Module ops 

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

Note on scalars: in the old design op scalars were typed T (tied to the tensor’s element type). Here a Float tensor’s precision is a runtime DType, so all op scalars are stored as f64 — the widest float — and cast back to the tensor’s actual precision when the op runs or its gradient is computed.

Re-exports

pub use construct::DEFAULT_FLOAT;

pub use construct::DEFAULT_INT;

Modules

construct

Structs

Slice

A slice range with start, optional end (negative = from end, None = to end), and step.

Enums

BinaryOp

CmpOp

FloatUnaryOp

Indexer

One element of a fancy-indexing operation.

Op

A node in the (implicit) computation graph: the operation that produced a Float tensor, holding Arc references to its inputs. Only Float tensors record ops — Int/Bool tensors are never differentiated.

ReduceOp

UnaryOp

ViewOp

A view operation: it shares the source tensor’s storage under a new [Layout] without touching device kernels.

Traits

IndexOp

Trait for fancy indexing via the .i() method.

IndexingDTypeKind

NumericDTypeKind

ReduceDTypeKind

ShapeDTypeKind

TransferDTypeKind

Kind-level dispatch for Tensor::to_device.