Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

3.7 KB
Raw Blame
substitutions:
  _REGION: us-central1
  _RUN_SA: avatar-reconstruction-sa@aerial-vehicle-466722-p5.iam.gserviceaccount.com
  _SERVICE: model-trellis2
  _REPO: model-trellis2
  _GCS_BUCKET: three-ws-avatar-reconstructions
  _WEIGHTS_BUCKET: three-ws-model-weights
  _WEIGHTS_BUCKET_PREFIX: trellis2-4b
  # Scale-to-zero by default. The us-central1 L4 grant is shared by every GPU
  # lane and warm pins elsewhere already hold it (docs/ops/gcp-credits-plan.md),
  # so a warm floor here is a quota decision, not a default. The deploy still
  # works at 0: the first request pays the weight stream and model load, which
  # the API reports as a cold start (BACKENDS.trellis2.coldStartSeconds). Raise
  # it with `gcloud run services update model-trellis2 --min-instances=1` once
  # the grant allows, or add trellis2 to FORGE_KEEPWARM_LANES.
  _MIN_INSTANCES: "0"
  _MAX_INSTANCES: "2"

steps:
  - id: build
    name: gcr.io/cloud-builders/docker
    dir: workers/model-trellis2
    args:
      - build
      - --tag
      - ${_REGION}-docker.pkg.dev/$PROJECT_ID/${_REPO}/server:$SHORT_SHA
      - --tag
      - ${_REGION}-docker.pkg.dev/$PROJECT_ID/${_REPO}/server:latest
      - --cache-from
      - ${_REGION}-docker.pkg.dev/$PROJECT_ID/${_REPO}/server:latest
      - .

  - id: push
    name: gcr.io/cloud-builders/docker
    args:
      - push
      - --all-tags
      - ${_REGION}-docker.pkg.dev/$PROJECT_ID/${_REPO}/server

  - id: deploy
    name: gcr.io/google.com/cloudsdktool/cloud-sdk:slim
    entrypoint: gcloud
    args:
      - run
      - deploy
      - ${_SERVICE}
      - --image=${_REGION}-docker.pkg.dev/$PROJECT_ID/${_REPO}/server:$SHORT_SHA
      - --region=${_REGION}
      - --service-account=${_RUN_SA}
      - --platform=managed
      - --port=8080
      - --gpu=1
      - --gpu-type=nvidia-l4
      - --no-gpu-zonal-redundancy
      - --no-cpu-throttling
      - --cpu=8
      - --memory=32Gi
      - --min-instances=${_MIN_INSTANCES}
      - --max-instances=${_MAX_INSTANCES}
      - --timeout=900
      - --allow-unauthenticated
      # The checkpoints (about 15 GiB) do not fit the in-memory filesystem next to
      # the model, so main.py streams them from WEIGHTS_GCS_URI one at a time
      # (download, load, delete). The gcsfuse volume below is the fallback read
      # path and carries no cold-start cost until a file is opened.
      - --add-volume=name=weights,type=cloud-storage,bucket=${_WEIGHTS_BUCKET}
      - --add-volume-mount=volume=weights,mount-path=/weights
      - --set-env-vars=GCS_BUCKET=${_GCS_BUCKET},WEIGHTS_DIR=/weights/${_WEIGHTS_BUCKET_PREFIX},MAX_CONCURRENT=1,WEIGHTS_GCS_URI=gs://${_WEIGHTS_BUCKET}/${_WEIGHTS_BUCKET_PREFIX},WEIGHTS_LOCAL_DIR=/tmp/trellis2-weights
      - --set-secrets=API_KEY=avatar-reconstruction-key:latest
      # Recycle an instance whose model load is permanently dead. main.py answers
      # /health with 503 only once the whole load-retry budget is spent, so an
      # ordinary cold load (200 while loading) is never killed mid-load.
      - --liveness-probe=httpGet.path=/health,initialDelaySeconds=60,periodSeconds=60,timeoutSeconds=10,failureThreshold=3

images:
  - ${_REGION}-docker.pkg.dev/$PROJECT_ID/${_REPO}/server:$SHORT_SHA
  - ${_REGION}-docker.pkg.dev/$PROJECT_ID/${_REPO}/server:latest

options:
  machineType: E2_HIGHCPU_32
  logging: CLOUD_LOGGING_ONLY

# CUDA extension compiles (nvdiffrast, CuMesh, FlexGEMM, o-voxel) push the image
# build well past Cloud Build's 10-minute default. Give it an hour.
timeout: 3600s

# The project's default compute SA was deleted: builds must pin the dedicated
# build account or submits fail with "Unknown service account".
serviceAccount: projects/aerial-vehicle-466722-p5/serviceAccounts/[email protected]