Add restart_server.sh (server.py + restart_server.sh + Korean README expansion)

restart_server.sh

@@ -0,0 +1,90 @@
+#!/bin/bash
+# One-shot rehydrate + relaunch for the Korean pest-detector server.
+# Idempotent: skips installs if the packages are already importable.
+# Run this after any pod restart (which wipes the container disk).
+#
+# Usage:   bash /workspace/restart_server.sh
+set -uo pipefail
+
+PORT="${PORT:-8080}"
+ADAPTER="${ADAPTER:-pfox1995/pest-detector-deploy}"
+LOAD_IN_4BIT="${LOAD_IN_4BIT:-true}"
+# RunPod auto-sets RUNPOD_POD_ID — derive proxy URL from it. Override with PUBLIC_URL=...
+PUBLIC_URL="${PUBLIC_URL:-https://${RUNPOD_POD_ID:-YOUR_POD_ID}-${PORT}.proxy.runpod.net}"
+
+cd /workspace
+mkdir -p /workspace/.tmp /workspace/.pip-cache
+export TMPDIR=/workspace/.tmp
+
+step() { echo; echo "=== $* ==="; }
+
+# ───────── Stage 1: stack ─────────
+step "1/3  Python stack"
+need_stack=0
+python3 -c "import unsloth, peft, fastapi, uvicorn, bitsandbytes, fla" 2>/dev/null || need_stack=1
+
+if [ $need_stack -eq 1 ]; then
+  echo "  installing unsloth + transformers + peft + fastapi + bnb + fla"
+  pip install --break-system-packages --cache-dir=/workspace/.pip-cache \
+    'unsloth[cu128-torch280]' 'transformers>=5.2,<6.0' 'peft==0.19.1' \
+    fastapi uvicorn python-multipart bitsandbytes flash-linear-attention \
+    2>&1 | tail -3
+else
+  echo "  ✓ already installed"
+fi
+
+# ───────── Stage 2: causal_conv1d (prebuilt wheel — skip the 9-arch CUDA build) ─────────
+step "2/3  causal_conv1d fast-path lib"
+if python3 -c "import causal_conv1d" 2>/dev/null; then
+  echo "  ✓ already installed: $(python3 -c 'import causal_conv1d; print(causal_conv1d.__version__)')"
+else
+  echo "  installing prebuilt wheel for torch 2.8 + cu12 + py312"
+  pip install --break-system-packages \
+    'https://github.com/Dao-AILab/causal-conv1d/releases/download/v1.6.1.post4/causal_conv1d-1.6.1%2Bcu12torch2.8cxx11abiTRUE-cp312-cp312-linux_x86_64.whl' \
+    2>&1 | tail -3
+  python3 -c "import causal_conv1d; print('  ✓ ccc', causal_conv1d.__version__)"
+fi
+
+# ───────── Stage 3: launch server in tmux ─────────
+step "3/3  launch server in tmux 'pest' on :$PORT"
+tmux kill-session -t pest 2>/dev/null || true
+sleep 1
+
+tmux new-session -d -s pest "
+  export HF_TOKEN=\$(cat /workspace/.cache/huggingface/token)
+  export HF_HOME=/workspace/.cache/huggingface
+  export ADAPTER='$ADAPTER'
+  export LOAD_IN_4BIT='$LOAD_IN_4BIT'
+  export PORT='$PORT'
+  python3 -u /workspace/deploy/server.py 2>&1 | tee /workspace/pest_server.log
+  exec bash
+"
+
+echo
+echo "  waiting for /health (Unsloth load + LoRA attach takes ~90s)..."
+ready_after=""
+for i in $(seq 1 90); do
+  if curl -sf -m 2 "http://localhost:$PORT/health" >/dev/null 2>&1; then
+    ready_after="$((i*2))"
+    break
+  fi
+  sleep 2
+done
+
+if [ -z "$ready_after" ]; then
+  echo "  ✗ server did not come up within 180s. Check:"
+  echo "    tmux capture-pane -t pest -p | tail -30"
+  echo "    tail -40 /workspace/pest_server.log"
+  exit 1
+fi
+
+echo "  ✓ ready after ${ready_after}s"
+echo
+echo "════════════════════════════════════════════════════════════"
+echo "  ✓ pest detector server is live"
+echo "  Local:  http://localhost:$PORT"
+echo "  Public: $PUBLIC_URL"
+echo "  VRAM:   $(nvidia-smi --query-gpu=memory.used,memory.total --format=csv,noheader,nounits | awk -F',' '{printf "%.1f / %.1f GB\n", $1/1024, $2/1024}')"
+echo "  Logs:   tail -f /workspace/pest_server.log"
+echo "  Stop:   tmux kill-session -t pest"
+echo "════════════════════════════════════════════════════════════"