mirror of
https://github.com/hanzoai/papers.git
synced 2026-10-29 07:34:31 +00:00
zeekay
3645e54c9f
Adopt the Metal lane's byte-comparable protocol: three verbatim workload classes, gamma sweep, multiple + vs-llama + lossless. Provisional Metal reference (1.23x WIN sustained-reuse @ g8), re-measured per backend at battery via spec_decode.sh, stamped to the spec cut (novel floor 0.47->0.71x at ml 0.11.91). results-spec.tex seam; kernel board untouched.
12 lines
633 B
TeX
12 lines
633 B
TeX
% Architecture-row (prompt-lookup) values. Provisional: the Metal-lane reference measurement
|
|
% (bench-protocols/spec-decode-metal.md, Qwen3-1.7B, engine 1.7.88, greedy, N=5--7). Re-measured
|
|
% independently per backend at battery time via scripts/spec_decode.sh and re-stamped to the
|
|
% spec cut (ml 0.11.91: a verify-round matvec fix lifts the novel floor 0.47->0.71x, lossless).
|
|
\renewcommand{\rSpecSustainMult}{1.43}
|
|
\renewcommand{\rSpecSustainVsLl}{1.23}
|
|
\renewcommand{\rSpecCodeMult}{0.87}
|
|
\renewcommand{\rSpecCodeVsLl}{0.73}
|
|
\renewcommand{\rSpecNovelMult}{0.48}
|
|
\renewcommand{\rSpecNovelVsLl}{0.42}
|
|
\renewcommand{\rSpecGamma}{8}
|