How flows run
A flow is a saved definition. Each Submit starts a new run with its own id, so one run never overwrites another.
From build to run
Section titled “From build to run”- Build: stages run top to bottom. A stage waits for the stage above it. A bar to the left of several stages runs them at the same time; the next stage waits for all of them. Each stage has a machine (if it needs one), resources, a time limit, inputs and outputs, and a script or prompt.
- Preview: shows what Submit would compile, without starting a run.
- Run: projectEureka compiles the flow and starts each stage in turn.
What each stage kind runs on
Section titled “What each stage kind runs on”| Kind | What it runs | Where |
|---|---|---|
| CPU / GPU | Your script. It can also run as a small Slurm job for srun, MPI, several nodes, or an array. |
A machine of the chosen instance type, tagged for cost |
| HPC | A job submitted to an existing Slurm cluster an admin registered | The external cluster, over SSH from the login node |
| LLM | One call to a language model, using a project credential | A cloud call from the control node; no machine |
| RFN | A model call repeated until the reply passes your checks, or until the turn limit (default 5, up to 20) | A cloud call from the control node |
| ML | One managed training job: Vertex AI on Google Cloud, SageMaker on AWS. Needs a project bucket. | The cloud’s training service |
| OCP | A file copy between project files, home, a bucket, or a machine | A dedicated transfer machine, or the login node for small copies |
| CTR | A container (SIF) run as a job. It comes from imported or example flows, not the palette. | A machine of the chosen instance type |
| Quantum (Qsim) | One OpenQASM 3 circuit (OpenQASM 2, qreg/creg, and include are refused) on a simulator, or on a QPU when an admin allows it. Braket runs on AWS only. On Kubernetes, only IonQ is available, and only when an admin sets an IonQ API key (a direct IonQ Cloud key). Other quantum services are not supported yet. A QPU run needs an admin to allow QPU for the environment and choose the IonQ QPU device, and the project must allow QPU; QPU runs are capped at 1,000 shots. Braket needs a bucket whose name starts with amazon-braket-. |
IonQ (any cloud or Kubernetes, with an IonQ API key) or Amazon Braket (AWS) |
Where output goes
Section titled “Where output goes”Each run is stored in the project under prj-storage/flows/<flow>/, one folder per run, with one folder per stage:
in/: files from earlier stages (in/from-<stage id>/)out/: files this stage publishes for later stageslog/: logs, including the reason a model, quantum, or ML stage failedwork/: scratch
Follow a run in Runs and Flow monitor. See Run a flow.
Run and stage states
Section titled “Run and stage states”- A run shows Running, Stopping, Stopped, Failed, or Done.
- A stage shows Waiting, Running, Done, Failed, Cancelled (stopped while running), or Skipped (never started).
When a stage fails
Section titled “When a stage fails”Each stage has If this stage fails, fail the flow, on by default.
- On: the run is marked failed, later stages do not start, and stages still running are stopped. Earlier results are kept.
- Off: the stage records the error and later stages still run.
Machine size
Section titled “Machine size”A CPU or GPU stage with no instance type is given the smallest enabled catalog type that meets the stage’s CPUs, memory, and GPUs. On Kubernetes a 1-CPU stage with a small memory request still gets a 1-core type (such as u1.medium) instead of the 2-core default (u1.large). The Kubernetes pick also has to fit on a worker, including KubeVirt overhead. If no enabled type fits, the stage fails at once and shows the reason; the pool is not left waiting.
A blank type means auto, except for GPU containers (minGpus 1 or more), which need a chosen type or an admin or project default. A stage on a GPU container with no type and no default is refused before it starts. The message names the container, the GPUs it needs, and the GPU types the project allows.
An X or VNC desktop, including the built-in Desktop, keeps a floor of 2 cores even when a catalog row says 1. A terminal or other non-desktop app may use 1 core. The same floor applies to a launcher. A saved type under that floor is not started. The message names the type, what it has, and what the app needs.
Kubernetes lists no GPU types. On Google Cloud the smallest GPU row is g2-standard-4 (1x L4). That catalog has no T4 row, TPU rows appear in the GPU list, and those GPU prices are empty. On AWS the smallest T4 is g4dn.xlarge (4 cores, 16 GB).
Starter launchers follow those minimums. See Create a project.
On Kubernetes
Section titled “On Kubernetes”- ML examples are hidden, and ML stages do not run. There is no Vertex or Bedrock call.
- There is no spot choice. Machines are on demand.
- The flow bucket picker is hidden unless the flow already names a bucket, so an imported bucket can still be seen and cleared. Files move through the project directory.
- Quantum is IonQ only, and only the simulator, unless an admin allows a QPU and sets an IonQ API key. Braket is not offered.
- A stage that never starts shows the reason. The monitor does not show an empty provider.
- A pool the cluster cannot schedule is removed, and the run is marked failed. If the machines never become ready within the wait (one hour, unless the stage sets a shorter wait), the run is stopped and marked timed out.
On AWS and Google Cloud, a stage can run as spot. Turning Spot on in the stage’s Capacity setting requests a spot machine. A saved AWS max price also means the stage is requested as spot, and the Capacity control follows that saved choice. Switching back to On-demand clears the max price.
If Google Cloud preempts a spot VM, the VM is stopped, not deleted. It stays in Environment overview as stopped. Kubernetes has no spot choice; machines are on demand.
Time limits
Section titled “Time limits”Each stage runs under a time limit. A stage that boots a machine is stopped at the limit. Model, quantum, ML, login-node, and external-cluster stages are marked timed out, but the remote work is not cancelled. See Set time limits.
From the CLI
Section titled “From the CLI”eureka run workflow.json --project NAME submits a flow document and saves it on the project flow list. The project is --project, then EUREKA_PROJECT, then your default (eureka project default NAME or ~/.eureka/default-project), then the sole project in the environment. If none of those resolve a project, the command exits 1 with the sentence that a project is required. It does not print a server error.