MDEngine

Balance in USD (`balance_usd`), the part of it held by your unfinished jobs (`reserved_usd`, `held_jobs`) and what a new submit can use (`available_usd` = balance - reserved; a submit needs wall_limit_s x rate of it), how jobs are priced (`pricing.mode` job = the deck's own work in atom-steps at per-class prices, capped at wall_limit_s x rate; metered = per second of pod time), the rate table, the key id of the API key in use, and `packs` (when the endpoint sells credit): [{id, name, usd, gpu_h, url}] to top up THIS key by appending ?client_reference_id=<key_id> to the pack url.

No key needed. Sends a study request to a person and returns a request_id, the tier that fits, the price band, what we still need from you, the next step and the claim scope. Nothing is run or charged. Call it after reading a guide.

Cancel a job that is not finished. A running job is billed up to the cancel time; its pod is terminated.

Capability manifest of the hosted runners: LAMMPS version, installed packages, and every style by category with gpu=true (KOKKOS-accelerated) or gpu=false (exists, but runs on the pod's CPU cores at the GPU rate). Default = compact summary; runner=lammps&full=true returns the whole style table. Use preflight_deck to check a specific deck.

Step 1 of the two-step path for big decks: validates the spec, reserves a job id, returns a presigned upload_url. PUT the deck as a .tar.gz (<= 2 GB, relative paths, input at `input`) to upload_url, then call start_job.

For a finished job: delete its deck and results tarballs immediately instead of at the automatic 30-day purge. Metadata and billing records are kept; results can no longer be downloaded.

No key needed. Without customer_type: the list of guides (one line each), the start page and how to get a key. With customer_type (mineral-processing | dark-proteins | alloys): the full guide, which says what to supply, what runs, what comes back and what it costs. With template (openmm-stability): the script text of that template, ready to save and submit.

The last <= 20 thermo/log lines the running pod reported (30 s heartbeat). Full log.lammps is in the results tarball.

For a done/failed job: a presigned download_url (valid ~7 days) for the results tarball (work/, log.lammps, exitcode). Results are deleted 30 days after the run.

State (created|uploaded|queued|launching|running|uploading|done|failed|cancelled), GPU, rate, billed seconds, cost so far, exit code, error, last thermo lines, and the post-run verdict: clean | warnings (a done run whose log shows QEq CG failures, lost atoms, a fix halt, or no end_marker line; counts in `warnings`).

Jobs of this API key, newest first. Compact rows (id, state, label, created, finished, gpu, cost_usd, error) unless full=true.

Dry run of the check submit_job performs: which styles the deck asks for are MISSING on every hosted LAMMPS image (the run would exit at startup), which are CPU-only, whether its pair style will use the GPU at all, and which image (`runner`) the job will be routed to — decks needing packages beyond the fast default image run on lammps-full automatically. Nothing is created or billed.

Step 2: queue a job whose deck tarball has been uploaded. A GPU pod is launched; billing starts when it reports running.

One call: create a hosted GPU job, upload the deck given INLINE as {relative_path: text}, and queue it. Total inline size <= 8 MB; for larger decks use create_job, PUT the tarball to upload_url, then start_job. Billing starts at the first heartbeat (state running) and stops at done/failed/cancelled. The deck is executed as a program on an isolated GPU pod.