Struct Capabilities 

pub struct Capabilities {
    pub supports_fp16: bool,
    pub supports_bf16: bool,
    pub supports_fp8: bool,
    pub tensor_cores: bool,
    pub peer_access: bool,
    pub unified_memory: bool,
    pub cluster_launch: bool,
    pub async_copy: bool,
    pub max_threads_per_block: u32,
    pub max_shared_mem_per_block: u32,
    pub warp_size: u32,
}

Per-backend capability report.

A concrete backend populates this from its driver during ComputeBackend::capabilities. Consumers use it to decide whether a mixed-precision or Tensor-Core path is available before issuing work.

The Default value is the conservative CPU profile: no accelerated precisions, no Tensor Cores, no peer access, but generous thread/shared-memory limits that any host can honour.

Fields

supports_fp16: bool

Native FP16 (half precision) compute is available.

supports_bf16: bool

Native BF16 (bfloat16) compute is available.

supports_fp8: bool

FP8 (E4M3/E5M2) compute is available.

tensor_cores: bool

Dedicated matrix/Tensor-Core (WMMA / mma.sync) units are present.

peer_access: bool

Peer-to-peer access between devices of this backend is possible.

unified_memory: bool

Unified / managed memory (host and device share an address space).

cluster_launch: bool

Thread-block clusters (Hopper sm_90+) are supported.

async_copy: bool

Asynchronous global→shared copy (cp.async, Ampere sm_80+).

max_threads_per_block: u32

Maximum threads per block / workgroup.

max_shared_mem_per_block: u32

Maximum shared / local memory per block in bytes.

warp_size: u32

Warp / wavefront / subgroup width in lanes.

Implementations

impl Capabilities

pub const fn cpu() -> Self

Conservative capability profile for a pure-CPU reference backend.

No accelerated precisions or matrix units, but limits large enough that the consumer never rejects a launch for being “too big”.

pub const fn supports_mixed_precision(&self) -> bool

Returns true if any reduced-precision compute path is available.

pub const fn can_use_tensor_cores(&self, m: usize, n: usize, k: usize) -> bool

Returns true if a GEMM of the given shape can plausibly use a Tensor-Core path: matrix units present and all three dimensions are multiples of the smallest standard WMMA tile (16).

Trait Implementations

impl Clone for Capabilities

fn clone(&self) -> Capabilities

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for Capabilities

impl Debug for Capabilities

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for Capabilities

fn default() -> Self

Returns the “default value” for a type.

impl Display for Capabilities

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Eq for Capabilities

impl PartialEq for Capabilities

fn eq(&self, other: &Capabilities) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for Capabilities