Overview
Connect to your control plane
Paste an access token. Running locally? It is in local/credentials.env (GRYVIA_ADMIN_TOKEN). On the Mac mini: ~/gryvia/state/credentials.env. Just looking around? Press Try demo.
What Gryvia does
You have one or more Apple Silicon Macs. Gryvia is a queue and a traffic controller for them: you hand it AI jobs (run a model, make embeddings, evaluate…), and it decides which Mac runs each one, when, without overloading memory. You watch and manage everything from this dashboard, the Mac app, or the command line.
Your first job in 3 minutes
- Start the server.
make run-server(or it is already running on your Mac mini). - Start a worker.
make run-agentturns this Mac into a worker. For another Mac runmake deploy MINI=hostname. - Open this page and connect with the admin token from
local/credentials.env, or press Try demo to explore with simulated data. - Run a job. Press Run demo job on Workloads. It finishes in about 3 seconds.
- Check the result. Click Logs on the job. Press ⌘K any time to search commands.
Words you will see
- Mac / node
- One machine running the Gryvia agent. Shown in Mac fleet.
- Workload / job
- One task to run: a runtime (mlx, llama, coreml, demo), a model label, and arguments.
- Reservation
- Memory (GiB) a job sets aside on a Mac. It is bookkeeping for fairness, not a live GPU meter.
- Queued → running → succeeded
- Life of a job. Failed and cancelled are the other endings.
- Tenant / quota
- A team and its memory allowance, so one team cannot starve the others.
- Drain
- Stop sending new jobs to a Mac (for maintenance). Running jobs finish. Resume undoes it.
- Priority
- 0–100. Higher goes first when jobs are waiting.
Common tasks
- Run something
- Workloads → Submit, or
gryvia.py submit job.json --wait. - Repeat or tweak a job
- Clone fills the form from an old job; Retry resubmits it.
- Save a favourite setup
- Save as template on the submit form.
- Something is stuck
- Check Alerts. The job's Placement column says why it is waiting.
- Maintenance on a Mac
- Mac fleet → Drain node.
- See it in Grafana
- Scrape
/metricswith the admin token. - Is everything OK?
gryvia.py doctor
Command line cheat sheet
export GRYVIA_ADMIN_TOKEN=… # from credentials.env gryvia.py doctor # is everything healthy? gryvia.py top # live view of Macs and jobs gryvia.py submit job.json --wait # run and wait for the result gryvia.py jobs --state failed # find problems gryvia.py logs JOB_ID -f # follow a job's output gryvia.py drain MAC_ID # stop new work on a Mac
Welcome to Gryvia: three steps to your first job
- Connect a Mac. A Mac joins by running the Gryvia agent. Locally:
make run-agent. On another Mac:make deploy. It appears in Mac fleet within seconds. - Run a job. No setup needed: sends a 3-second test workload to your fleet.
- Watch it finish. Open Workloads to see it go queued → running → succeeded, then click Logs.
Scheduling, live
queue → admission → MacsScheduler decisions
How to read this
Each Apple Silicon Mac has one GPU made of many cores. The grid lights cores in proportion to the GPU's measured busy percentage.
Utilisation is the whole Mac's GPU load (it includes other apps). Job reservations are the memory each job set aside for admission; the marker shows what the GPU is actually using.
Fleet capacity
Queue health
Reservations are admission bookkeeping. macOS controls hardware GPU scheduling.
Trend
Recent workloads
Mac fleet
Drain nodes safely before maintenanceWorkloads
Submit native workload
Timeline
Last 30 minutes · reservation windows per MacModel hub
Bring a model onto your Macs, check it fits, benchmark it, and test it against your own prompts.
Register an existing model
Playground
Chat with a model that has an inference replica: deploy one from Deployments, or register origins in the Model hub. Replies stream as they are generated.
No reply yet.
Datasets
Data you train and evaluate with. Files live on a Mac under its managed models folder; Gryvia records the size, row count and SHA-256.
Add from Hugging Face
Register a file already on a Mac
Training
LoRA fine-tuning on one Mac with MLXNew fine-tune
Training uses this Mac's single GPU; the run appears in Live and Workloads like any job. Loss charts update as MLX reports them.
Runs
Sweep
Loss
Workflows
Chain steps: a step starts only after the ones it depends on succeedNew pipeline
Deployments
Models served by llama-server on a Mac (loopback)A deployed model is reachable by the control plane on that Mac and works in the Playground. Remote access needs an ssh tunnel; it is not exposed to your network.
Tenants & quotas
Reservation metering
Worker time and reserved GiB-seconds are allocation estimates, not measured GPU utilization or invoices.
Set tenant policy
Alerts
Activity trail
Settings
Appearance
Accent
About
Gryvia Apple 0.3.0 · native Apple Silicon workload scheduling.
Licensed under the Business Source License 1.1. Copyright © 2026 Zyvor AI Labs Private Limited.
Prometheus metrics at /metrics (admin token). Press ? for keyboard shortcuts.