#!/usr/bin/env bash
# Steering planner VLM for the Space: native vLLM, Qwen3-VL-30B-A3B, A100 80GB replicas, billed to the lerobot org.
# Needs a token with inference endpoint write access on lerobot (HF_TOKEN or `hf auth login`).
# Usage: bash scripts/deploy_planner.sh show     -> hardware and prices
#        bash scripts/deploy_planner.sh deploy   -> create the endpoint
#        bash scripts/deploy_planner.sh status   -> state and URL (set it as the Space variable PLANNER_URL)
#        bash scripts/deploy_planner.sh pause    -> stop billing
set -euo pipefail
NAME=steering-planner
NAMESPACE=lerobot
MIN_REPLICA=1
MAX_REPLICA=4
# With MIN_REPLICA=1 the endpoint never scales to zero (the API rejects --scale-to-zero-timeout then);
# for cold starts instead of an always-on replica, set MIN_REPLICA=0 and add --scale-to-zero-timeout.

case "${1:-show}" in
  show) hf endpoints hardware --namespace "$NAMESPACE" --vendor aws --accelerator gpu; exit 0 ;;
  pause) hf endpoints pause "$NAME" --namespace "$NAMESPACE"; exit 0 ;;
  status) hf endpoints describe "$NAME" --namespace "$NAMESPACE"; exit 0 ;;
  deploy) ;;
  *) echo "unknown action $1"; exit 1 ;;
esac

hf endpoints deploy "$NAME" \
  --namespace "$NAMESPACE" \
  --repo Qwen/Qwen3-VL-30B-A3B-Instruct \
  --framework custom --engine vllm --custom-image vllm/vllm-openai:v0.31.0 \
  --vendor aws --region us-east-1 \
  --accelerator gpu --instance-type nvidia-a100 --instance-size x1 \
  --min-replica "$MIN_REPLICA" --max-replica "$MAX_REPLICA" \
  --scaling-metric pendingRequests \
  --container-args "--max-model-len 32768" \
  --type authenticated