chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

fix: address CP HiCache review regressions

Browse Source
This commit is contained in:
谢骞
2026-05-08 04:11:25 +08:00
parent 8844c303e2
commit f38937723e
4 changed files with 104 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
python/sglang/srt/mem_cache/hiradix_cache.py
View File

@@ -732,23 +732,25 @@ class HiRadixCache(RadixCache):
device_indices=node.value,
node_id=node.id,
)
metadata = getattr(result, "metadata", None)
required_host_slots = 0
if getattr(result, "metadata", None) is None:
if metadata is None:
required_host_slots = result.required_host_slots
self._evict_host_for_physical_slots(
required_host_slots,
synchronize_across_ranks=getattr(self, "tp_world_size", 1) > 1,
)
if getattr(result, "metadata", None) is None:
if metadata is None:
result = self.cache_controller.write(
device_indices=node.value,
node_id=node.id,
)
if getattr(result, "metadata", None) is None:
metadata = getattr(result, "metadata", None)
if metadata is None:
return 0
node.host_len = len(node.value)
node.cp_hicache = result.metadata
node.cp_hicache = metadata
node.host_value = None
self.ongoing_write_through[node.id] = node
if not write_back:
@@ -1100,7 +1102,9 @@ class HiRadixCache(RadixCache):
self, required_host_slots: int, synchronize_across_ranks: bool = False
) -> int:
synchronize_across_ranks = (
synchronize_across_ranks and getattr(self, "tp_world_size", 1) > 1
synchronize_across_ranks
and getattr(self, "tp_group", None) is not None
and getattr(self, "tp_world_size", 1) > 1
)
if required_host_slots <= 0 and not synchronize_across_ranks:
return 0
@@ -1148,9 +1152,7 @@ class HiRadixCache(RadixCache):
if x.host_ref_counter > 0:
continue
host_indices = (
x.cp_hicache.host_indices if x.cp_hicache is not None else None
)
host_indices = self._node_host_evict_indices(x)
physical_count = len(host_indices) if host_indices is not None else 0
self._record_remove_event(x)
@@ -1197,7 +1199,9 @@ class HiRadixCache(RadixCache):
# Block deleted entirely (GPU already evicted, now CPU freed) --
# emit BlockRemoved so the router removes this block from its index.
self._record_remove_event(x)
num_evicted += self.cache_controller.evict_host(x.host_value)
num_evicted += self.cache_controller.evict_host(
self._node_host_evict_indices(x)
)
parent = self._remove_host_leaf(x)
@@ -1581,8 +1585,10 @@ class HiRadixCache(RadixCache):
while last_node.evicted:
host_hit_length += len(last_node.host_value)
last_node = last_node.parent
while not last_host_node.backuped:
while last_host_node != self.root_node and not last_host_node.backuped:
last_host_node = last_host_node.parent
if not last_host_node.backuped:
last_host_node = self.root_node
return MatchResult(
device_indices=value,

6
python/sglang/srt/server_args.py
View File

@@ -745,15 +745,15 @@ class ServerArgs:
Orchestrates the handling of various server arguments, ensuring proper configuration and validation.
"""
# Normalize load balancing defaults early (before dummy-model short-circuit).
self._handle_load_balance_method()
# Validate SSL arguments early (before dummy-model short-circuit).
self._handle_ssl_validation()
# Validate CP shared KV constraints early (before dummy-model short-circuit).
self._handle_cp_shared_kv_validation()
# Normalize load balancing defaults early (before dummy-model short-circuit).
self._handle_load_balance_method()
if self.model_path.lower() in ["none", "dummy"]:
# Skip for dummy models
return