Driving gipper from another repo
Writing a job that runs on any provider
- Paths: use paths relative to the job directory. The job runs in
~/jobon the machine, which is/root/jobin containers (Runpod, Vast) and/home/shadeform/jobon Shadeform. Don't hardcode/rootor/workspace;/workspaceexists only with a Runpod network volume. - GPU names: plain names in
prefer("RTX 4090", "A40") match every provider's spelling. Variants are separate cards: "RTX 4090" does not match "RTX 4090D". - Runpod-only fields:
[gpu] cloudonly affects Runpod, and[pod] volumepins a job to Runpod. - CUDA and the driver: hosts vary. On Vast a 535 driver showed up in Quebec,
and some images carry CUDA forward-compat libraries that break
cuInit(error 804). gipper checkscuIniton every host and works around 804. If the job needs a specific runtime, set[placement] min_cuda(filters Vast offers and becomes Runpod'sallowedCudaVersions) andmin_compute_capability(checked on the machine). Install torch from$GIPPER_TORCH_INDEX. - UDP:
[pod] ports = ["5004/udp"], then advertise$GIPPER_PUBLIC_IP:$GIPPER_PORT_5004_UDP. The outside port is not 5004 on Vast. - Telling gipper about the job:
[placement] compute_s,setup_download_gbanddownload_gbmake the first estimates honest; after a few runs, memory takes over for compute time. Give runs of the same work the sameworkloadso they share that memory.
The Python surface
These names are stable within 0.x:
| Name | What |
|---|---|
gipper.job.load(path) |
a Job |
gipper.run.plan(job, provider=, gpu=, throughput=) |
(offers, hourly ceiling, notes); each offer carries .estimate |
gipper.run.run(job, provider=, dry_run=, log_dir=, gpu_override=, throughput=) |
(run_id, outcome): succeeded, failed, timeout, capped, interrupted, error, termination_pending |
gipper.run.estimate(job, offer=o) |
the whole-run estimate for one offer |
gipper.run.preflight(remote) |
GPU idle check; wrap it to add your own. Raise gipper.run.Unsuitable when the host works but doesn't suit the job (wrong architecture for a pinned runtime), and gipper.run.BadHost when it's broken. Both replace the host; only BadHost counts against it in memory |
gipper.run.settings |
looked up at call time, so a wrapper can change caps for one run |
gipper.providers.load(name) |
.offers(job), .list_ours(), .get(id) (Vast), .terminate(id), .account(), .billed(since) |
gipper.providers.runpod.Runpod().api() |
the raw Runpod client |
gipper.ledger.read(since=, kinds=, job=) / .write(kind, payload, job=) |
receipts |
gipper.history.History.load() |
memory: .failure(p, machine), .ready_s, .mbps, .compute_s(workload, gpu), .billing_factor(p) |
gipper.ssh.Remote(host, port, user) |
.run, .put, .put_secrets, .get over gipper's key. .put records a custody manifest first and can raise gipper.custody.Refused |
Kept for scripts written before 0.2: gipper.runpod.api.Api and
gipper.runpod.run._wait_ready. Any other gipper.runpod.run name raises an
error pointing at gipper.run, so a script that patches the old pick_gpu fails
instead of silently patching nothing.
Receipts a guard can rely on
pod.create:provider,pod,machine,country,gpu,usd_per_h,deadlinehost.rtt: round trip fromorigin(this machine's hostname) to the rented machinepod.closure_verified: the provider no longer lists the machine. Until it appears there may be apod.termination_pending; treat that machine as still billing.spend(computed at teardown) andspend.billed(the provider's bill, fromreconcile; the latest one per machine counts).gipper.billing.spend_since(t)combines them by machine.custody.machine,custody.upload,custody.uploaded,custody.secret,custody.note: what reached a machine (custody.md).- Receipts without
providerpredate 2026-10-26 and are Runpod's.
An editable install imports code directly from its checkout. Use an isolated worktree for changes if clients are using that checkout; see AGENTS.md.