This commit is contained in:
@@ -388,9 +388,9 @@ class FlexKVRadixCache(RadixCache):
|
|||||||
return
|
return
|
||||||
|
|
||||||
# Compute the committed prefix mirroring LMCRadixCache's logic.
|
# Compute the committed prefix mirroring LMCRadixCache's logic.
|
||||||
from sglang.srt.server_args import get_global_server_args
|
from sglang.srt.runtime_context import get_server_args
|
||||||
|
|
||||||
global_server_args = get_global_server_args()
|
global_server_args = get_server_args()
|
||||||
topk = global_server_args.speculative_eagle_topk
|
topk = global_server_args.speculative_eagle_topk
|
||||||
enable_kv_committed_len = topk is None or topk == 1
|
enable_kv_committed_len = topk is None or topk == 1
|
||||||
if enable_kv_committed_len:
|
if enable_kv_committed_len:
|
||||||
|
|||||||
@@ -35,6 +35,7 @@ def _make_ctx(
|
|||||||
server_args.radix_cache_backend = backend
|
server_args.radix_cache_backend = backend
|
||||||
server_args.enable_streaming_session = enable_streaming
|
server_args.enable_streaming_session = enable_streaming
|
||||||
server_args.enable_lmcache = enable_lmcache
|
server_args.enable_lmcache = enable_lmcache
|
||||||
|
server_args.enable_flexkv = False
|
||||||
return TreeCacheBuildContext(
|
return TreeCacheBuildContext(
|
||||||
server_args=server_args,
|
server_args=server_args,
|
||||||
params=MagicMock(),
|
params=MagicMock(),
|
||||||
|
|||||||
Reference in New Issue
Block a user