Crate luma_tensor 

luma-tensor — a tensor computation library with:

• Compile-time kind separation (Float, Int, Bool) so a Float tensor cannot accidentally participate in a Bool operation.
• Runtime precision (f32, f64, i32, …) within each kind, decided at construction time via DType.
• A device abstraction (Device) that lets the same code run on Cpu or (in the future) Cuda.
• A tape-based autograd engine that tracks only Float-kind tensors.

Re-exports

pub use device::cpu::Cpu;

pub use device::cpu::CpuBoolStorage;

pub use device::cpu::CpuFloatStorage;

pub use device::cpu::CpuIntStorage;

pub use device::BoolOps;

pub use device::Device;

pub use device::FloatOps;

pub use device::IntOps;

pub use dtype::Bool;

pub use dtype::DType;

pub use dtype::DTypeKind;

pub use dtype::Float;

pub use dtype::Int;

pub use dtype::KindTag;

pub use dtype::Storage;

pub use dtype::FloatDType;

pub use dtype::IntDType;

pub use dtype::BoolDType;

pub use dynamic::DynTensor;

pub use error::Error;

pub use error::Result;

pub use grad::FloatMeta;

pub use grad::GradStore;

pub use grad::NoGradGuard;

pub use grad::TensorMeta;

pub use grad::is_grad_enabled;

pub use grad::set_grad_enabled;

pub use ops::BinaryOp;

pub use ops::CmpOp;

pub use ops::FloatUnaryOp;

pub use ops::Op;

pub use ops::ReduceOp;

pub use ops::TransferDTypeKind;

pub use ops::UnaryOp;

pub use ops::ViewOp;

pub use ops::IndexOp;

pub use ops::Indexer;

pub use ops::Slice;

pub use scalar::Scalar;

pub use tensor::D;

pub use tensor::Dim;

pub use tensor::Dims;

pub use tensor::Layout;

pub use tensor::Shape;

pub use tensor::StorageIndices;

pub use tensor::Tensor;

pub use tensor::TensorId;

pub use tensor::TensorImpl;

Modules

device

dtype

dynamic

Type-erased tensor enum for runtime kind dispatch.

error

grad

ops

The autograd computation-graph node type Op and its op-kind enums.

scalar

Scalar values mirroring crate::DType variants.

tensor

Macros

bail

dispatch_float

Run $body with $data bound to the inner &Vec<t> of a CpuFloatStorage, then wrap the returned Vec<t> back into a CpuFloatStorage.

dispatch_float2

Dispatch two float storages of the SAME variant; errors on mismatch.

dispatch_float2_raw

Dispatch two float storages of the SAME variant, $body yields a raw value.

dispatch_float_raw

Like dispatch_float but $body yields a non-storage value (e.g. Vec).

dispatch_int

dispatch_int2

dispatch_int2_raw

dispatch_int_raw

no_grad

Temporarily disable gradient tracking within the current scope.

s

Create a Slice with python-like syntax.