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.

3.8 KB
Raw Blame
substitutions:
  _REGION: us-central1
  # Runtime identity. It holds secretAccessor on avatar-reconstruction-key,
  # objectViewer on the weights bucket and objectAdmin on the results bucket
  # (bucket-level grants, see README.md "One-time GCP setup").
  _RUN_SA: [email protected]
  _SERVICE: model-modly
  _REPO: model-modly
  _GCS_BUCKET: three-ws-avatar-reconstructions
  _WEIGHTS_BUCKET: three-ws-model-weights
  _WEIGHTS_BUCKET_PREFIX: modly
  # Both models by default. MODLY_MODELS=triposg/generate serves TripoSG alone
  # and halves the weights held in the in-memory /tmp.
  _MODLY_MODELS: triposg/generate,hunyuan3d-mini-turbo/generate
  # Scale-to-zero by default. The us-central1 L4 grant is shared by every GPU
  # lane (docs/ops/gcp-credits-plan.md), so a warm floor here is a quota
  # decision, not a default. The first request after idle pays the weight copy,
  # the Modly start and the warm-up generation, which the API reports as a cold
  # start (BACKENDS.modly.coldStartSeconds). Raise it with
  # `gcloud run services update model-modly --min-instances=1` once the grant
  # allows, or add modly to FORGE_KEEPWARM_LANES.
  _MIN_INSTANCES: "0"
  _MAX_INSTANCES: "2"

steps:
  - id: build
    name: gcr.io/cloud-builders/docker
    dir: workers/modly
    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=${_RUN_SA}
      - --platform=managed
      - --port=8080
      - --gpu=1
      - --gpu-type=nvidia-l4
      - --no-gpu-zonal-redundancy
      - --no-cpu-throttling
      - --cpu=8
      # About 12 GB of weights live in the in-memory /tmp next to two model
      # processes (one per extension, each loaded on first use).
      - --memory=32Gi
      - --min-instances=${_MIN_INSTANCES}
      - --max-instances=${_MAX_INSTANCES}
      - --timeout=900
      # The worker authenticates every task route itself (API_KEY below); only
      # /health and / are open.
      - --allow-unauthenticated
      - --set-env-vars=^@^GCS_BUCKET=${_GCS_BUCKET}@MAX_CONCURRENT=1@WEIGHTS_GCS_URI=gs://${_WEIGHTS_BUCKET}/${_WEIGHTS_BUCKET_PREFIX}@MODLY_MODELS=${_MODLY_MODELS}
      - --set-secrets=API_KEY=avatar-reconstruction-key:latest
      # Recycle an instance whose bring-up is permanently dead. main.py answers
      # /health with 503 only once the whole bring-up retry budget is spent, so
      # an ordinary cold start (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

# Torch, two extension venvs and the diso CUDA compile 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]