[PD] Support fake decode for PD disaggregation without prefill node (#14628)

Co-authored-by: sunhailiang <sunhailiang@baidu.com>

python/sglang/srt/disaggregation/decode.py

@@ -323,7 +323,11 @@ class DecodePreallocQueue:
             req.retraction_mb_id = None
             self.retracted_queue.append(req)
         else:
-            if req.bootstrap_host == FAKE_BOOTSTRAP_HOST:
+            # 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
+            ):
                 kv_receiver_class = get_kv_class(
                     TransferBackend.FAKE, KVClassType.RECEIVER
                 )

python/sglang/srt/managers/data_parallel_controller.py

@@ -489,6 +489,16 @@ class DataParallelController:
                 self.workers
             )
         else:
+            # Set default bootstrap_room if in FAKE auto mode and room is None
+            if (
+                req.bootstrap_room is None
+                and self.server_args.disaggregation_decode_enable_fake_auto
+            ):
+                req.bootstrap_room = self.round_robin_counter
+                self.round_robin_counter = (self.round_robin_counter + 1) % len(
+                    self.workers
+                )
+
             assert (
                 req.bootstrap_room is not None
             ), "req.bootstrap_room should not be None. Do not send requests directly to prefill or decode instances, but send to the router instead."

python/sglang/srt/server_args.py

@@ -602,6 +602,8 @@ class ServerArgs:
     disaggregation_prefill_pp: Optional[int] = 1
     disaggregation_ib_device: Optional[str] = None
     disaggregation_decode_enable_offload_kvcache: bool = False
+    # Enable auto FAKE mode for decode node testing, no need to pass bootstrap_host in request
+    disaggregation_decode_enable_fake_auto: bool = False
     num_reserved_decode_tokens: int = 512  # used for decode kv cache offload in PD
     # FIXME: hack to reduce ITL when decode bs is small
     disaggregation_decode_polling_interval: int = 1
@@ -4261,6 +4263,12 @@ class ServerArgs:
             action="store_true",
             help="Enable async KV cache offloading on decode server (PD mode).",
         )
+        parser.add_argument(
+            "--disaggregation-decode-enable-fake-auto",
+            action="store_true",
+            help="Auto enable FAKE mode for decode node testing, "
+            "no need to pass bootstrap_host and bootstrap_room in request.",
+        )
         parser.add_argument(
             "--num-reserved-decode-tokens",
             type=int,