@stuart_rowlands

QuantCDN founder, Salsa Digital partner

 Australia
quantcdn.io
Joined April 2011
  • Tweets228
  • Following242
  • Followers141
We built a custom in-browser runtime for Astro. Full tsx/mdx compilation, visual CMS, live preview, AI editing agents - all running in the browser. Create a branch, edit content visually, preview instantly, merge & deploy to the edge.
2
3
127
Result on GLM-4.7-Flash with a 2K to 16K token prompt: the checking step goes from 414 ms to 28 ms, and speculative decoding goes from 7.1 to 64.4 tokens a second. Before the fix it was slower than not using it. After, it is 1.35× faster than plain decoding.
1
28
more replies
Both fixes are up as pull requests, mlx-lm #1817 and mlx-vlm #2141. Full write-up with the charts: stuartrowlands.com/posts/mla…

One guard, 110× off the verify step. Speculative decoding on MLA models in MLX.

Every MLA model in mlx-lm only used its fast attention path when there was exactly one query, which made speculative decoding slower than not speculating. Here is what the fix is and what it measured...

stuartrowlands.com
1
35
Follow-up: mlx-vlm #2141 is closed, the work is now @Blaizzy/mlx-vlm#2150 and covers ten MLA models instead of one. On Moonlight-16B at 32K context a two-token verify step goes 380.8 ms → 22.1 ms. @Prince_Canuma CI needs an approving run when you get a chance 👀
2
1
33
The fix is to work out, per model, how many tokens the fast path stays fast for. That number depends on the model's shape, not on how long your prompt is. It is 170 for the DeepSeek-style models and 398 or 512 for GLM. Then use the fast path up to that number.
1
22
The models have a fast path and a slow path for attention, and the fast path was only switched on when handling exactly one token. Checking guesses always handles two or more. So every check took the slow path.
1
16
Speculative decoding on DeepSeek-lineage models in mlx-lm was slower than not speculating. On GLM-4.7-Flash at long context it ran at 0.15× plain decode. The cause was one comparison operator. Write-up with all the measurements and fix:
1
56
Speculative decoding works by guessing a few tokens ahead, then checking them all in one go. The checking step is supposed to be cheap. On these models it was re-doing work over the entire conversation every time, so the longer your chat, the slower it got.
1
33
Easy decision 👌
1
55
yeah we are. here's my take.. did not predict running companies through an RTS game in 2026 but here we are
Aaron Slodov
@aphysicist31 Dec 2025
millennial gamers are the best prepared generation for agentic work, they've been training for 25 years
3
1
1
339
Stuart Rowlands retweeted
Andras Bacsai
@heyandras10 Jan 2024
🛠️ The next version of @coolifyio will include two amazing features, thanks to @stuart_rowlands 👏 #buildinpublic #opensource
8
4
47
3,111
Stuart Rowlands retweeted
QuantCDN@quantcdn1 Nov 2023
A brief history of Drupal 7 and an overview of end-of-life options quantcdn.io/blog/articles/br… #Drupal #Drupal7 #DrupalPlanet #CMS #Migration #OpenSource #EndOfLife

Drupal 7 End of Life (EOL): A brief history of Drupal 7 and overview of end-of-life options

With any technology, there will be doomsday sayers and there will be the normal evolution of growth and decline. Knowing the power of Drupal and the character of the Drupal community, I’m optimistic...

quantcdn.io
2
6
486
Stuart Rowlands retweeted
QuantCDN@quantcdn4 Apr 2022
Introducing ApexWise.io. The easiest way to manage root / base / apex domain redirects (with TLS). Free for up to 3 domains!
3
7
Stuart Rowlands retweeted
Section@sectionio3 Mar 2022
Proud to help power @quantcdn as they upend the traditional #cdn paradigm. #edgecomputing #staticweb section.io/blog/quantcdn-lev…

Webscale Commerce Infrastructure and Agentic AI Platform

Webscale powers high performance commerce with secure cloud hosting, application delivery, and agentic AI that boosts speed, uptime, and conversions for Adobe, Magento, Shopware, and Woo stores.

webscale.com
2
5
We have a colleague in Ukraine paying money out of his own pocket to help people around him. If you want to make a difference to real people on the ground please donate here: fundrazr.com/nonmilitary-sup… ❤️🇺🇦

Non-military support to people of Ukraine

People of Kharkiv, the second largest city of Ukraine, have been under massive attacks by Russian forces. Donations will go directly to a person in Kharkiv who will buy necessity goods: clothes,...

fundrazr.com
4
6
This belongs in every webdev audit toolkit alongside securityheaders.com, lighthouse and the like. Nicely done @wiifm
Sean Hamlin@wiifm11 Nov 2021
Introducing my new tool, that made to distil my caching knowledge cachingscore.com - please give it a whirl and let me know any feedback

About Caching Score

Assess how strong the caching capabilities of any given site is.

cachingscore.com
2
We archived Yahoo! Answers.. a big chunk of it, at least. quantcdn.io/blog/we-archived…
4
1
4
more replies
There are 10s of millions of pages in the archive, and terabytes of data. Still likely only represents a portion of the total. If others managed to capture a more complete snapshot and want to collaborate on a more complete archive let me know
1
We'll also make it searchable / filterable if there's interest! Content is currently a little difficult to surface without it.
Load more