Skip to main content

​
InferCrane v1 consolidated release acceptance

The v1 workflow freezes one RC and performs all paid/manual evidence once. It is resumable by run ID, rejects concurrent paid runs, stops on the first missing observation, cleans each deployment before moving on, and records logs under .infercrane/v1-acceptance/RUN_ID/.
preflight performs local qualification plus read-only provider/configuration checks. qualify requires a clean checkout at local tag v1.0.0-rc.1; it never creates a stable tag or publishes an artifact. Reuse the same run ID after disconnect or interruption. Do not start a replacement run while an external create response is unresolved.

​
Required infrastructure

​
RunPod

  • readable API-key file in RUNPOD_KEY_FILE
  • immutable Serverless vLLM template ID in INFERCRANE_RUNPOD_SERVERLESS_TEMPLATE_ID
  • secure elastic GPU capacity for a maximum of two replicas
  • public, digest-pinned qualified worker image
The existing RunPod harness runs elastic/serverless traffic, benchmark, autoscaling, Guard, disconnect/restart/delete disruption, lost-create-response adoption, cancellation, and cleanup.

​
AWS BYOC

Create the narrow resources documented in AWS EC2 BYOC: role with external ID, private subnet/security group, qualified GPU AMI and instance type, instance profile, worker secret, and immutable runtime images. Copy .env.production.example plus .env.aws.example into one private environment file and set:
The spec directory must contain vllm.yaml, sglang.yaml, and custom-oci.yaml. The harness copies each spec into its restricted run state and replaces the top-level deployment name with a unique, run-scoped name; it never mutates the source files or reuses a fixed public identity. Every image and model revision must be immutable. The control plane is started through the provider-neutral production Compose file plus the AWS overlay. Create one restricted file containing a random credential of at least 32 characters and install the same value in the AWS Secrets Manager and Kubernetes worker Secrets referenced by the provider configuration:
The harness reads this file without copying its value into evidence. Do not reuse a personal or long-lived production credential.

​
Kubernetes

Provide a restricted kubeconfig, namespace/RBAC from deploy/kubernetes, runtime service account and worker secret, GPU nodes/device plugin, and immutable runtime images. Copy .env.production.example plus .env.kubernetes.example into one private file and set:
Use the same three filenames. Select deployment or standard kserve explicitly; advanced Dynamo/llm-d topologies are not implied by this gate.

​
Runtime spec example

Use the provider/runtime documentation to construct SGLang and OCI specs; do not reuse placeholder digests from repository examples.

​
Evidence and cleanup

The portable stages prove doctor, readiness, buffered/streaming traffic, vLLM tool/structured output, AIPerf, deletion, InferCrane orphan inventory, and direct provider inventory. The portable harness captures the existing managed AWS/Kubernetes inventory before mutation and requires an identical inventory afterward. This detects run-owned leaks without requiring an otherwise-used account or namespace to be empty. RunPod directly lists run-owned Pods and endpoints. After the report says real_infrastructure: passed, independently confirm:
Provider deletion is not proven by absent local rows. If cleanup was interrupted, resume qualify with the same run ID so it can use persisted identities; do not erase its PostgreSQL volume first.

​
What automated local evidence covers

The frozen RC already carries hermetic evidence for explicit governed fallback, budget exhaustion, privacy acknowledgement, oscillation control, no-replay streaming, provider create-response loss, runtime translation, migration history, and tenant isolation. A real OpenRouter request is not silently added to final qualification: it would transmit content and incur separate cost. If an operator elects to add one, it must remain an explicit, redacted, hard-budgeted supplemental record. No fake worker, Kind cluster, or provider simulator may be cited as real GPU qualification. No provider timing, cost, or performance result may be inferred from an operator timeout.