[Fix] Unify pinned host pool release on graceful shutdown (#32029)

This commit is contained in:
Liangsheng Yin
2026-07-22 12:29:34 -07:00
committed by GitHub
parent f5dcbe8f14
commit 5c6a29b3c3
7 changed files with 29 additions and 5 deletions
@@ -105,6 +105,9 @@ class DecodeKVCacheOffloadManager:
self.offload_inflight = {}
logger.info("Enable offload kv cache for decode side")
def release_host_resources(self) -> None:
self.decode_host_mem_pool.destroy()
def _mark_offload_started(self, rid):
self.offload_inflight[rid] = self.offload_inflight.get(rid, 0) + 1
+3 -5
View File
@@ -1465,11 +1465,9 @@ 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()
self.tree_cache.release_host_resources()
if self.decode_offload_manager is not None:
self.decode_offload_manager.release_host_resources()
def run_event_loop(self) -> None:
"""Run the scheduler's event loop.
@@ -236,6 +236,13 @@ class BasePrefixCache(ABC, PrefixCacheTrait):
)
self.metrics_collector.increment_eviction_num_tokens(num_evicted)
def release_host_resources(self) -> None:
"""Release pinned host buffers in userspace on graceful shutdown.
Kernel-side unpinning during process reclaim can stall teardown for
tens of seconds (see HostKVCache.destroy). Idempotent.
"""
@abstractmethod
def reset(self):
pass
@@ -207,6 +207,9 @@ class HiMambaRadixCache(MambaRadixCache):
)
super().reset()
def release_host_resources(self) -> None:
self.host_pool_group.destroy()
def write_backup(self, node: TreeNode, write_back=False) -> int:
# Backup invariant (for write-through mode): backed-up nodes must form a
# contiguous prefix from root — no gaps. Skip if parent isn't backed
@@ -775,6 +775,10 @@ class HiRadixCache(RadixCache):
self.evictable_host_leaves.clear()
super().reset()
def release_host_resources(self) -> None:
if self.token_to_kv_pool_host is not None:
self.token_to_kv_pool_host.destroy()
def get_height(self, node: TreeNode):
height = 0
while node != self.root_node:
@@ -1573,6 +1573,10 @@ class HostPoolGroup:
for entry in self.entries:
entry.host_pool.clear()
def destroy(self) -> None:
for entry in self.entries:
entry.host_pool.destroy()
def available_size(self):
return self.anchor_entry.host_pool.available_size()
@@ -371,6 +371,7 @@ class UnifiedRadixCache(KVCacheEventMixin, BasePrefixCache):
# HiCache D↔H defaults (overridden by init_hicache)
self.cache_controller: Optional[HybridCacheController] = None
self.host_pool_group = None # set by attach_hybrid_pool_to_unified_cache
self.write_through_threshold = 256
self.prefetch_stop_policy = "best_effort"
self.prefetch_threshold = 256
@@ -568,6 +569,10 @@ class UnifiedRadixCache(KVCacheEventMixin, BasePrefixCache):
def register_sidecar_pool(self, spec: SidecarPoolSpec) -> None:
self.sidecar_pool_specs.append(spec)
def release_host_resources(self) -> None:
if self.host_pool_group is not None:
self.host_pool_group.destroy()
def match_prefix(self, params: MatchPrefixParams) -> MatchResult:
result = self.session.try_match_prefix(params)
if result is not None: