Huge thanks to Google TPU Research Cloud for providing me with access to TPU chips!

As promised, this is a follow up post with a cleaner way to track Google’s TPU performance in MLFlow while training your models.

This will spawn a background process (not thread, jaxlib didn’t like that) that will periodically send to the MLFlow Tracking server the relevant TPU information.

import multiprocessing
import mlflow

class TPUMonitoring(multiprocessing.Process):
    def __init__(self, run_id, sampling_interval_seconds=1):
        super().__init__()
        self.run_id = run_id
        self.sampling_interval = sampling_interval_seconds
        self._stop_event = multiprocessing.Event()
        self.client = mlflow.client.MlflowClient()

    def log_tpu_mlflow(self, step):
        chip_type, _, data = self.tpu_system_metrics()
        for tpu in data:
            for metric, val in

Huge thanks to Google TPU Research Cloud for providing me with access to TPU chips!

MLFlow sadly doesn’t support TPU monitoring out of the box like with NVIDIA cards and pynvml.

Luckily it was very easy to add this feature with the following snippet:

def log_tpu_mlflow(step):
    chip_type, count, data = tpu_system_metrics()
    for tpu in data:
        for metric, val in tpu.items():
            if metric != "device_id":
                mlflow.log_metric(f"system/{chip_type}/{tpu['device_id']}/{metric}", float(val), step=step)


def tpu_system_metrics():
    try:
        import tpu_info
        chip_type, count = tpu_info.device.get_local_chips()
        data = tpu_info.metrics.get_chip_usage(chip_type)
        data = [{"device_id": d.device_id, "memory_usage": d.memory_usage, "total_memory": d.total_memory, "duty_cycle_pct": d.duty_cycle_pct} for d in data]
        return

We have come a long way since Termux and proot :)

Android 15 and 16 on Pixel phones allow you to run a fully fledged virtual machine on your device. As a happy new owner of a Pixel 10 Pro I was eager to find out how well it works. The short summary is that it works very well! But…

The terminal app itself is lacking any customization, not even font size, so I decided to dig a little bit deeper into how it works. The default theme of Courier (?) font and a grey background is fine, but what if you want to change it?

Looking at the official docs it seems that the Linux Terminal App is actually a WebView that connects to ttyd service running inside the VM. That is an interesting approach, and to be honest I didn’t even notice it is a WebView.

In order to customize your session you need to edit the /etc/systemd/system/ttyd.service with your favorite editor.

You need to modify the ExecStar...

Another great Fedora release and another great Fedora Release Party! :)

After a (sadly) long break, we are back. University of Novi Sad, Faculty of Sciences was once again home for a day for all Fedora, Linux and FOSS enthusiasts. Since our last Release Party for Fedora 31 in December 2019 a lot has changed, but our message and drive did not. We are still striving to build a strong community around Fedora!

In total, we had five awesome talks. After the welcome, I was first to present what is new about Fedora 37 and what has changed since our last meet. We had a lot of new faces so a short introduction to Fedora was also in order. :)

The second talk was given by our Faculty colleague Dusan Simic, and the talk was about Toolbox and how it can be used for isolated clean development environments. A very useful talk for our students, for sure.

For the third talk we discussed about...

Everytime I get to use Red Hat’s Open Shift I think to myself: “I should use OpenShift more”. It is a really great tool not just for high availability clusters but for general software development also. Its Web interface is super intuitive and nice to use, and you can provision all sorts of stuff in minutes. And everything is open source. Its really great.

So naturally I decided to run it locally, especially after the news that starting from version 4.8 you can use it on single node architecture also. I decided to run it virtualized in libvirt so I can destroy it if I no longer needed. Hooray!

Sadly my “Hooray” was short lived, since the installation proved difficult for me.

The official guide calls for either using the Assisted installer which is only available for OpenShift, not for OKD or using the manual way which is not that difficult, but I just couldn’t get it to work. The...

Writing Python on the go is not easy or something you should do often. But if you have a long commute, why not? :)

There are several cool options already, like PyDroid3 (among others) allowing you to have a Python interpreter wherever you go. But what if you want more capabilities, like testing Jupyter notebooks or full control over your installed packages (with pip or conda). Or… What if you just want to mess around with Jupyter on your phone/tablet? Well, the answer is that it is very easy to try it out and see if you like it. Here are the steps.

1. We will need to install Termux, and you can also get it from the Play Store if you do not use F-Droid

2. We will also need Termux-Fedora (which I am very proud of, with its 150+ stars on Github). The instructions are in the readme file there. It is very simple nowadays, it only takes a couple of minutes on a relatively modern phone.

3. Once...

Here we go again! :)

After a short break we are back with another local event meant to celebrate Fedora and Free Open Source Software.

First let’s cover the usual:

• Location was, as always before, Department of Mathematics and Informatics, Faculty of Sciences, University of Novi Sad. This time we opted for newly renovated and colorful amphitheater A7 :)
• People: Momcilo (FedoraUser) and myself were “ambassadoring” the event while we welcomed local Linux and FOSS users, students and others who wanted to hear what is new in the Fedora universe

Talks

1. The first talk was about what is new in F31 (and F30 because we missed that release event). We talked about changes in GNOME, Fedora Toolbox, name changes and other stuff happening at the moment.

2. The second talk was presented by Ivan Pribela, and it was about (Un)Secure Boot, why it doesn’t work, how to make it work and why not to turn...

The new Dell XPS 7590 is a great machine. Like every great machine it has some drawbacks, but since it is much greater than the sum of its features you bond with it and forgive it.

Take mine for example. Even though the keyboard is far from perfect (with backspace and enter keys making some really weird noises) and it arrived with some scratches on the beautiful aluminium chamfered edges,
everything else about it is just… well perfect.

The 4K OLED display – perfect. The six-core powerhouse 9750H – perfect. The fast NVIDIA GTX1650 – perfect. The carbon/aluminium chassis – perfect. And the best thing about it: It runs Linux almost 100% out of the box.

Installation

Chances are that your Dell XPS will ship with Windows. If you want to dual-boot (which I recommend for BIOS updates) you will need to go through some steps.

To avoid repeating everthing, there is a great post here...

If you are not aware, PyTorch XLA project is an effort to run PyTorch on TPU (Tensor Processing Unit) architecture which offers even higher performance in training Deep Learning models compared to GPU’s.

In my previous post I showed you how to use serveo.net to SSH into your Google Colab provided container.

In this short post I will show you how to use the same principle to connect to provided free TPUs. Beware that PyTorch XLA is in its inception phase and you may run into bugs.

1. Open a Colab notebook and set it up (Runtime -> Runtime Type -> Select Python 3 and TPU)

2. Create a code cell and paste the following:

   import os
   tpu_addr = os.environ['COLAB_TPU_ADDR']
   print('Local TPU address', tpu_addr)
   import random, string, getpass
   password = ''.join(random.choice(string.ascii_letters + string.digits) for i in range(20))
   alias = ''.join(random.choice(string.ascii_letters +

   ...