Theme concepts for the Deskbar shell

Twelve looks to pick from: eight classic desktops rebuilt as faithfully as CSS allows, and four new ones. Each card lists what it borrows, how it maps onto the theme's palette, window style, wallpaper and dock model, and the gzipped size of its CSS here. control-panel.css has about 4.6 KB of its 12 KB budget left, so each look should ship as its own lazy stylesheet, loaded the same way. Click any screen to enlarge it, Esc to close. The Spatial card follows the pointer, and the Chroma card has a working light and dark switch.

Classic

Unix workstations, NeXT, the Mac, Amiga and Atari

NeXTSTEP 3.3

Four exact greys, Helvetica, black key title bars, scrollers on the left and the tile dock down the right edge.

kv-cache-quantisation.rtf ~/Posts

Bringing K/V Context Quantisation to Ollama

Explaining the concept of K/V context cache quantisation, why it matters and the journey to integrate it into Ollama.

Quantising the K/V cache to q8_0 roughly halves the memory the context uses, with a negligible hit to quality. That means larger context windows, or bigger models, on the same hardware.

With q4_0 the savings are closer to three quarters, though the quality trade-off starts to show on longer contexts.

Save

Apps

Library

Posts

Talks

agent-skills.rtf

mcp-devtools.rtf

nvidia-p2p.rtf

sampling.rtf

Name:

kv-cache-quantisation

Cancel OK

Notes.rtf

Talks

Mac OS 9 Platinum

Pinstriped title bars, lavender scroll thumbs, the rainbow Apple menu, Control Strip and white-boxed icon labels.

FileEditViewWindowHelp 9:41 PM

Appearance

kv-cache quantisation

Bringing K/V Context Quantisation to Ollama

Explaining the concept of K/V context cache quantisation, why it matters and the journey to integrate it into Ollama.

Quantising the K/V cache to q8_0 roughly halves the memory the context uses, with a negligible hit to quality. That means larger context windows, or bigger models, on the same hardware.

With q4_0 the savings are closer to three quarters, though the trade-off starts to show on longer contexts.

Appearance

ThemesAppearanceFontsDesktop

Select a theme:

Mac OS Default

Deskbar

Hi-Tech

DeleteSave Theme...

Macintosh HD

Posts

Trash

Mac OS X 10.1 Aqua

Pinstripes, gel traffic lights and scroll thumbs, a pulsing default button in a sheet, and the white plexiglass Dock with black triangles.

TextEditFileEditFormatWindowHelpThu 9:41 PM

Posts

Back

Computer

Home

Favorites

Applications

Documents

Library

Movies

Posts

Talks

Photos

agent-skills

kv-cache

nvidia-p2p

kv-cache-quantisation.rtf

Bringing K/V Context Quantisation to Ollama

Quantising the K/V cache to q8_0 roughly halves the memory the context uses, with a negligible hit to quality. That means larger context windows, or bigger models, on the same hardware.

With q4_0 the savings are closer to three quarters.

Save changes to this document before closing? If you don't save, your changes will be lost.

Don't SaveCancelSave

Macintosh HD

Posts

CDE on Solaris

The Crimson palette, chunky Motif frames with resize grooves, a bitmap backdrop and the Front Panel with its four workspace buttons.

Mailer

Calendar

Terminal

WindowEditOptionsHelp

ultra5% uname -sr SunOS 5.8 ultra5% cd ~/posts; ls agent-skills/ kv-cache/ nvidia-p2p/ sampling/ ultra5%

Text Editor - kv-cache.txt

FileEditFormatOptionsHelp

Bringing K/V Context Quantisation to Ollama

Explaining the concept of K/V context cache quantisation, why it matters and the journey to integrate it into Ollama.

Quantising the K/V cache to q8_0 roughly halves the memory the context uses, with a negligible hit to quality. That means larger context windows, or bigger models, on the same hardware.

Line: 14Col: 40INS

Sun OpenWindows (OPEN LOOK)

Pushpin menus, L-shaped resize corners, oval menu buttons, window-mark buttons and the elevator-on-a-cable scrollbar.

cmdtool - /bin/csh

sparc10% cd posts sparc10% ls agent-skills kv-cache nvidia-p2p sampling sparc10% more kv-cache/intro sparc10%

Text Editor - kv-cache.txt (edited), dir; ~/posts

FileViewEditFindSave

Bringing K/V Context Quantisation to Ollama

Explaining the concept of K/V context cache quantisation, why it matters and the journey to integrate it into Ollama.

Quantising the K/V cache to q8_0 roughly halves the memory the context uses, with a negligible hit to quality. That means larger context windows, or bigger models, on the same hardware.

Text Editor V3Edited

mailtool

cmdtool

Amiga Workbench 1.3

The four-colour blue, white, black and orange palette, Topaz-style type, depth gadgets, complement-highlighted menus and the System Request.

WorkbenchDiskSpecial

NotePad: kv-cache

Bringing K/V Context Quantisation to Ollama

Explaining the concept of K/V context cache quantisation, why it matters and the journey to integrate it into Ollama.

