Provision a 5090-tuned container and bake in cuDNN 9.17 fix

- config: batch_size_per_card 14 -> 32 (5090 32GB headroom)
- setup_server.sh: pin nvidia-cudnn-cu13>=9.17 to match the sm_120 wheel
  (without it conv2d hits "Cannot load symbol cublasLtCreate" abort)
- new scripts/recreate_container.sh: one-shot rebuild with --shm-size 8g,
  preserves /root/.netrc so wandb auth survives, runs setup_server.sh

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

configs/kr_lp_pgnet.yml

@@ -95,7 +95,7 @@ Train:
   loader:
     shuffle: True
     drop_last: True
-    batch_size_per_card: 14
+    batch_size_per_card: 32
     num_workers: 8
 
 Eval:

scripts/recreate_container.sh

@@ -0,0 +1,50 @@
+#!/usr/bin/env bash
+# kr_lp_pgnet 컨테이너 재생성 (--shm-size 8g 적용 등)
+#
+# 호스트에서 실행:
+#   bash scripts/recreate_container.sh
+#
+# 동작:
+#   1. /root/.netrc 백업 (wandb 인증 보존)
+#   2. 기존 컨테이너 강제 삭제
+#   3. 새 컨테이너 시작 (--shm-size 8g, --gpus all, bind mount)
+#   4. .netrc 복원
+#   5. setup_server.sh 자동 실행 (paddle 등 재설치)
+
+set -euo pipefail
+
+NAME=kr_lp_pgnet
+IMAGE=ubuntu:24.04
+SHM=8g
+WORKSPACE=/home/cuuva/workspace
+
+# 1. .netrc 백업 (wandb 로그인 보존용)
+NETRC_BAK=/tmp/${NAME}_netrc.bak
+if docker exec "$NAME" test -f /root/.netrc 2>/dev/null; then
+    docker cp "$NAME":/root/.netrc "$NETRC_BAK"
+    echo "  ✓ /root/.netrc backed up → $NETRC_BAK"
+fi
+
+# 2. 기존 컨테이너 강제 정리
+docker rm -f "$NAME" 2>/dev/null || true
+
+# 3. 새 컨테이너 (--shm-size 8g 핵심)
+docker run -d --name "$NAME" --gpus all \
+    --shm-size="$SHM" \
+    -v "$WORKSPACE:/workspace" \
+    -w /workspace \
+    --restart unless-stopped \
+    "$IMAGE" sleep infinity
+echo "  ✓ 새 컨테이너 시작 (shm-size=$SHM)"
+
+# 4. .netrc 복원
+if [ -f "$NETRC_BAK" ]; then
+    docker cp "$NETRC_BAK" "$NAME":/root/.netrc
+    rm "$NETRC_BAK"
+    echo "  ✓ /root/.netrc 복원 (wandb login 유지)"
+fi
+
+# 5. setup_server.sh 자동 실행 (paddle/PaddleOCR/cuDNN 등 재설치)
+echo
+echo "  → setup_server.sh 실행..."
+docker exec "$NAME" bash /workspace/kr_lp_pgnet/scripts/setup_server.sh

scripts/setup_server.sh

@@ -54,8 +54,11 @@ while IFS= read -r line; do
     $PIP install $PIP_OPTS --ignore-installed "$line" 2>/dev/null || echo "  skip: $line"
 done < /tmp/kr_lp_req.txt
 
-echo "[5/7] OpenCV (numpy2 호환) + numpy<2 (PaddleOCR release/2.7 호환성) + wandb"
+echo "[5/7] OpenCV (numpy2 호환) + numpy<2 (PaddleOCR release/2.7 호환성) + wandb + cuDNN 9.17"
 $PIP install $PIP_OPTS 'opencv-python>=4.10' 'opencv-contrib-python>=4.10' wandb
+# paddle sm_120 wheel은 cuDNN 9.17 빌드라 paddle deps의 9.13.0.50을 9.17로 upgrade 필요.
+# (안 하면 conv2d에서 cublasLtCreate 심볼 로드 실패 → process abort)
+$PIP install $PIP_OPTS --upgrade 'nvidia-cudnn-cu13>=9.17,<9.18'
 # imgaug 등이 numpy 1.x API(np.sctypes)에 의존하므로 numpy 1.x로 핀.
 # paddle 3.3.0.dev는 numpy 1.x도 호환.
 $PIP install $PIP_OPTS 'numpy<2' --force-reinstall