[CI] Rewrite killall_sglang as Python with CI/local dual mode (#21331)

Co-authored-by: Alison Shao <alison.shao@mac.lan>
Co-authored-by: Alison Shao <alison.shao@MacBook-Pro-D2W773R9CD.local>
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>

scripts/ci/cuda/ci_install_dependency.sh

@@ -84,9 +84,16 @@ mark_step_done "Host / runner detection"
 # Kill existing processes
 # ------------------------------------------------------------------------------
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-bash "${SCRIPT_DIR}/../../killall_sglang.sh"
+REPO_ROOT="$(cd "${SCRIPT_DIR}/../../.." && pwd)"
+python3 "${REPO_ROOT}/python/sglang/cli/killall.py"
+KILLALL_EXIT=$?
 echo "CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-}"
 
+if [ $KILLALL_EXIT -ne 0 ]; then
+    echo "ERROR: killall.py detected uncleanable GPU memory. Aborting CI."
+    exit 1
+fi
+
 mark_step_done "Kill existing processes"
 
 # ------------------------------------------------------------------------------

scripts/killall_sglang.sh

@@ -1,5 +1,10 @@
 #!/bin/bash
 
+# DEPRECATED: This script will be migrated to python/sglang/cli/killall.py.
+# CI mode is already handled there. This script remains for local/non-CI usage.
+#
+# TODO: Migrate remaining modes (rocm, all, gpus) to killall.py and remove this file.
+#
 # Usage:
 #   ./killall_sglang.sh              - Kill SGLang processes only (NVIDIA mode)
 #   ./killall_sglang.sh rocm         - Kill SGLang processes only (ROCm mode)