Provider profiles

InferCrane separates the stable cloud identity from the adapter that implements a particular provider product. For example, aws-ec2, aws-asg, and aws-sagemaker are different profiles even though their cloud is aws. This prevents managed endpoint semantics from being forced into a one-resource-per-replica contract.

Cloud	Profile	Local state	Ownership boundary
AWS	`aws-ec2`	Locally qualified	One private VM per replica intent
AWS	`aws-asg`	Registered	ASG owns instances and replacement
AWS	`aws-eks`	Registered	Reuse namespaced Kubernetes lifecycle
AWS	`aws-sagemaker`	Registered	SageMaker owns endpoint children
AWS	`aws-bedrock`	Registered external	Customer-governed external capacity
GCP	`gcp-compute`	Locally qualified	One private VM per replica intent
GCP	`gcp-mig`	Registered	MIG owns instances and rolling replacement
GCP	`gcp-gke`	Registered	Reuse namespaced Kubernetes lifecycle
GCP	`gcp-vertex`	Registered	Vertex owns deployed-model children
CoreWeave	`coreweave-cks`	Registered	Reuse namespaced Kubernetes lifecycle; CKS owns GPU platform services

Registered means the capability and ownership boundary is modeled and inspectable. It does not mean InferCrane can execute that profile. Local and real qualification are separate evidence states. Use infercrane integrations rather than relying on this prose for the running binary’s exact state.

Exact selection

Use an adapter only when a cloud/runtime pair has multiple installed implementations:

infercrane deploy Qwen/Qwen3-8B \
  --cloud gcp \
  --provider-adapter gcp-compute \
  --region europe-west4 \
  --gpu nvidia-l4

The adapter identity is persisted in immutable revision state. Reconciliation therefore resumes the same implementation after a control-plane restart instead of guessing from currently installed plugins.

⌘I