Ai

Daytona 0.181.0 Delivers Per-Sandbox GPU Assignment on Multi-GPU Runners

Daytona 0.181.0 Delivers Per-Sandbox GPU Assignment on Multi-GPU Runners

Daytona's newest release, version 0.181.0, brings a long-requested capability: assigning individual physical GPUs to each sandbox when running on multi-GPU infrastructure. It's a small change in the codebase but a big leap for teams building GPU-heavy AI workloads.

What Changed

Before this update, if a runner had multiple GPUs, they were treated as a single pool. Sandboxes could claim any GPU, leading to contention and inefficient scheduling. Now, the API and runner coordinate to lock a specific physical GPU to each sandbox. The pull request (#4792) touches both the API and runner components, implementing a per-sandbox GPU assignment strategy. It's a targeted change: no new configuration flags, no breaking API shifts. The runner simply asks the API which GPU to use for each sandbox.

Alongside this feature, the team bumped the Go toolchain to 1.25.7 across runner and proxy components, and patched OpenSSL and PCRE2 dependencies. These are housekeeping moves, but they keep the platform secure and performant. The SDK-Go was also updated to v0.181.0.

Why It Matters

For teams running multiple development or test sandboxes on a shared GPU server, this is a game-changer. Imagine a machine with four NVIDIA A100s. Earlier, two sandboxes could accidentally trample each other's memory by both grabbing the same GPU. Now each sandbox gets its own dedicated GPU, guaranteeing isolation. That means no more mysterious OOM kills during model training. Performance becomes predictable. Debugging becomes saner.

It's also a sign of maturation for Daytona. The project started as a simple dev environment manager. Now it's handling real hardware constraints. I've seen similar features in enterprise platforms like Gitpod and GitHub Codespaces, but Daytona's open-source nature means any team can inspect and trust the allocation logic.

One thing I'd love to see next: dynamic GPU sharing for workloads that don't need a full card. But that's a problem for another release. For now, this is a solid step.

Official Source: https://github.com/daytonaio/daytona/releases/tag/v0.181.0