Module dtype 

The dtype system: Dtype/Float/tensor traits kernel code is generic over.

Re-exports

pub use bytes::FloatBytes;

Modules

bytes

Host-only float byte-conversion helpers (kept separate from Float so the DSL-embedded dtype/mod.rs source doesn’t need them — see teeny-triton’s crate docs). Host-side float byte serialization.

Traits

AddOffsets

Adding a tensor of integer offsets to Self (e.g. a pointer).

BF16

bfloat16: 16-bit float with an 8-bit exponent (same range as f32, reduced precision).

Bool

The boolean dtype.

BoolTensor

A tensor of bool.

Comparison

Element-wise comparison against Rhs, producing a boolean-valued tensor.

Dtype

Base marker trait for all types that can flow through the system as tensor elements.

EagerTensor

Marker trait for eager (non-symbolic) tensors.

F8E4M3FN

8-bit float, 4 exponent + 3 mantissa bits.

F8E4M3FNUZ

8-bit float, 4 exponent + 3 mantissa bits, no infinities (unsigned zero variant).

F8E5M2

8-bit float, 5 exponent + 2 mantissa bits.

F8E5M2FNUZ

8-bit float, 5 exponent + 2 mantissa bits, no infinities (unsigned zero variant).

Float

Floating-point dtypes.

I4

4-bit integer.

I32Tensor

A tensor of i32.

Int

Integer dtypes.

Num

Numeric scalar dtypes; BITS is used for device buffer allocation sizing.

Pointer

A device pointer to elements of dtype D. A Pointer is itself a Dtype (so it can be stored in tensors), but has no BITS requirement.

RankedTensor

A tensor with a statically-known shape.

Tensor

A tensor of dtype D and rank RANK.