Docs / Run and ship
Runs and jobs
Running a blueprint: the run panel, placing blocks on several workers, saved profiles, tags, where a run is, data copied to the worker that needs it, the time limit, the queue, and the sessions coming next.
Work starts from a project's blueprint: press Run in the editor, choose where it runs in the run panel, and the run becomes a job in the queue, with one task per worker block (more when a block runs in parallel). Computing > Jobs lists every job of the workspace, running and finished; a project shows its own under Runs, and the Overview shows the latest with their project and flow.
The run panel
Run opens a panel from the right with the workers of the project and the blocks of the flow. Start is at the top and at the bottom; most runs need nothing else.
- Workers it may use. Tick the workers the run may use. Unticked workers take no block of this run. All also takes workers added later.
- Where each block runs. Each worker block says in one line where it will go, for example On Lab CPU, Lab GPU or Laptop, whichever is free first or On Gianluca De Novi PC, where its dataset lives. Change opens its choices: any of the ticked workers (EsaSim picks), one worker, workers with a tag, and rules (accelerator, GPU memory, RAM, sharing a worker with another block, how parallel tasks spread).
- Saved profiles. When you have some, the top of the panel offers them: a pool and its rules kept under a name (for example Home lab), for this project or for the whole workspace. Save these choices as a profile… at the bottom makes one.
- The check. At the bottom, Every block has a worker ready means the run can start at once; otherwise the panel says which block cannot run and why (no ticked worker fits, a worker is offline or busy, not enough GPU memory).
- Stop after. The run's time limit. It starts from the longest a block asks for (a training block's Minutes) with a margin, and the panel warns when it is shorter. When the time is up the run is stopped, and a block still running loses what it had not saved. Blocks can read the time left with
ctx.time_left().
Saving a profile and running with one also work from AI assistants (run_flow with a profile, list_profiles, list_workers with tags): see AI assistants (MCP).
Where a run is
While a run is shown in the blueprint, a button by Stop (and in the console, and on the View page) says where it is: On Server1 PC, On 2 workers, Waiting for a worker, and once it is over Ran on …. It opens the list of the run's workers: each with whether it is working on the run, done with its part, online or offline, and the blocks it runs or ran with their progress; the blocks still waiting for a worker are listed apart. A click on a block shows it on the grid. Each block's badge on the grid names its worker too (Running 40% · Server1 PC).
Tags
A worker's page (Computing > Local computing > the worker) has Tags: words of your own such as gpu-big, lab or robot-bridge. Rules that pick workers by tag keep working when you add or replace a worker: tag the new one and it joins.
Which worker takes each block
Each worker block says what it can run on: any worker, or only workers with a given accelerator (NVIDIA, AMD, Intel, Huawei Ascend, Tenstorrent, Apple). Some blocks narrow this by a setting: Text to video goes only to a worker with an NVIDIA GPU unless its Accelerator is set to CPU, and naming an accelerator in Text to image or Embed images sends them only to workers that have it. A task also goes only to workers the project may use (Project > Compute), to the workers the run panel allows, and to workers with a free slot.
When no worker that fits is online, the task waits in the queue rather than failing: the Blueprint's console says so before you run (can run on X, but it is offline: a run waits until one is online). A task whose worker stops reporting for three minutes goes back in the queue for another worker, then is marked lost.
Data follows the work
A block that reads a dataset kept on one worker's disk used to run on that worker only. Now, when your choices leave that worker out (it is not ticked, or a rule sends the block elsewhere), EsaSim copies the data to the worker that runs the block, up to 2 GB, and the block starts once the copy is in place. Larger datasets keep the block on the worker that holds them; the run panel says so. Datasets in cloud storage (S3, Google Drive and the others) are read from any worker.
