mirror of
https://github.com/hanzoai/papers.git
synced 2026-10-22 17:20:17 +00:00
zeekay
ff05c1b547
Decode was measured stochastic vs llama greedy (an unstated control). Corrected greedy: Metal 0.72->0.92, ROCm 0.72->0.85, Vulkan 0.52->0.60, CUDA 0.84 (device sampler, tax~0). Adds the erratum, makes sampling parity a required protocol control, revises the abstract/ BW/model-size/conclusion, updates the roofline (Vulkan 120 vs 199 GB/s, measured 213 peak). The correction increases the decode standing and the paper's credibility.
114 lines
4.4 KiB
TeX
114 lines
4.4 KiB
TeX
% GENERATED by scripts/assemble_paper.sh -- do not hand-edit.
|
|
\renewcommand{\rRocmDecodeHci}{1.1}
|
|
\renewcommand{\rRocmDecodeH}{113.7}
|
|
\renewcommand{\rRocmDecodeL}{133.8}
|
|
\renewcommand{\rRocmDecodeR}{0.85}
|
|
\renewcommand{\rRocmDecodeV}{LOSS}
|
|
\renewcommand{\rRocmEffBW}{146}
|
|
\renewcommand{\rRocmPrefillHci}{14.0}
|
|
\renewcommand{\rRocmPrefillH}{3693.6}
|
|
\renewcommand{\rRocmPrefillLongHci}{47.0}
|
|
\renewcommand{\rRocmPrefillLongH}{3506.1}
|
|
\renewcommand{\rRocmPrefillLongL}{4551.3}
|
|
\renewcommand{\rRocmPrefillLongR}{0.77}
|
|
\renewcommand{\rRocmPrefillLongV}{LOSS}
|
|
\renewcommand{\rRocmPrefillL}{4742.5}
|
|
\renewcommand{\rRocmPrefillMaxHci}{93.0}
|
|
\renewcommand{\rRocmPrefillMaxH}{2840.8}
|
|
\renewcommand{\rRocmPrefillMaxL}{4103.3}
|
|
\renewcommand{\rRocmPrefillMaxR}{0.69}
|
|
\renewcommand{\rRocmPrefillMaxV}{LOSS}
|
|
\renewcommand{\rRocmPrefillRaggedHci}{19.4}
|
|
\renewcommand{\rRocmPrefillRaggedH}{3603.8}
|
|
\renewcommand{\rRocmPrefillRaggedL}{4678.0}
|
|
\renewcommand{\rRocmPrefillRaggedR}{0.77}
|
|
\renewcommand{\rRocmPrefillRaggedV}{LOSS}
|
|
\renewcommand{\rRocmPrefillR}{0.78}
|
|
\renewcommand{\rRocmPrefillV}{LOSS}
|
|
\renewcommand{\rRocmReach}{85\%}
|
|
\renewcommand{\rRocmRoofBW}{172}
|
|
\renewcommand{\rVulkanDecodeHci}{1.7}
|
|
\renewcommand{\rVulkanDecodeH}{93.9}
|
|
\renewcommand{\rVulkanDecodeL}{155.4}
|
|
\renewcommand{\rVulkanDecodeR}{0.60}
|
|
\renewcommand{\rVulkanDecodeV}{LOSS}
|
|
\renewcommand{\rVulkanEffBW}{120}
|
|
\renewcommand{\rVulkanPrefillHci}{10.9}
|
|
\renewcommand{\rVulkanPrefillH}{3042.3}
|
|
\renewcommand{\rVulkanPrefillLongHci}{37.6}
|
|
\renewcommand{\rVulkanPrefillLongH}{1475.0}
|
|
\renewcommand{\rVulkanPrefillLongL}{4256.8}
|
|
\renewcommand{\rVulkanPrefillLongR}{0.35}
|
|
\renewcommand{\rVulkanPrefillLongV}{LOSS}
|
|
\renewcommand{\rVulkanPrefillL}{4547.7}
|
|
\renewcommand{\rVulkanPrefillMaxHci}{33.3}
|
|
\renewcommand{\rVulkanPrefillMaxH}{922.8}
|
|
\renewcommand{\rVulkanPrefillMaxL}{3850.3}
|
|
\renewcommand{\rVulkanPrefillMaxR}{0.24}
|
|
\renewcommand{\rVulkanPrefillMaxV}{LOSS}
|
|
\renewcommand{\rVulkanPrefillRaggedHci}{241.4}
|
|
