skills.rest
pollinations
pollinationsOFFICIAL
Explore all skills by this author →
Website𝕏@pollinations_aiGitHub

manage-vast-gpu-fleet

Inspect, price, canary, and replace Vast.ai GPU workers through an approval-gated fleet workflow.

5.0k|1.0k|Updated Apr 15, 2021
Search Tags:#cost optimization#canary deployment#Infrastructure#Vast.ai#devOps#gpu fleet#instance replacement
One-click install
npx skills add https://github.com/pollinations/pollinations --skill manage-vast-gpu-fleet
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: manage-vast-gpu-fleet
Source: https://github.com/pollinations/pollinations/tree/main/.claude/skills/manage-vast-gpu-fleet
Command: npx skills add https://github.com/pollinations/pollinations --skill manage-vast-gpu-fleet

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Managing a fleet of rented Vast.ai GPU instances manually risks overspending on idle machines, breaking production during replacements, and losing track of what is actually deployed. This Skill provides one repository-owned workflow to audit fleet costs, qualify cheaper offers, test isolated replacement candidates, and cut over to production only after explicit human approval.

Core Features & Use Cases

  • Fleet Auditing: Rebuilds live state from Vast instances, utilization metrics, and production attribution, then reconciles it against the GPU_INSTANCES.md inventory to flag stale docs, idle paid resources, and unregistered workers.
  • Offer Qualification & Canary Preparation: Scores marketplace offers against savings, reliability, hardware, and spend policies, then rents and validates isolated candidates with per-target leases, TTLs, and spend ceilings.
  • Approval-Gated Promotion: Stops at READY FOR APPROVAL with full cost and test evidence, then performs serialized cutovers, destroys replaced instances, and opens a documentation PR.
  • Use Case: A scheduled daily task runs in Audit mode to reconcile fleet economics, while a separate offer scout in Prepare mode rents a cheaper RTX 4090 candidate, runs the model's verification script, and waits for a human to approve the production cutover.

Quick Start

Ask the agent to read .claude/skills/manage-vast-gpu-fleet/SKILL.md and run an Audit of the current Vast GPU fleet, reporting costs, utilization, and any stale documentation without changing anything.

Frequently Asked Questions about manage-vast-gpu-fleet

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I safely replace a Vast.ai GPU instance in production?▼

Use the three-mode workflow: Audit the fleet, Prepare an isolated canary candidate that passes all verification tests, then Promote only after explicit human approval naming the workload and instance. The old instance is destroyed only after the new one confirms real production traffic.

How do I audit GPU fleet costs and utilization on Vast.ai?▼

Run the skill in Audit mode, which lists every Vast instance with its rate, reliability, and utilization, maps instances to workloads, measures demand and latency, and reconciles findings against GPU_INSTANCES.md without changing any external state.

Can a scheduled task automatically promote a cheaper GPU candidate?▼

No. Scheduled tasks may only run in Audit or Prepare mode, and Prepare requires explicit spend authorization. Promotion always requires a separate human approval naming the exact candidate, workload, and target production instance.

What happens when two automation runs target the same GPU slot?▼

An atomic lease keyed by the exact production instance prevents duplicates. A second run can prepare candidates for other unlocked targets but can never create a second candidate for an already locked production slot.

What happens if a canary candidate fails its verification tests?▼

The failed candidate is destroyed, its compute and storage billing is verified as ended, and only its target lease is released. Production and other canaries remain untouched, and transient host failures are recorded with a time-bounded cooldown.

Related Skills

Explore More Software Engineering Skills→
raja21068raja21068
2

vast-gpu

Finds and provisions cost-optimized vast.ai GPU instances for AI training workloads.

Community
Advanced
rogsmerogsme

gsd-update-reviewers

Manage reviewer model entries in shared registries across AI review systems.

Community
Intermediate
vorynkavitaliyvorynkavitaliy

grafana-dashboards

Create production-grade Grafana dashboards for system and application metrics.

Community
Intermediate
BagelHoleBagelHole
46

gpu-kubernetes-operations

Manage GPU-backed Kubernetes clusters for AI inference and training.

Community
Advanced