swm pod

Full lifecycle management for GPU pod instances.

Subcommands

Command	Description
pod create	Provision a new GPU instance
pod list	List instances across all providers
pod status <id>	Detailed status of one instance
pod start <id>	Start a stopped instance
pod stop <id>	Stop (pause billing, preserve volume)
pod terminate <id>	Permanently destroy instance
pod down <id>	Push workspace + terminate
pod prune	Remove stale config entries

pod create

swm pod create -p runpod -g "H100 SXM" -n my-project \
  --cuda 12.8 --gpu-count 4 --volume 200 \
  --lifecycle auto-down --idle-timeout 30 -y

Option	Description
-p, --provider	Provider (required)
-g, --gpu	GPU type
-n, --name	Pod name (required)
-w, --workspace	Workspace to restore from storage
-b, --bucket	Storage bucket (provider:bucket)
--no-storage	Skip storage configuration entirely
--volume	Persistent volume size in GB
--disk	Container disk size in GB
--gpu-count	Number of GPUs
--image	Docker image (provider default if empty)
--cuda X.Y	Auto-pick the newest provider image matching this CUDA major.minor (e.g. 12.8). Ignored if --image is set.
--cloud-type	RunPod cloud type: SECURE, COMMUNITY, ALL
--ports	Ports to expose (default: 22/tcp,8888/http,8188/http)
--region	Datacenter/region ID
--lifecycle	Guard mode: auto-down, auto-stop, remind, manual
--idle-timeout N	Idle timeout in minutes
-x, --exclude	Glob pattern to exclude from pull (repeatable)
-y	Skip confirmation prompt

If a workspace is configured, pod create runs the full bootstrap over SSH:

1. Install s5cmd and configure storage
2. Pull the workspace (or initialize empty for a new one)
3. Start the inotify watcher
4. Start the auto-sync daemon (default interval 60s)

If any step fails, swm persists the pod ↔ workspace mapping unconditionally and prints exactly the commands you need to retry the missing steps:

⚠ Bootstrap incomplete. Re-run the remaining steps when ready:
  Storage configuration:  swm setup storage runpod:abc123
  Workspace pull:         swm sync pull runpod:abc123
  Auto-sync start:        swm sync auto runpod:abc123

The --cuda flag complements the Min CUDA column in swm gpus. swm cross-checks the resolved image against the GPU’s minimum CUDA and warns if the image is too old for the hardware.

pod down

Push workspace to storage and terminate in one command:

swm pod down my-project
swm pod down my-project --no-sync       # terminate without pushing