[PD-HiCache]: Support Async Offloading KVCache In Decode Side (#10192)

Signed-off-by: Shangming Cai <csmthu@gmail.com>
Co-authored-by: Shangming Cai <csmthu@gmail.com>

python/sglang/srt/managers/scheduler.py

@@ -44,6 +44,9 @@ from sglang.srt.disaggregation.decode import (
     DecodeTransferQueue,
     SchedulerDisaggregationDecodeMixin,
 )
+from sglang.srt.disaggregation.decode_kvcache_offload_manager import (
+    DecodeKVCacheOffloadManager,
+)
 from sglang.srt.disaggregation.prefill import (
     PrefillBootstrapQueue,
     SchedulerDisaggregationPrefillMixin,
@@ -755,6 +758,24 @@ class Scheduler(
                     eviction_policy=server_args.radix_eviction_policy,
                 )
 
+        if (
+            server_args.disaggregation_mode == "decode"
+            and server_args.disaggregation_decode_enable_offload_kvcache
+        ):
+            self.decode_offload_manager = DecodeKVCacheOffloadManager(
+                req_to_token_pool=self.req_to_token_pool,
+                token_to_kv_pool_allocator=self.token_to_kv_pool_allocator,
+                tp_group=(
+                    self.attn_tp_cpu_group
+                    if self.server_args.enable_dp_attention
+                    else self.tp_cpu_group
+                ),
+                tree_cache=self.tree_cache,
+                server_args=self.server_args,
+            )
+        else:
+            self.decode_offload_manager = None
+
         self.decode_mem_cache_buf_multiplier = (
             1
             if self.spec_algorithm.is_none()