\renewcommand{\rVulkanPrefillRaggedH}{1098.9}
|
|
\renewcommand{\rVulkanPrefillRaggedL}{4449.6}
|
|
\renewcommand{\rVulkanPrefillRaggedR}{0.25}
|
|
\renewcommand{\rVulkanPrefillRaggedV}{LOSS}
|
|
\renewcommand{\rVulkanPrefillR}{0.67}
|
|
\renewcommand{\rVulkanPrefillV}{LOSS}
|
|
\renewcommand{\rVulkanReach}{60\%}
|
|
\renewcommand{\rVulkanRoofBW}{199}
|
|
\renewcommand{\rCudaDecodeHci}{1.0}
|
|
\renewcommand{\rCudaDecodeH}{86.2}
|
|
\renewcommand{\rCudaDecodeL}{102.8}
|
|
\renewcommand{\rCudaDecodeR}{0.84}
|
|
\renewcommand{\rCudaDecodeV}{LOSS}
|
|
\renewcommand{\rCudaEffBW}{111}
|
|
\renewcommand{\rCudaPrefillHci}{608.2}
|
|
\renewcommand{\rCudaPrefillH}{7873.6}
|
|
\renewcommand{\rCudaPrefillLongHci}{745.2}
|
|
\renewcommand{\rCudaPrefillLongH}{7285.6}
|
|
\renewcommand{\rCudaPrefillLongL}{7326.8}
|
|
\renewcommand{\rCudaPrefillLongR}{0.99}
|
|
\renewcommand{\rCudaPrefillLongV}{PARITY}
|
|
\renewcommand{\rCudaPrefillL}{6190.0}
|
|
\renewcommand{\rCudaPrefillMaxHci}{217.1}
|
|
\renewcommand{\rCudaPrefillMaxH}{7474.8}
|
|
\renewcommand{\rCudaPrefillMaxL}{6871.3}
|
|
\renewcommand{\rCudaPrefillMaxR}{1.09}
|
|
\renewcommand{\rCudaPrefillMaxV}{WIN}
|
|
\renewcommand{\rCudaPrefillRaggedHci}{228.7}
|
|
\renewcommand{\rCudaPrefillRaggedH}{7961.3}
|
|
\renewcommand{\rCudaPrefillRaggedL}{7250.2}
|
|
\renewcommand{\rCudaPrefillRaggedR}{1.10}
|
|
\renewcommand{\rCudaPrefillRaggedV}{WIN}
|
|
\renewcommand{\rCudaPrefillR}{1.27}
|
|
\renewcommand{\rCudaPrefillV}{WIN}
|
|
\renewcommand{\rCudaReach}{84\%}
|
|
\renewcommand{\rCudaRoofBW}{132}
|
|
\renewcommand{\rMetalDecodeHci}{1.5}
|
|
\renewcommand{\rMetalDecodeH}{233.4}
|
|
\renewcommand{\rMetalDecodeL}{253.8}
|
|
\renewcommand{\rMetalDecodeR}{0.92}
|
|
\renewcommand{\rMetalDecodeV}{LOSS}
|
|
\renewcommand{\rMetalEffBW}{299}
|
|
\renewcommand{\rMetalPrefillHci}{11.6}
|
|
\renewcommand{\rMetalPrefillH}{3386.0}
|
|
\renewcommand{\rMetalPrefillLongHci}{4.3}
|
|
\renewcommand{\rMetalPrefillLongH}{3301.5}
|
|
\renewcommand{\rMetalPrefillLongL}{3696.9}
|
|
\renewcommand{\rMetalPrefillLongR}{0.89}
|
|
\renewcommand{\rMetalPrefillLongV}{LOSS}
|
|
\renewcommand{\rMetalPrefillL}{4097.5}
|
|
\renewcommand{\rMetalPrefillMaxHci}{113.6}
|
|
\renewcommand{\rMetalPrefillMaxH}{2767.3}
|
|
\renewcommand{\rMetalPrefillMaxL}{2863.0}
|
|
\renewcommand{\rMetalPrefillMaxR}{0.97}
|
|
\renewcommand{\rMetalPrefillMaxV}{PARITY}
|
|
\renewcommand{\rMetalPrefillRaggedHci}{9.4}
|
|
\renewcommand{\rMetalPrefillRaggedH}{3345.3}
|
|
\renewcommand{\rMetalPrefillRaggedL}{3988.9}
|
|
\renewcommand{\rMetalPrefillRaggedR}{0.84}
|
|
\renewcommand{\rMetalPrefillRaggedV}{LOSS}
|
|
\renewcommand{\rMetalPrefillR}{0.83}
|
|
\renewcommand{\rMetalPrefillV}{LOSS}
|
|
\renewcommand{\rMetalReach}{92\%}
|
|
\renewcommand{\rMetalRoofBW}{325}
|