What a project needs travels with it
Workers come and go: they only cache what they download. Everything a project needs to work (its robots, models, textures, sounds) goes in its assets (Data > Assets), which are saved with the project on GitHub. A block that runs on a worker reads them through a Project assets block: name the files or a pattern (models/*.glb) and wire it to the block's input; whichever worker runs the block downloads them from EsaSim before it starts. Datasets are the data the project works on, connected to the project from where they live (EsaSim storage, a bucket, a drive or a folder of a machine). A dataset connected from a machine's folder is there only while that machine is online: if it is offline, the run does not start and says which machine to turn on.
Your workers for a project (sessions)
Workers belong to the workspace, and a project runs only on its own workspace's workers. Inside it, each person keeps the workers they work with:
- Keeping workers. The first time you run a flow of a project, the workers you tick become yours for that project (all those you may use and that are free, when you leave them all ticked). Every later run of any flow of the project goes to them, each block to one that has what it needs; the run panel then shows Your workers for this project instead of the list, with Change workers and Release. A button by Run in the top bar says how many you keep and how long they stay.
- One person at a time. While you keep a worker, nobody else's run goes to it; others see it as In use: your session. Each person keeps their own, so nobody waits behind another person's runs. Runs and results are still the project's: everyone in it sees them.
- How long. They stay yours while the project's blueprint is open, while any run of yours goes on, and for an idle time after (1 h unless you choose 15 min to 8 h in the panel). Ten minutes before the end the panel and the button say so, with Keep. With nobody there they go after 12 h at most, never while a run of yours goes on. An admin can release anyone's workers from the worker's page.
- Who may use which worker. On a worker's page, Who may use it: everyone in the workspace, or only the people an admin ticks (and whoever paired it). The run panel shows each person only the workers they may use.
- Results stay on the workers. A run's results stay on the worker that made it: EsaSim keeps their list and fetches a file from the worker when you open or download it (worker 0.13 and later send nothing else). In the project's Settings › Results and cache you choose whether results move to GitHub when a run ends (
results/run-<number>/in the project's repository), whether you are asked to download them, and when the project's files are cleared from the workers: after each run (once its results went to GitHub and the download question was answered), when the project is closed, or by hand. Compute shows each worker's hardware and what it holds, with Wipe cache for anything left. The software they installed stays, so the next run is ready at once. - Your hardware for someone else's project: copy the project into your workspace and send the results back; keys and data connections stay with its owner (see Projects).
Preset jobs
Before blueprints, jobs started from presets (GPU benchmark, Isaac Sim scenario sweep, Isaac Lab RL training, MuJoCo simulation, Data curation). New work starts from a blueprint, where the same workloads are blocks, so Computing > Jobs no longer has a New job button. Preset jobs already run stay in the list and can be rerun.
Queue and statuses
New jobs enter the queue. The queue assigns a job when a matching worker or unit is free.
| Status | Meaning |
|---|---|
queued | Waiting for a free worker or unit that matches the target and resources. |
assigned | A worker accepted the job. |
starting | The worker prepares the environment, for example pulling a container image. |
running | The workload is executing. |
succeeded | The workload finished without error. |
failed | The workload exited with an error or exceeded its maximum runtime. |
canceled | A user canceled the job. |
lost | The worker stopped responding while the job was assigned or running. |
A lost job usually means the worker slept, rebooted or lost its network connection. Check the worker, then rerun the job.
Cancel and run again
- Cancel stops a queued or running job. Its status becomes
canceled. - To run a flow again, press Run in the editor: the run panel offers the placement of the last run, or a saved profile.
- Rerun, on a preset job's page, launches a new job with exactly the same settings.
A new run is a new job. The earlier job and its outputs stay unchanged.
Limits
From your plan
- Free: your workers only, up to 2 workers, 1 GPU per job, 3 projects, no replicas.
- Pro and Enterprise: your workers and shared GPU servers, multi-GPU compute units, replicas, unlimited projects.
From organization admins
In an organization, owners and admins can set per-member feature permissions and GPU limits. For example, a member may be allowed to use their own workers only, or to use at most 2 GPUs per job. Runs use only the workers you are allowed to use. See Organizations and roles.