TGI fast startup (WIP)

cloudbuild.yaml

steps:
- name: 'gcr.io/cloud-builders/docker'
  args: ['build', '-t', 'us-central1-docker.pkg.dev/$PROJECT_ID/containers/tgi', '.']
- name: 'gcr.io/cloud-builders/docker'
  args: ['push', 'us-central1-docker.pkg.dev/$PROJECT_ID/containers/tgi']
images:
- us-central1-docker.pkg.dev/$PROJECT_ID/containers/tgi
options:
 machineType: 'N1_HIGHCPU_32'
 diskSizeGb: '500'

Dockerfile

# Use official Huggingface TGI image (See https://huggingface.co/docs/text-generation-inference/en/quicktour)
FROM ghcr.io/huggingface/text-generation-inference:1.4

# Model to use. Customize with docker build --build-arg MODEL_HUB_ID=your model .
ARG MODEL_HUB_ID=tiiuae/falcon-7b-instruct
# Port to listen to
ARG PORT=8080

# Download model
RUN text-generation-server download-weights $MODEL_HUB_ID

# Start the server at container startup
ENTRYPOINT text-generation-launcher --model-id $MODEL_HUB_ID --port $PORT