[Feature] Radix Tree in C++ (#7369)
This commit is contained in:
@@ -569,7 +569,23 @@ class Scheduler(
|
||||
page_size=self.page_size,
|
||||
)
|
||||
else:
|
||||
if self.enable_hierarchical_cache:
|
||||
if os.environ.get("SGLANG_EXPERIMENTAL_CPP_RADIX_TREE") == "1":
|
||||
# lazy import to avoid JIT overhead
|
||||
from sglang.srt.mem_cache.radix_cache_cpp import RadixCacheCpp
|
||||
|
||||
self.tree_cache = RadixCacheCpp(
|
||||
disable=False,
|
||||
use_hicache=self.enable_hierarchical_cache,
|
||||
req_to_token_pool=self.req_to_token_pool,
|
||||
token_to_kv_pool=self.token_to_kv_pool_allocator,
|
||||
tp_cache_group=self.tp_cpu_group,
|
||||
page_size=self.page_size,
|
||||
hicache_ratio=server_args.hicache_ratio,
|
||||
hicache_size=server_args.hicache_size,
|
||||
hicache_write_policy=server_args.hicache_write_policy,
|
||||
enable_kv_cache_events=self.enable_kv_cache_events,
|
||||
)
|
||||
elif self.enable_hierarchical_cache:
|
||||
self.tree_cache = HiRadixCache(
|
||||
req_to_token_pool=self.req_to_token_pool,
|
||||
token_to_kv_pool_allocator=self.token_to_kv_pool_allocator,
|
||||
|
||||
Reference in New Issue
Block a user