CosmicAC Logo
Configuration referenceJob configurationManaged Inference Job

vLLM Managed Inference Job configuration

Fields you set when you create a vLLM Managed Inference Job.

A vLLM Managed Inference Job takes the following fields, as well as the common fields and GPU configuration that every job takes.

FieldRequiredCLI flagDescription
ModelYes--modelThe Hugging Face model ID to serve, for example MiniMaxAI/MiniMax-M2.5.
Runtime image (CUDA)Yes--runtime-imageThe serving runtime image, as a Docker image reference such as vllm/vllm-openai:v0.15.1. The web interface and interactive cosmicac jobs create use the model master's image. In non-interactive mode, if you omit --runtime-image, the CLI sends vllm/vllm-openai:v0.17.1.
Data typeYes--data-typeThe numeric precision the model runs at, for example BF16 or Auto. If you omit --data-type, the CLI sends Auto.
QuantisationNo--quantisationThe quantisation scheme, for example fp8 or awq.
Tensor parallelNo--tensor-parallelThe number of GPUs to split the model across. One of 1, 2, 4, 8, or none. See Tensor parallel.
GPU memory utilizationNo--gpu-memory-utilizationThe fraction of GPU memory to use, from 0 to 1.
Max concurrent sequencesNo--max-concurrent-sequencesThe maximum number of requests the model handles at once, from 1 to 512.
Max model lengthNo--max-model-lengthThe maximum context length, in tokens. The web interface and interactive cosmicac jobs create use the model master's value. In non-interactive mode, if you omit --max-model-length, the CLI sends 32768. See max_tokens in the Inference API reference.
Reasoning parserNo--reasoning-parserThe parser that separates a model's thinking from its final response. The value default means no parser.
Video & image inputNo--multimodalWhether the model accepts multimodal input. true or false.
Root disk sizeNo--root-disk-size-gbThe root disk size of the virtual machine, in GB. See Root disk size.
Environment variablesNo--envVariables for the model server and the inference agent. A variable named after a vllm serve flag sets that option. See vLLM serving options.
Endpoint nameYes--endpoint-nameThe name of the inference endpoint, which forms a path segment in its URL. The name has 3 to 63 characters, and uses only lowercase letters, digits, and hyphens. It must be unique across every Managed Inference Job in the deployment, including failed jobs that you haven't deleted.
ReplicasNo--replicaThe number of endpoint replicas. One of 1, 2, or 4.
Require Authorization headerNo--require-auth-header / --no-auth-headerWhether callers must send an authorization header. true or false.

Tensor parallel

The tensor parallel value must divide the job's GPU count exactly. Otherwise, CosmicAC rejects the job with a 400 status and the message tensor_parallel must evenly divide the GPU count per instance.

If a job has no tensor parallel value, CosmicAC uses the model's per_replica_gpu_count, listed in Recommended model parameters. The job's GPU count equals that number, so the value always divides it.

The web interface doesn't show the field, and sends the model master's per_replica_gpu_count. The CLI preselects the same value in interactive mode, and reads --tensor-parallel in non-interactive mode.

Root disk size

The minimum root disk size depends on the model. It's the model's disk_gb, listed in Recommended model parameters. If the value is smaller, CosmicAC rejects the job with the message root_disk_size_gb (X) is below model master disk_gb (Y).

On this page