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-pixal3d
  _REPO: model-pixal3d
  _GCS_BUCKET: three-ws-avatar-reconstructions
  _WEIGHTS_BUCKET: three-ws-model-weights
  _WEIGHTS_BUCKET_PREFIX: pixal3d
  _MVADAPTER_BUCKET_PREFIX: pixal3d-mvadapter
  # Scale-to-zero by default; see workers/model-trellis2/cloudbuild.yaml's
  # identical comment about the shared us-central1 L4 grant
  # (docs/ops/gcp-credits-plan.md). The first request after a cold start pays
  # the weight stream and pipeline load (BACKENDS.pixal3d.coldStartSeconds).
  _MIN_INSTANCES: "0"
  _MAX_INSTANCES: "2"

steps:
  - id: build
    name: gcr.io/cloud-builders/docker
    dir: workers/model-pixal3d
    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=${_MAX_INSTANCES}
      - --timeout=900
      - --allow-unauthenticated
      # The checkpoint tree (about 46 GiB across both pipelines) does not fit the
      # in-memory filesystem next to a loaded model, so main.py streams each
      # sub-model from WEIGHTS_GCS_URI one at a time (download, load, delete). The
      # gcsfuse volume below is the fallback read path and carries no cold-start
      # cost until a file is opened.
      - --add-volume=name=weights,type=cloud-storage,bucket=${_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/pixal3d-weights,MVADAPTER_WEIGHTS_DIR=/weights/${_MVADAPTER_BUCKET_PREFIX}
      - --set-secrets=API_KEY=avatar-reconstruction-key:latest
      # Recycle an instance whose weight staging 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 staging) 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, natten) push the
# image build well past Cloud Build's 10-minute default. Give it an hour, same as
# workers/model-trellis2.
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]