Inference

Batch / offline inference

Throughput-first offline scoring or embedding generation, ideal for spot capacity.

RayArgo WorkflowsvLLM (offline)TensorRT

Architecture

A work queue feeds GPU batch workers that read from and write to object storage. Idempotent, resumable work runs safely on preemptible nodes.

Sizing

Dataset size ÷ target completion time; scale workers horizontally at high batch utilization.

Networking

North-south to object storage; keep workers close to data.

Storage

Object in, object out; cache the model locally; shard work for parallelism.

Scaling

Scale workers to zero when the queue drains; checkpoint/resume for preemption.

Monitoring

Items/sec, cost per item, queue depth, preemption/retry rate.

Sample manifest
starting point
apiVersion: batch/v1
kind: Job
metadata:
  name: batch-embed
spec:
  parallelism: 8
  completions: 8
  template:
    spec:
      restartPolicy: OnFailure
      containers:
        - name: worker
          image: your-registry/embed:latest
          resources: { limits: { nvidia.com/gpu: 1 } }
Recommended GPUs