> ## Documentation Index
> Fetch the complete documentation index at: https://infercrane.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Terraform provider

> Manage logical InferCrane deployments without transferring provider-resource ownership to Terraform.

The Terraform provider manages `infercrane_deployment` through the control-plane API. InferCrane
continues to own replicas, provider request identity, drain fencing, revisions, Release Guard, and
cleanup.

`infercrane_slo_policy` manages the deterministic policy used by `infercrane recommend`:

```hcl theme={"theme":{"light":"github-light-default","dark":"vesper"}}
resource "infercrane_slo_policy" "qwen" {
  deployment               = infercrane_deployment.qwen.name
  max_ttft_p95_ms          = 250
  max_error_rate           = 0.01
  min_output_tokens_second = 30
}
```

Destroying this resource deletes the policy but preserves benchmark and recommendation history.
Recommendations remain an explicit evaluation rather than a Terraform apply.

<Warning>
  The provider is not yet published in the Terraform Registry. Local qualification builds it from
  `integrations/terraform`; registry installation begins only after release publication.
</Warning>

```hcl theme={"theme":{"light":"github-light-default","dark":"vesper"}}
terraform {
  required_providers {
    infercrane = {
      source  = "infercrane/infercrane"
      version = "2.0.0"
    }
  }
}

provider "infercrane" {
  endpoint = "https://infercrane.internal"
  # Prefer INFERCRANE_API_KEY; api_key is sensitive when set here.
}

resource "infercrane_deployment" "qwen" {
  name         = "qwen-prod"
  model        = "Qwen/Qwen3-8B"
  runtime      = "vllm"
  cloud        = "runpod"
  compute_mode = "elastic"
  gpu          = "L40S"
  min_replicas = 1
  max_replicas = 4

  operation_timeout_seconds = 900
}
```

## Lifecycle behavior

| Terraform transition | InferCrane behavior                                                       |
| -------------------- | ------------------------------------------------------------------------- |
| Create               | Submit one idempotent durable deployment operation and wait               |
| Read                 | Refresh logical state and active revision through the API                 |
| Update               | Candidate → provision → Release Guard → promote; a rejection fails safely |
| Delete               | Submit durable logical deletion and wait for fenced cleanup               |
| Import               | `terraform import infercrane_deployment.qwen qwen-prod`                   |

If Terraform is interrupted, InferCrane continues the operation. The next apply derives the same
idempotency identity, adopts the original logical deployment, and resumes observation. Terraform
state never stores a provider API key supplied through `INFERCRANE_API_KEY`, and it does not contain
provider-owned replica identifiers as desired inputs.

<Note>
  An update can remain unpromoted when Release Guard lacks sufficient evidence or rejects the
  candidate. Inspect the persisted reason with `infercrane rollout inspect qwen-prod`; Terraform does
  not bypass the guard to force convergence.
</Note>