CosmicAC Logo
Configuration reference

Job configuration reference

Fields you set when you create a GPU Container Job or a Managed Inference Job.

This page lists the fields that every job takes when you create it, in the web interface or with cosmicac jobs create. For the create flow, see Create a GPU Container Job and Create a Managed Inference Job.

Each job type also takes its own fields, listed on the following pages.

Job typeFields
GPU Container JobGPU Container Job configuration
vLLM Managed Inference JobvLLM Managed Inference Job configuration
Parakeet Managed Inference JobParakeet Managed Inference Job configuration

The field tables on these pages use the following columns.

  • Required: whether the CosmicAC API rejects a job without the field. The web interface and the CLI supply a value for each optional field.
  • CLI flag: the cosmicac jobs create flag that sets the field in non-interactive mode.

Common fields

Every job type takes these fields.

FieldRequiredCLI flagDescription
Job typeYes--typeThe kind of job to create, either GPU Container or Managed Inference.
Job nameYes--nameThe name that identifies the job.
TagsYes--tagsOne or more labels for the job. The CLI takes a comma-separated list.
LocationYes--locationThe location where the job runs, for example PT. The CLI lists the locations that your racks report.
NotificationsNo--notifyThe job lifecycle events that the job reports. The value is a comma-separated list of job_failed, job_degraded, job_recovered, and job_restart_storm, or all, or none. An event reaches your webhook only if its switch is also on in Settings → Notifications. See What controls delivery.

GPU configuration

These fields select the job's hardware.

FieldRequiredCLI flagDescription
GPUYes--gpu-typeThe GPU to use, for example GH100_H100_SXM5_80GB. The CLI lists the GPU types that your racks report.
GPU countYes--gpu-countThe number of GPUs for one replica. One of 1, 2, 4, or 8. A vLLM Managed Inference Job also takes 16 for a multi-node replica.
CUDA / driverNo--driverThe GPU driver version, for example CUDA 13.0. Your model masters set the available versions.
CPU cores per GPUNo--cpu-cores-per-gpuThe number of CPU cores to reserve for each GPU. If you omit this field, CosmicAC uses the value that your rack reports for the GPU type.
RAM per GPUNo--mem-gb-per-gpuThe system memory, in GB, to reserve for each GPU. Use a whole number. If you omit this field, CosmicAC uses the value that your rack reports for the GPU type.

To set the GPU type and count in one flag, use --gpu TYPE=COUNT, for example --gpu H100=2. This flag replaces --gpu-type and --gpu-count.

On a node with 8 GPUs, a replica with 1, 2, 4, or 8 GPUs runs on one node, and a replica with 16 GPUs takes two whole nodes. K8S_GPU_PER_NODE in your deployment configuration sets the number of GPUs per node, and .env.example sets it to 8. Only a vLLM Managed Inference Job can run a multi-node replica. See Multi-node replicas.

For a Managed Inference Job, the GPU count must equal the model's per_replica_gpu_count, listed in Recommended model parameters. Otherwise, CosmicAC rejects the job with a 400 status and the message gpu.count must equal per_replica_gpu_count (N).

On this page