module data

Typed data entry: build a tensor from a slice, with the element type naming the dtype. Half-precision elements ride bit-carrier newtypes (F16, BF16); the payload stays at the runtime's dtype end to end, and any numeric conversion runs through the runtime (f_tensor_to_dtype), never through host float math.

Three entry forms: Api::tensor_from_slice COPIES the slice in (the tensor owns its memory from then on); Api::f_tensor_from_blob / Api::f_tensor_from_blob_strided BORROW the slice as a view (no copy; the tensor cannot outlive the &mut borrow); and Api::f_tensor_from_blob_owned ADOPTS a Vec (no copy; the runtime drops it when the last reference goes away); the blob forms come as an f_ Result primary plus a panicking twin. The write half is Api::f_tensor_copy_from: host data into an existing tensor.

let x = api.tensor_from_slice(&[1.0f32, 2.0, 3.0, 4.0], &[2, 2])?;
let mut buf = [1.0f32; 8];
let view = api.f_tensor_from_blob(&mut buf, &[8])?; // addresses buf, no copy

BF16 (struct)

A bfloat16, carried as its bit pattern.

Implements: Clone, Copy, Debug, Element, Eq, PartialEq, StructuralPartialEq

F16 (struct)

A 16-bit IEEE-754 half, carried as its bit pattern.

Implements: Clone, Copy, Debug, Element, Eq, PartialEq, StructuralPartialEq

Element (trait)

An element type a slice can enter the tensor world as. DTYPE is the runtime dtype the bits are read at.

Safety

An implementation asserts that Self's in-memory representation is exactly DTYPE's element encoding, densely packed; the runtime reads the slice's bytes at that dtype.