Quantising the K/V cache to q8_0 roughly halves the memory the context uses, with a negligible hit to quality. That means larger context windows, or bigger models, on the same hardware.

System Request

Please insert volume
Posts: in any drive

RetryCancel

Workbench1.3

RAM DISK

Trashcan

Atari ST GEM

Green desktop, drawer-shaped drive icons, dithered move bars, the info line, dashed menu separators and a double-bordered alert.

DeskFileViewOptions

FLOPPY A

HARD DISK

TRASH

C:\POSTS\*.*

40312 bytes used in 5 items.

AGENTS.DOC

KVCACHE.DOC

P2P.DOC

1ST_WORD: KVCACHE.DOC

Page 1 Line 4 Col 12 INSERT

[----!----!----!----!----!----!--]

Bringing K/V Context Quantisation to Ollama

Quantising the K/V cache to q8_0 roughly halves the memory the context uses, with a negligible hit to quality.

That means larger context windows, or bigger models, on the same hardware.

Open

Show Info...

------------------

New Folder...

Close Window

------------------

Format...

The document KVCACHE.DOC
has been changed.
Save it before closing?

DiscardSave

GNOME 2 Clearlooks

Two panels, the foot menu, Metacity's blue gradient title bars, Tango icons, Verdana-like Vera type and a GTK info bar.

ApplicationsPlacesSystem Thu 24 Sep, 9:41 PMsam

Computer

sam's Home

posts - File Browser

FileEditViewGoBookmarksHelp

Places

sam

Desktop

File System

Trash

agent-skills

kv-cache

nvidia-p2p

index.md

kv-cache.md (~/posts) - gedit

FileEditViewSearchToolsDocumentsHelp

The file has changed on disk.IgnoreReload

kv-cache.md

# Bringing K/V Context Quantisation to Ollama Explaining the concept of K/V context cache quantisation, why it matters and the journey to [integrate it into Ollama]. Quantising the K/V cache to q8_0 roughly halves the memory the context uses.

MarkdownLn 5, Col 42INS

posts - File Browser kv-cache.md - gedit

Modern

New looks that use CSS nobody had in 1995: registered custom properties, oklch, 3D transforms, scroll timelines, view transitions

Phosphor

A P3 amber tube: text-mode windows with bloom, scanlines, a rolling refresh band, burn-in and a function-key bar for a dock.

PHOSPHOR/OSFileEditViewTermLOAD 0.42 MEM 61% 21:41

tty2[x]

$ nvidia-smi topo -p2p r GPU0 GPU1 GPU0 X OK GPU1 OK X $ vllm serve qwen3.5-35b \ --tensor-parallel-size 2 INFO P2P access: enabled INFO 2x RTX 3090, NVLink off $ █

nvidia-p2p.md[-][x]

Patching NVIDIA's driver for P2P

NVIDIA artificially restricts peer-to-peer GPU communication to their enterprise cards. Turns out this is a software limitation, not a hardware one.

I patched my drivers, hacked vLLM to use it, and got a 10-30% throughput improvement on dual RTX 3090s.

[ Open ][ Copy link ]

BACKUP

FEEDS

1Help2Posts3View4Edit5Term6Feeds7Find9Menu10Quit

Spatial

Glass panes float at different depths over a soft landscape. Move the pointer: near panes and their shadows shift more than far ones, and the glass catches the light.

Photos

9:41

Posts

Photos

27 Apr 20269 min read

Measuring Model Quantisation Quality with KL Divergence

KL divergence against a known-good reference answers "how much did this quant change the model's behaviour?" rather than "how good is this model overall?".

Read postShare

Broadsheet

Neo-brutalist newsprint: ink rules, hard offset shadows, Space Grotesk headlines over Literata body, and a scroll-driven reading bar.

SKILLS

SMCLEODPostsTalksPhotosAboutNo.142 / 21:41

Index

07.2026Agent Skills: common pitfalls

04.2026Quant quality with KL divergence

03.2026Apple Silicon 4K HiDPI limits

02.2026NVIDIA P2P on consumer GPUs

Reading

Writing and Reviewing Agent Skills

Sam McLeod04 Jul 20268 min

I write and review a lot of Agent Skills, and find myself pushing back in reviews, as many authors assume they're writing "just another markdown prompt" without considering that they're working with one component of an agentic system.

A skill is one component of an agentic system.

The feedback usually starts with the description, because that is the only part the model sees before deciding to load it.

Read on

RSS

CV.PDF

Chroma

Material-You-style dynamic colour from a living wallpaper: an animated oklch mesh gradient drives every surface, chip and button in step, in light and dark.

Thu 24 Sep Search posts 21:41

Chiptunes

Space DebrisCaptain, 1991, 4-ch MOD

Posts

LLMsQuantisation

Measuring Model Quantisation Quality with KL Divergence

KL divergence against a known-good reference answers "how much did this quant change the model's behaviour?" rather than "how good is this model overall?".

38%

Listen

Feeds

About