refactor FAKE transfer backend and remove --disaggregation-decode-enable-fake-auto parameter (#18345)

python/sglang/srt/disaggregation/ascend/conn.py

@@ -26,7 +26,6 @@ class AscendKVManager(MooncakeKVManager):
             hostname=local_ip,
             npu_id=self.kv_args.gpu_id,
             disaggregation_mode=self.disaggregation_mode,
-            disaggregation_decode_enable_fake_auto=self.disaggregation_decode_enable_fake_auto,
         )
 
     def register_buffer_to_engine(self):

python/sglang/srt/disaggregation/ascend/transfer_engine.py

@@ -27,7 +27,6 @@ class AscendTransferEngine(MooncakeTransferEngine):
         hostname: str,
         npu_id: int,
         disaggregation_mode: DisaggregationMode,
-        disaggregation_decode_enable_fake_auto: bool,
     ):
         if import_error is not None:
             logger.warning(
@@ -38,9 +37,6 @@ class AscendTransferEngine(MooncakeTransferEngine):
         self.engine = TransferEngine()
         self.hostname = hostname
         self.npu_id = npu_id
-        self.disaggregation_decode_enable_fake_auto = (
-            disaggregation_decode_enable_fake_auto
-        )
 
         # Centralized storage address of the AscendTransferEngine
         self.store_url = os.getenv("ASCEND_MF_STORE_URL")
@@ -76,12 +72,6 @@ class AscendTransferEngine(MooncakeTransferEngine):
                 output_tensor_list, tmp_tensor, group=get_tp_group().device_group
             )
         """Initialize the ascend transfer instance."""
-        if self.disaggregation_decode_enable_fake_auto:
-            logger.info(
-                "Ascend Transfer Engine is not initialized in decode fake transfer mode."
-            )
-            return
-
         ret_value = self.engine.initialize(
             self.store_url, self.session_id, self.role, self.npu_id, trans_op_type
         )

python/sglang/srt/disaggregation/common/conn.py

@@ -52,9 +52,6 @@ class CommonKVManager(BaseKVManager):
         self.kv_args = args
         self.is_mla_backend = is_mla_backend
         self.disaggregation_mode = disaggregation_mode
-        self.disaggregation_decode_enable_fake_auto = (
-            server_args.disaggregation_decode_enable_fake_auto
-        )
         # for p/d multi node infer
         self.bootstrap_host = server_args.host
         self.bootstrap_port = server_args.disaggregation_bootstrap_port

python/sglang/srt/disaggregation/decode.py

@@ -344,7 +344,7 @@ class DecodePreallocQueue:
             # Auto enable FAKE mode if configured
             if req.bootstrap_host == FAKE_BOOTSTRAP_HOST or (
                 req.bootstrap_host is None
-                and self.scheduler.server_args.disaggregation_decode_enable_fake_auto
+                and self.scheduler.server_args.disaggregation_transfer_backend == "fake"
             ):
                 kv_receiver_class = get_kv_class(
                     TransferBackend.FAKE, KVClassType.RECEIVER
@@ -759,7 +759,7 @@ class DecodeTransferQueue:
 
         if decode_req.req.bootstrap_host == FAKE_BOOTSTRAP_HOST or (
             decode_req.req.bootstrap_host is None
-            and self.scheduler.server_args.disaggregation_decode_enable_fake_auto
+            and self.scheduler.server_args.disaggregation_transfer_backend == "fake"
         ):
             # Warm up or fake transfer mode
             pass

python/sglang/srt/disaggregation/fake/__init__.py

@@ -1 +1,5 @@
-from sglang.srt.disaggregation.fake.conn import FakeKVReceiver, FakeKVSender
+from sglang.srt.disaggregation.fake.conn import (
+    FakeKVManager,
+    FakeKVReceiver,
+    FakeKVSender,
+)

python/sglang/srt/disaggregation/fake/conn.py

@@ -8,13 +8,27 @@ from sglang.srt.disaggregation.base.conn import (
     BaseKVManager,
     BaseKVReceiver,
     BaseKVSender,
+    KVArgs,
     KVPoll,
 )
+from sglang.srt.disaggregation.utils import DisaggregationMode
+from sglang.srt.server_args import ServerArgs
 
 logger = logging.getLogger(__name__)
 
 
-# For warmup reqs, we don't kv transfer, we use the fake sender and receiver
+# For warmup reqs, we don't kv transfer, we use the fake manager, sender and receiver
+class FakeKVManager(BaseKVManager):
+    def __init__(
+        self,
+        args: KVArgs,
+        disaggregation_mode: DisaggregationMode,
+        server_args: ServerArgs,
+        is_mla_backend: Optional[bool] = False,
+    ):
+        super().__init__(args, disaggregation_mode, server_args, is_mla_backend)
+
+
 class FakeKVSender(BaseKVSender):
     def __init__(
         self,

python/sglang/srt/disaggregation/utils.py

@@ -335,10 +335,15 @@ def get_kv_class(
         return class_mapping.get(class_type)
     elif transfer_backend == TransferBackend.FAKE:
         from sglang.srt.disaggregation.base import KVArgs
-        from sglang.srt.disaggregation.fake import FakeKVReceiver, FakeKVSender
+        from sglang.srt.disaggregation.fake import (
+            FakeKVManager,
+            FakeKVReceiver,
+            FakeKVSender,
+        )
 
         class_mapping = {
             KVClassType.KVARGS: KVArgs,
+            KVClassType.MANAGER: FakeKVManager,
             KVClassType.SENDER: FakeKVSender,
             KVClassType.RECEIVER: (FakeKVReceiver),
         }