Drop default preload to fit free-tier memory

hf-entrypoint.sh

@@ -3,18 +3,30 @@
 # Single-container layout for Hugging Face Spaces.
 set -euo pipefail
 
-# Models to preload at SIE startup. Same trio the local compose uses.
-PRELOAD="microsoft/Florence-2-base,naver-clova-ix/donut-base-finetuned-cord-v2,urchade/gliner_multi-v2.1"
+# Models to preload at SIE startup. Empty by default on HF Spaces free tier
+# (16 GB advertised but the cgroup-imposed memory ceiling is tight enough
+# that preloading multiple model classes triggers LRU eviction during boot,
+# which makes the first user click much slower than a cold-load anyway).
+#
+# Override via the PRELOAD env var in the Space's Settings if you upgrade to
+# CPU-Upgrade (32 GB) or a GPU tier:
+#   PRELOAD="microsoft/Florence-2-base,naver-clova-ix/donut-base-finetuned-cord-v2,urchade/gliner_multi-v2.1"
+PRELOAD="${PRELOAD:-}"
 
-echo "[hf-entrypoint] starting sie-server on 127.0.0.1:8080 with preload=$PRELOAD"
-sie-server serve \
-  --host 127.0.0.1 \
-  --port 8080 \
-  --preload "$PRELOAD" &
+SIE_ARGS=(serve --host 127.0.0.1 --port 8080)
+if [ -n "$PRELOAD" ]; then
+  echo "[hf-entrypoint] starting sie-server with preload=$PRELOAD"
+  SIE_ARGS+=(--preload "$PRELOAD")
+else
+  echo "[hf-entrypoint] starting sie-server with no preload (lazy-load on first request)"
+fi
+
+sie-server "${SIE_ARGS[@]}" &
 SIE_PID=$!
 
-# Wait up to 20 minutes for SIE to come up. First boot pulls model weights
-# from HF; subsequent restarts reuse the /data cache and start in seconds.
+# Wait up to 20 minutes for SIE to come up. Without preload this should be
+# ~30 seconds. With preload the first user has to download model weights
+# from HF before SIE marks itself ready.
 echo "[hf-entrypoint] waiting for /healthz"
 for i in $(seq 1 1200); do
   if curl -fsS http://127.0.0.1:8080/healthz > /dev/null 2>&1; then