[Fix] Release hierarchical cache host pool on graceful shutdown (#31746)

python/sglang/srt/managers/scheduler.py

@@ -1460,6 +1460,11 @@ class Scheduler(
         # HostKVCache.destroy. Called from run_scheduler_process's finally.
         if self.hisparse_coordinator is not None:
             self.hisparse_coordinator.destroy()
+        # A plain HiRadixCache (no hisparse) also holds a large pinned host KV
+        # pool; unregister it here too, else the kernel unpins it during reclaim.
+        host_pool = getattr(self.tree_cache, "token_to_kv_pool_host", None)
+        if host_pool is not None:
+            host_pool.destroy()
 
     def run_event_loop(self) -> None:
         """Run the scheduler's event loop.

test/registered/hicache/test_hicache_storage.py

@@ -3,6 +3,7 @@ from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 register_cuda_ci(est_time=99, stage="base-b", runner_config="1-gpu-small")
 register_amd_ci(est_time=300, suite="stage-b-test-1-gpu-small-amd")
 
+import subprocess
 import time
 import unittest
 
@@ -47,6 +48,14 @@ class TestHiCache(CustomTestCase, MMLUMixin):
 
     @classmethod
     def tearDownClass(cls):
+        # Graceful stop first so the server unregisters its large pinned host KV
+        # pool in userspace; a bare SIGKILL leaves the kernel to unpin it during
+        # reclaim, holding GPU memory long enough to fail the next test.
+        cls.process.terminate()
+        try:
+            cls.process.wait(timeout=60)
+        except subprocess.TimeoutExpired:
+            pass
         kill_process_tree(cls.process.pid)
         time.sleep(5)