: &Storage, rhs: &Storage) -> Result<Storage>

Performs element-wise addition.

fn sub(&self, lhs: &Storage, rhs: &Storage) -> Result<Storage>

Performs element-wise subtraction.

fn mul(&self, lhs: &Storage, rhs: &Storage) -> Result<Storage>

Performs element-wise multiplication.

fn div(&self, lhs: &Storage, rhs: &Storage) -> Result<Storage>

Performs element-wise division.

fn sum( &self, storage: &Storage, shape: &Shape, axis: Option<usize>, ) -> Result<Storage>

Computes the sum of elements along an optional axis.

fn mean( &self, storage: &Storage, shape: &Shape, axis: Option<usize>, ) -> Result<Storage>

Computes the mean of elements along an optional axis.

fn transpose(&self, storage: &Storage, shape: &Shape) -> Result<Storage>

Transposes a 2D tensor.

fn to_vec_f32(&self, storage: &Storage) -> Result<Vec<f32>>

Converts the storage to a vector of f32 values.

fn matmul( &self, lhs: &Storage, rhs: &Storage, lhs_shape: &Shape, rhs_shape: &Shape, ) -> Result<Storage>

Matrix multiplication for 2D tensors.

fn bmm( &self, lhs: &Storage, rhs: &Storage, lhs_shape: &Shape, rhs_shape: &Shape, ) -> Result<Storage>

Batched matrix multiplication for 3D tensors.

fn exp(&self, storage: &Storage) -> Result<Storage>

Element-wise exponential function.

fn log(&self, storage: &Storage) -> Result<Storage>

Element-wise natural logarithm.

fn sqrt(&self, storage: &Storage) -> Result<Storage>

Element-wise square root.

fn pow(&self, storage: &Storage, power: f32) -> Result<Storage>

Element-wise power function.

fn sin(&self, storage: &Storage) -> Result<Storage>

Element-wise sine function.

fn cos(&self, storage: &Storage) -> Result<Storage>

Element-wise cosine function.

fn relu(&self, storage: &Storage) -> Result<Storage>

Element-wise ReLU activation function.

fn sigmoid(&self, storage: &Storage) -> Result<Storage>

Element-wise sigmoid activation function.

fn tanh(&self, storage: &Storage) -> Result<Storage>

Element-wise tanh activation function.

Provided Methods

fn is_available(&self) -> bool

Checks if this backend is available on the current system.

Default implementation returns true. Backends should override this to perform actual availability checks (e.g., CUDA device presence).

Implementors