Baseten Terraform provider

Experimental Terraform provider for Baseten custom model deployments.

Development

nix develop
scripts/test.sh ci

Useful test suites:

scripts/test.sh unit
scripts/test.sh lint
scripts/test.sh vuln
scripts/test.sh build
scripts/test.sh docs
scripts/test.sh release-check
TF_ACC=1 BASETEN_API_KEY=... scripts/test.sh acceptance

CI runs scripts/test.sh ci on pushes and pull requests. Tags matching v* run the release workflow, which creates draft GitHub release artifacts with GoReleaser.

Acceptance tests provision real Baseten resources. Set BASETEN_ACC_ACCELERATOR to override the default A10G test accelerator.

Custom model example

See examples/custom-model for a complete fixture that is also exercised by the local provider plan test.

terraform {
  required_providers {
    baseten = {
      source = "polymath-as/baseten"
    }
  }
}

provider "baseten" {
  # Or set BASETEN_API_KEY.
  api_key = var.baseten_api_key
}

locals {
  model_path = "${path.module}/model"
  model_files = sort(fileset(local.model_path, "**"))
  model_hash = sha256(join("", [
    for file in local.model_files : filesha256("${local.model_path}/${file}")
  ]))
}

resource "baseten_custom_model" "example" {
  name        = "example-custom-model"
  source_path = local.model_path
  source_hash = local.model_hash

  raw_config  = file("${local.model_path}/config.yaml")
  config_json = jsonencode(yamldecode(file("${local.model_path}/config.yaml")))

  environment_name = "production"

  min_replica        = 0
  max_replica        = 1
  scale_down_delay   = 120
  concurrency_target = 2

  timeouts {
    create = "90m"
    update = "10m"
    delete = "10m"
  }
}

min_replica = 0 enables scale-to-zero. Update source_hash when local model contents change so Terraform replaces the Baseten deployment with a new archive. Long builds can raise the timeouts.create value.

Existing deployments can be imported by model and deployment ID:

terraform import baseten_custom_model.example model-123:deployment-456

Keep the local model block configured after import; Baseten cannot reconstruct local-only inputs such as source_path, source_hash, and config_json.

Large model artifacts

Archives are streamed into Baseten's temporary S3 upload location with multipart upload, so the provider does not keep the full tarball in memory. For models with hundreds of GB of weights, prefer keeping weights in external storage or a model registry and referencing them from config.yaml; use the Terraform upload path for deployment code and smaller artifacts.