zeekay dcec1ab7c0 paper: Hanzo ML + Engine unified multi-backend inference-stack overview ...

Measured head-to-head vs llama.cpp on the same silicon: the flash-attn dense-prefill
result on GB10 (+42%, still 0.77-0.84x, stated honestly as the one trailing axis),
the CUDA decode-graph / MoE-prefill / GDN-kernel leads, AMD gfx1151 prefill+decode,
and cross-backend decode parity. Compiles clean (13 pp).

2026-06-27 23:54:22 +00:00

261 KiB

Raw History

View Raw