manage-prime-gpu-pods

name: manage-prime-gpu-pods description: Manage Prime Intellect on-demand GPU pods using only the prime CLI. Use when work requires discovering non-spot GPU availability, selecting the cheapest matching configuration, creating pods, checking pod status, connecting with SSH, terminating pods, or auditing pod history from terminal commands.

Manage Prime Gpu Pods

Overview

Manage Prime Intellect GPU pods end-to-end with CLI-only workflows. Prefer automation-safe commands with --output json and use the bundled script to filter and select on-demand (non-spot) GPU configurations.

Workflow

1. Verify CLI access and authentication.

prime --version
prime config view
# Authenticate if needed
prime login --headless
# Or configure key directly
prime config set-api-key

2. Discover matching GPU inventory in JSON.

prime availability list \
  --gpu-type H100_80GB \
  --gpu-count 1 \
  --no-group-similar \
  --output json

3. Select on-demand configurations and generate a create command.

prime availability list \
  --gpu-type H100_80GB \
  --gpu-count 1 \
  --no-group-similar \
  --output json \
| python "$CODEX_HOME/skills/manage-prime-gpu-pods/scripts/select_on_demand_gpu.py" \
    --top 5 \
    --output table

prime availability list \
  --gpu-type H100_80GB \
  --gpu-count 1 \
  --no-group-similar \
  --output json \
| python "$CODEX_HOME/skills/manage-prime-gpu-pods/scripts/select_on_demand_gpu.py" \
    --output create-cmd

4. Create the pod from a selected ID.

prime pods create --id <ID> --name <pod-name> --yes

If disk, vcpus, memory, or image is not provided, the CLI prompts interactively for missing values.

5. Monitor readiness and connect.

prime pods list
prime pods status <pod-id>
prime pods ssh <pod-id>

6. Terminate and review history.

prime pods terminate <pod-id> --yes
prime pods history --output json

Execution Rules

• Use prime availability list --output json for machine parsing instead of table output.
• Use --no-group-similar when selecting a concrete purchasable configuration.
• Treat on-demand as is_spot == false.
• Sort matching on-demand candidates by price_value ascending when user asks for cheapest.
• Use prime pods create --id <ID> as the canonical launch path from availability results.
• Use prime pods status <pod-id> --output json for automation polling.
• Use prime pods terminate <pod-id> --yes for deterministic cleanup.
• Respect team context:
  • Persisted team: prime config set-team-id <team-id>
  • Per-command override: prime pods create --team-id <team-id> ...

References

• Read references/prime-cli-command-map.md for a compact command/flag inventory derived from this repository.
• Run scripts/select_on_demand_gpu.py to filter and rank non-spot configurations from prime availability list --output json.