Benchmarking vLLM on 2× RTX 3090 — Post-NVLink: The Verdict Flips
Installing an NVLink bridge on 2× RTX 3090 and re-running the sweep — dense models cross the 2500ms TTFT budget, and TP=2 becomes the default across the board.
// the full back-catalogue
44 posts on systems, data, and AI — dating all the way back.
Installing an NVLink bridge on 2× RTX 3090 and re-running the sweep — dense models cross the 2500ms TTFT budget, and TP=2 becomes the default across the board.
Six sweeps on 2× RTX 3090 without NVLink — which coding models actually work at TP=2, what the TTFT floor is, and why dense models fail without the bridge.
Hardware, software, and methodology for a systematic TP=2 coding model sweep on an Ampere dual-GPU rig — what the sweep does, how it works, and what 'best config' actually means.
What a small to medium enterprise actually needs to do AI now, in a way that makes sense.
The ELK stack has been becoming more and more popular, specifically Elasticsearch
In any web project it often happens that you have an entity, say a Group, to which you want to add a linked entity, say a Student.
We had the opportunity to muck around with a couple of technologies at Tutuka
I presented this at Tech4Africa 2013
Usually when you're working with data you expect some kind of default ordering.
I've been working on teach-a-rific.
Symfony 2 comes with a great security component
Reading the PHP docs, it's not immediately clear how the `array_merge_recursive`
The Swiftmailer
It can become quite difficult to keep PHP code structured and clean.
While working on Backend-PHP I needed to do a proper implementation of the Decorator Pattern in PHP.
I've been getting more into Testing (specifically Unit Testing) the past few weeks, and although I'm still struggling to firstly get it right, and secondly see the long term benefits, I know that the quality of my code is already better.
When you `json_encode` an object, it will ignore all protected and private properties if the `json_encode` isn't called from within the object.
I'm trying to use puppet to setup standard dev and (eventually) production environments.
I like frameworks.
I really like working with PHP.
Johnny (an asthma sufferer and fictitious character) died yesterday.
I hate repetitive work.
The last two weeks I had the privilege of staying in a house on the beach near Mosselbay.
I have the privilege to be the Head of Development of Branded Internet, a white label ISP in South Africa.
I'm working on some statistics surrounding BI's API Documentation coverage: Just how much user documentation have we written?
Looking for a new challenge, I started a small project in node.
I need to parse Emails in node.
A a student I went to a career psychologist to explore my career needs and drives, and got told that I like solving problems in a creative way.
I've setup zacoders.
If you like simplicity, you might like to simplify your website's URL.
At long last I've been able to find the time to port my website from drupal to backend
I've been working on backend
I'm busy creating a site for a client who wants to know when certain internal links get clicked.
I'm playing around with [Google Gears][gears], particularly the database part, just to see what it's usefull for, and if I can use it in an app I'm working on.
I want to generate CSS on the fly for a personal project.
In one of my pet projects, I redirect all requests to index.
I'm playing around with Google Gears
Bakend: PHP framework, as it's now called, now has it's own website!