Jacob Tomlinson

Training Large Language Models (LLMs) requires a vast amount of input data, and the higher the quality of that data the better the model will be at producing useful natural language.

Abstract This year I built a library that already exists.

Dask is a flexible library for parallel computing in Python.

I’ve been working on kr8s for a while now and one of my core goals is to build a Python library for Kubernetes that is the most simple, readable and produces the most maintainable code.

The Dask blog is a bit neglected these days.

By using Dask to scale out RAPIDS workloads on Kubernetes you can accelerate your workloads across many GPUs on many machines.

Since upgrading to a UniFi Dream Machine (UDM) Pro I’ve had a problem with some of my UniFi devices showing as offline.

This post will be updated with notes from the livestream throughout the day.

For the last few months I’ve been tinkering with a new Kubernetes client library for Python called kr8s.

When writing tests the balance between avoiding indirection and DRY-ness should be much more weighted towards avoiding indirection than in the code it is testing.

As software engineers we should all be able to communicate the things we have built to others, but giving a formal demo of something you’ve been working on can be daunting.

The more we scale up our workloads the more we run into bugs that only appear at scale.

I was invited to join the Kubernetes Batch Working Group (k8s-batch-sig) to give an overview of how Dask runs on Kubernetes, especially for batch style workloads with DaskJob.

Did you know that the Dask scheduler has a --jupyter flag that will start a Jupyter server running within the Dask Dashboard?

When writing and speaking about linux container technologies I’m trying to be more intentional with the words I use, which means often avoiding the word docker.

Sometimes I want to oversubscribe the GPUs in my Kubernetes cluster.

Sometimes when I give live demos with Kubernetes clusters I want to make sure that the container images I’m going to use are already pulled onto all of the nodes in my cluster.

The RAPIDS suite of open source software libraries gives you the freedom to execute end-to-end data science and analytics pipelines entirely on GPUs with minimal code changes and no new tools to learn.

This week I’ve been working on some custom Sphinx extensions for a documentation site.

Over the last few years, many open-source Python projects that I work on have switched to CalVer.

1.

1.

1.

Last week I was on holiday, so I skipped a week of newsletter, and this one is a little light.

In July I published a blog post on using Dask on KubeFlow with the Dask Kubernetes Operator.

1.

Thank you for reading my very first newsletter!

In the machine learning and MLOps world, GPUs are widely used to speed up model training and inference, but what about the other stages of the workflow like ETL pipelines or hyperparameter optimization?

When using GPUs with Kubernetes it can be important to know which driver and CUDA versions are installed on the nodes.

Kubeflow is a popular Machine Learning and MLOps platform built on Kubernetes for designing and running Machine Learning pipelines for training models and providing inference services.

A big frustration for me when reviewing Pull Requests on GitHub is coming back to a PR you’ve already reviewed to check on recent changes and be greeted with “We went looking everywhere, but couldn’t find those commits”.

When someone opens a Pull Request (PR) on your GitHub project it can be helpful for a bot to comment on the PR.

When you contribute code to collaborative projects, whether they are open source community projects or large internal projects inside organisations, the feeling of having your code running inside a large application can be very rewarding.

Kubeflow is a popular MLOps platform built on Kubernetes for designing and running Machine Learning pipelines for training models and providing inference services.

Writing GPU code in Python is easier today than ever!

Writing GPU code in Python is easier today than ever!

Kubeflow is a popular MLOps platform built on Kubernetes for designing and running Machine Learning pipelines for training models and providing inference services.

Talk abstract There are many powerful libraries in the Python ecosystem for accelerating the computation of large arrays with GPUs.

When working with Dask clusters you often need the remote worker environment to match you local environment.

Today I found myself needing a Go application’s main thread to stop and wait until the user wants it to exit with a ctrl+c keyboard interrupt.

Today is the deadline for the license changes to Docker Desktop for Mac and Windows.

In a couple of days Docker will begin charging employees of companies with >250 employees to use Docker Desktop.

This week I’ve been playing around with Kubeflow as part of a larger effort to make it simpler to use Dask and RAPIDS in MLOps workflows.

This post has been superseded with this tutorial that no longer requires any code changes.

Before I was a maintainer of open source software I was a user of open source software, and I sometimes behaved badly.