pub fn bench_format(quant_format: &str) -> &strExpand description
The bench format key an engine quant name is measured under.
The benchmark keys its numbers by expert format, not by model,
because the CPU-MoE-vs-PCIe-gather ratio the choice rides on is
dominated by (format, hardware) – so a profile taken on one
workload transfers to any model with the same expert format on the
same card. Most engine names are already the bench name; mxfp4 is
benched under its kernel’s name, mxfp4_triton.
Anything not in the table is passed through unchanged, which is
deliberate rather than lossy: an unmapped name finds no entry in the
profile, the lookup yields None, and the caller keeps its safe
offload default. Only the offload-family formats that have a CPU MoE
weight path can ever resolve to hybrid, and those are exactly the
ones listed here.