🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

🚀 New in dstack 0.19.9 *CPU, memory and GPU metrics are now retained and visible after runs complete. *The CLI now displays container exit codes for failed jobs, making it easier to understand why runs failed. Release notes: github.com/dstackai/dstac…

Running @NVIDIA or AMD clusters for distributed training? We’ve put together a brief guide on setting up fast interconnects with dstack – whether you’re scaling AI workloads in the cloud or on-prem. dstack.ai/docs/guides/cl…

ROCm experience Part 2. Tested verl project on AMD: * AMD example uses Ubuntu 20, but most clouds run Ubuntu 22—glibc errors with Broadcom RoCE * Broadcom NICs untested—only works with Mellanox. Building custom Ubuntu 22 + ROCm 6.2 + vLLM 0.6.3 didn’t fix RDMA * AMD example

Anush Elangovan We plan to support the default AMD Docker image with dstack so users can simply specify ROCm version: github.com/dstackai/dstac… The problem is most AI frameworks today do not simply work require and use their own custom images very custom ones, incl. verl project,

🚀 New in dstack 0.19.10: smarter task scheduling. Run configs now support a priority field to control scheduling order. If you're still using Slurm for ML development, time to try dstack. github.com/dstackai/dstac…

Announcing a new example on how to use RAGEN on dstack to train reasoning agents. Under the hood the example uses verl project and ray for running Reinforcement Learning across multiple nodes. Huge thanks to Lambda for providing access to a 16xH100 1CC cluster.

At NVIDIA GTC, Electronic Arts showed how their teams use dstack to spin up GPUs and orchestrate training and development at scale. 👉 Read the full case study + slides: dstack.ai/blog/ea-gtc25/

Wonder how dstack can accelerate your ML team and cut GPU costs? Check out our latest case study with Electronic Arts, presented NVIDIA GTC 👇

The verl project docs now include a guide on using dstack for distributed RL training across multiple nodes, without needing Kubernetes or Slurm. More on docs: verl.readthedocs.io/en/latest/inde…

RAGEN is a new RL framework for training reasoning agents — and now dstack is featured in their repo as a way to run multi-node training (as an alternative to K8S/Slurm). Check it out: github.com/RAGEN-AI/RAGEN

🚀 v0.19.11 is out! If you don’t specify a Docker image, dstack now lets you use `uv` (in addition to `pip`). And it’s time to say goodbye to `conda`! `uv` is not only more convenient — it’s also a lot faster. Read more: github.com/dstackai/dstac…

I've been a long time fan of conda, but I think its time to go! Would love to see more ML projects to switch to uv! Kudos to Astral and Charlie Marsh for amazing project!

Incredible work. Open-source wins. Congrats DeepSeek!

We’ve just published a new example of using dstack with Hugging Face TRL to train across multiple nodes! It lets you quickly run training on any cloud GPU or on-prem cluster—no K8s or Slurm needed. Check it out: dstack.ai/examples/distr…