pub fn w4a16_v2_kernel(gpu: &dyn GpuBackend) -> KernelHandleExpand description
Try to load an optional kernel, logging at debug level if it’s not found.
Returns KernelHandle(0) (null) on failure — callers must check before use.
Debug (not warn) because misses are expected when a model doesn’t use a
given feature: e.g. Qwen3-Coder-Next (GDN+attention) never calls MLA
kernels, but the layer builder still probes them. Warning on expected
misses drowned out genuine problems in startup logs.
Resolve the w4a16_gemm_t_m128_v2 handle honoring ATLAS_W4A16_VARIANT.
One resolver for the THREE sites that dispatch on this handle (attention
projections, dense-FFN prefill, SSM batched decode), so variant policy and
rollback live in exactly one place. Default (unset) resolves to a ZERO
handle — v1 everywhere — because the 27B port measured SLOWER than v1
(see body). ATLAS_W4A16_VARIANT=v2 opts in on all three sites at once;
requesting it on a target without the kernel is a HARD startup error
(fail fast, not a silent fallback discovered in a perf regression).