remote-gpu-trainer v1.0.0 — first tagged release
An Agent Skill for the full DL experiment lifecycle on a GPU you own or rent: RUN → VERIFY → DELIVER.
First tagged, stable release. The complete RUN→VERIFY→DELIVER reorganization landed in 2026-06 and has been hardened since with verified teardown/billing rules and event-driven monitoring fixes.
What it covers
- RUN — local and rented GPUs (AutoDL, RunPod, vast.ai, Lambda, Paperspace, Chinese platforms, bare SSH, Slurm, K8s): SSH transport, detached & resumable launch, spot-preemption resilience, checkpoint→durable + idempotent resume, disk / inode / OOM survival, durable monitoring.
- VERIFY — is the number real: seed / determinism, delta-vs-noise, collapse / smoke artifacts, data leakage, proxy agreement, confound control.
- DELIVER — reproducible single-source figures & tables, a provenance manifest,
EVIDENCE.json, one-command repro.
Highlights
- 7 platform profiles + a universal gotcha catalog (U1–U43), each written Symptom → Root cause → Fix.
- Host-portable monitoring & scheduling across Claude Code, Codex, Cursor, Trae, Antigravity, Gemini CLI, and a generic fallback — an event-driven background watcher is the primary wake signal; timer wake-ups are only a heartbeat fallback.
- Structural eval suite:
python evals/run_evals.py→ 15/15 reachability cases green.
Use
No build step, no dependencies to install. Point your Agent-Skills host at the repo (entry point: SKILL.md), or install through your agent's skill mechanism.
MIT licensed.