cheptsov

A nice idea! Is there a blog post with more examples and details on how it works? Would be great to read.

Environment variables are especially convenient IMO for parameterizing Docker images when running containers, e.g. using Docker or Kubernetes, etc.

Also, for configuring all sort of credentials for libraries, e.g. AWS, WandB, HuggingFace, etc.

Sorry if the question is stupid but maybe someone can help understand it better. I’m using direnv (1). I like the tool for its simplicity. How are dotenv or dotenvx different?

(1) https://direnv.net/

Right now, mostly VSCode - mainly because it’s a desktop IDE and it also supports notebooks.

That will be interesting. Happy to test and share feedback. Previously, I was a part of the DataSpell and PyCharm team building notebook support. Now working with dstack, where we support dev environments and super interested in remote support.

Can I use it to connect to a remote Jupyter notebook server?

I'd love to see more individual researchers openly exploring AI safety from a scientific and humanitarian perspective, rather than just the technical or commercial angles.

With dstack you can either utilize multiple affordable cloud GPU providers at once to get the cheapest GPU offer or also use an own cluster of on-prem servers. Dstack supports both altogether. Disclaimer: I’m a core contributor to dstack

A good clickbait title. One should credit the author for that.

As to the topic, IMO, there is a contradiction. The only way to handle big data is to divide it into chunks that aren’t expensive to query. In that sense, no data is "big" as long as it’s handled properly.

Also, about big data being only a problem for 1 percent of companies: it's a ridiculous argument implying that big data was supposed to be a problem for everyone.

I personally don’t see the point behind the article, with all due respect to the author.

I also see many awk experts here who have never been in charge of building enterprise data pipelines.

We build a similar tool except we focus on AI workloads. Also support on-prem clusters now in addition to GPU clouds. https://github.com/dstackai/dstack

I‘m a happy user of direnv. Hard to imagine my life without it. The only problem is not to forget to include it to .gitignore.

Always wondered why not spend the same amount of effort on talking to users, and improving the product, finding early users?

Are there any reads on whether it’s worth at all to get in to YC these days?

Certainly interested in having more solutions helping use SQLite. At dstack.ai, we use SQLite and love it.

- For our hosted version, we use Litestream; we lack a UI for accessing data.

Not exactly this, but something related. At https://github.com/dstackai/dstack, we build an alternative to K8S for AI infra.

Why not just show it per layer and folder via plain text?

Would be great if you could share TL;DR with how it works and other details

You can use https://github.com/dstackai/dstack to deploy your model to the most affordable GPU clouds. It supports auto-scaling and other features.

Disclaimer: I’m the creator of dstack.

Interesting project! Many people face the problem and look up for general advice. Copilot seems to be the best interface, in my opinion, to get started investing and gradually dive deeper into it.

Thank you for sharing the announcement here! I'm the founder of dstack.

To share more context:

dstack is an open-source tool designed for managing AI infrastructure across various cloud platforms. It's lighter and more specifically geared towards AI tasks compared to Kubernetes. Due to its support for multiple cloud providers, dstack is frequently used to access on-demand and spot GPUs across multiple clouds.

To democratize access to GPUs, and to streamline the process of managing multiple clouds, we introduce dstack Sky, a managed service that enables users to access GPUs from multiple providers through dstack – without needing an account in each cloud provider.

We launched dstack Sky today on Product Hunt: https://www.producthunt.com/posts/dstack-sky

Happy to hear feedback, and answer questions!

Same question here. Anyone can point to the source that says they are going to publish the weights?

Sounds promising! Gonna check it out.

I have a lot of respect for Erik and his work with Modal, which I've heard a lot of good feedback about. What Erik says about serverless and code over configuration can benefit many users and companies. However, I strongly disagree on the main points and certainly have a different wishlist for infrastructure. My main point would be on that list – open-source and vendor-agnosticism.

Finally, I believe simple configuration can coexist with code.

P.S.: At dstack, we are building an open-source platform to manage AI infra – a more lightweight and AI-friendly alternative to Kubernetes.

Any chance you plan to offer the API to cloud LPUs? And not just the LLM API? It would be cool run custom code (training, serving, etc).

Great job on the project! It looks fantastic. Thanks to your post, I discovered Fly's GPUs. We are currently developing https://github.com/dstackai/dstack to enable users to run any model on any cloud. I am curious if it would be possible to add support for Fly.io as well. If you are interested in collaborating on this, please let me know!

Congrats on the launch! Excited to hear that you go fully open-source with this! There is a certain need for great visualization tools that enable building apps.

To be frank, I’m not a fan of Kubernetes as long as GPU is concerned. Kubernetes has a very large legacy. Why we think dstack can do better: 1. Lightweight-ness - that it very easy to integrate dstack with modern cloud GPU providers. 2. AI-friendly interface built-in - no need to things like KubeFlow and alike.

FTR, we integrate dstack with Kubernetes too - already [1] But our native cloud integrations can be a lot more efficient. For example, when it comes to auto-scaling - this part is in work

1. https://dstack.ai/changelog/0.15.1/

Vast.ai is supported. RunPod not yet but on our list! The full list of all supported targets can be found at https://dstack.ai/docs/installation/

Yup, that’s what we are building now. Tasks will allow to specify how many nodes you need, and dstack will automatically provision a cluster or will run it using the one that is already provisioned. Please feel free to ping me on our Diacord if you’d like to test an early version!

This is from https://dstack.ai/blog/2023/06/29/say-goodbye-to-managed-not... Yup, that's why we support dev environments as first-class citizens. Now, also support tasks (for fine-tuning, other batch jobs) and services (deployment, incl. OpenAI compatibility for LLMs).

Thank you! Yes, that's precisely our intent. Most AI platforms offer nice yet proprietary solutions for training and deployment. We wish there were a simple and open standard that any provider could support.