dstack

Why we included this project

dstack is worth a look for anyone juggling multiple GPU providers or mixing rented capacity with on-prem hardware. Instead of hand-rolling per-cloud scripts, you describe a training run, dev environment, or inference workload in a small YAML file, and dstack provisions the matching instance, runs the job, and cleans up afterward. Because the orchestration layer sits above the hardware, the same config can target NVIDIA, AMD, TPU, or Tenstorrent machines without rewiring anything. It also handles distributed training with torchrun and Ray across multi-node fleets, which removes a chunk of the infrastructure glue small teams usually end up writing themselves.