lvrged-factory

GPU infrastructure control for Pi — v2. Install one package and your agent can provision, deploy, run, monitor, pause, and destroy GPU video workloads on RunPod, with MiniMax H3 on an RTX PRO 6000 as the zero-friction default lane: Turbo LoRA v4 (8 steps, euler/beta) + SageAttention2 + INT8 on the cu130 image. Other providers plug in via a table entry + skill recipe; no release needed.

pi install npm:@lvrged/lvrged-factory

Then, inside Pi:

Set me up to make H3 videos.

The agent runs the onboarding skill, asks exactly two questions (batch size, budget), and everything else is preset — GPU, image, disk, ports, workflow stack, capacity fallback. Every default was verified in a live RunPod session on 2026-08-12.

The lane (what you get without choosing anything)

| Layer | Preset | | --- | --- | | Provider / GPU | RunPod · RTX PRO 6000 96GB (--gpu-id "NVIDIA RTX PRO 6000 Blackwell Server Edition") | | Image | runpod/comfyui:cuda13.0 — official cu130 build, live-verified layout; real SageAttention2 built at install (cu128 runs H3 INT8 ~2x slower) | | Disk | 80GB container disk, no network volume (volumes DC-lock the install) | | Weights | Comfy-Org/MiniMax-H3 pruned INT8 set + larryvrh Turbo LoRA v4 step600 EMA (~41GB, public) | | Workflow | Turbo 8 steps · euler (MiniMaxH3TurboSampler, video shift 12 / audio shift 3) · beta scheduler · Sage2 ON | | Capacity | provision tool tries COMMUNITY across the DC rotation, then SECURE, shrinking disk when a machine doesn't fit — expect secure $2.09/hr in practice | | Cost | ~$0.25 per finished 480p minute at the 8-step anchor; ~$0.70 of each fresh pod is startup |

What you get

• lvrged_factory_* tools — the state machine: lvrged_factory_setup, eight resource-oriented tools: lvrged_factory_setup, _status, _ensure (the idempotent entry point), _pod (provision — capacity fallback built in — register, pause, resume, destroy), _job (run, progress, watch — background queue watcher with done/stuck notifications — finish), _ledger (job queries + cost rollups), _manifest (model/workflow registry), _policy (the spend trust boundary, kept deliberately separate).
• /lvrged-factory commands — the human dashboard (gpu status, gpu onboard, gpu jobs, gpu spend, ...).
• scripts/ — the golden path, shared by Pi and Claude Code users:

  POD=$(scripts/h3-pod-up.sh)     # provision with capacity fallback
  scripts/h3-install.sh "$POD"    # nodes + weights + sage + pod restart + verify
  scripts/h3-run.py --prompt "..."   # generate; timings recorded
  scripts/h3-benchmark.sh         # benchmark protocol v1 (cold + 5 warm, median)

• Skills — the knowledge, with every hard-won ops lesson folded in: the comfyui_args.txt + pod-restart convention (never hand-roll daemons), the SageAttention stub trap, the model path-prefix HTTP-400, the SSH tunnel fallback, capacity-error decoding, pause/resume semantics.
• Docs — docs/h3-economics.md (including benchmark protocol v1), per-provider playbooks, pricing snapshots, architecture.
• A persistent registry + job ledger in <project>/.pi/lvrged-factory/gpu/ — the agent restarts tomorrow and still knows what's deployed, what it costs per minute, and which workflow version made which video.

Using the same skills with Claude Code (no Pi)

claude-skills/ ships the same knowledge adapted to bash-only operation (no lvrged_factory_* tools), sharing the same JSON registry and the same scripts/:

cp -r claude-skills/* /path/to/your/project/.claude/skills/

Architecture

See docs/ARCHITECTURE.md. The rules that matter: the core abstraction is the deployment; the ledger is the source of truth; the extension is dumb but strict (state, spend policy, deterministic command templates — no reasoning); the agent executes, guided by skills; the golden path is scripted.

                Pi Agent  ──────────  Claude Code
                    │                      │
          lvrged-factory extension      claude-skills (bash)
                    │                      │
             provider capability        scripts/
                  tables                   │
                    └────────┬─────────────┘
                          RunPod
                    (more via adapters)

The spend policy (read this)

Extensions run arbitrary code — this one executes provider CLIs and spends real money. The trust boundary is .pi/lvrged-factory/gpu/policy.json:

{ "ceiling_per_job_usd": 5, "ceiling_daily_usd": 40, "ceiling_monthly_usd": 400, "confirm_above_usd": 1, "idle_shutdown_after_min": 30 }

Provisioning above confirm_above_usd always asks you in the UI (quoting the secure rate, since community stock-outs are the norm), and ceilings cap what the agent may do without checking in. The agent cannot weaken these silently. Review this package's source before installing — it's yours.

Also worth knowing before you spend: pods bill from creation (a stuck image pull still costs); a $0 RunPod balance kills pods and deletes disks — keep a $5–10 buffer; and the H3 license excludes US/EU/UK/South Korea — your read.

Development

npm install          # dev deps (typescript, pi types)
npx tsc --noEmit     # typecheck
bash -n scripts/*.sh && python3 -m py_compile scripts/h3-run.py
pi -e .              # load the extension from this dir in a scratch session

License

MIT (the package — the H3 model weights carry their own license; see above)