Efficient Multi-Scale Deformable Attention on GPUs

A journal offprint set by a careful compositor: serif prose in a 34em measure, italic notes in the margin, one thin red rule where the article begins.

research.khws.io/efficient-msdaEfficient Multi-Scale Deformable Attention on GPUs
Screenshot of https://research.khws.io/efficient-msda/, © the authors. Shown for reference. Request removal
Venue
TMLR 2026
Year
2026
First author
Kurt H. W. Stolle
Field
efficient deep learning / GPU kernels
Template
custom Nuxt site (research.khws.io)
Theme
Light

Checked against the rendered page. No figures, tables or video on the page. Right rail position is set by a script variable (--right-rail-left). Dark scheme exists via footer toggle; light documented as default.

Why it is here

Most distinctive page in the pool. Editorial serif type, a left metadata rail, italic margin sidenotes, numbered findings and ruled section dividers make it read like a well-set journal article. The measure is comfortable and the hierarchy is clear.

Overview

Color Palette

Typography

Source Serif 4
Efficient Multi-Scale Deformable Attention on GPUs Abstract · Method · Results · BibTeX — 0123456789
Figtree
Efficient Multi-Scale Deformable Attention on GPUs Abstract · Method · Results · BibTeX — 0123456789
Source Code Pro
Efficient Multi-Scale Deformable Attention on GPUs Abstract · Method · Results · BibTeX — 0123456789

Spacing & Shape

Page Modules

Guidelines

More like this