Struct Metal 

pub struct Metal;

The Metal backend (ferrox-metal).

Trait Implementations

impl BackendCaps for Metal

const GEMM_FALLBACK: &'static str = "Metal N x matvec batch"

apply_gpu_batch re-reads the whole weight matrix once per position, on the GPU. Still Metal, still the 13.7x shape.

fn matvec_kernel(kind: QuantKind) -> Option<&'static str>

As the kernel name [ferrox_metal::gpu::matvec_launch_meta] resolves.

This is the single source of truth for that question. It is not #[cfg(feature = "metal")]-gated deliberately: the table is a property of the kernel set, and gating it would make it untestable on the builds that run cargo test --workspace.

Duplicating this list is how IQ4_XS batched prefill silently ran on the CPU — metal_kind_supported and apply_gpu_batch’s kind table disagreed by exactly one entry, and the only symptom was a benchmark row 13.7x behind. Every Metal-kind question now routes through here.

fn gemm_supported(kind: QuantKind) -> bool

The *_mul_mm_sg simdgroup GEMMs.

The invariant that this set equals Metal::matvec_kernel’s is asserted by a test, so adding a matvec kernel without a GEMM fails the suite instead of a benchmark.

const ID: Backend = Backend::Metal

How crate::kernel_registry reports this backend. Dispatch and observability read the same constant, so a backend cannot be dispatched to under one name and reported under another.

const NAME: &'static str = "Metal"

Human-readable name, for the one message a dispatch failure prints.