[NPU] disaggregation_decode_enable_fake_auto parameter adaptation (#17811)

Co-authored-by: ronnie_zheng <zl19940307@163.com>
Co-authored-by: Even Zhou <even.y.zhou@outlook.com>

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

@@ -26,6 +26,7 @@ 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

@@ -21,7 +21,11 @@ logger = logging.getLogger(__name__)
 class AscendTransferEngine(MooncakeTransferEngine):
 
     def __init__(
-        self, hostname: str, npu_id: int, disaggregation_mode: DisaggregationMode
+        self,
+        hostname: str,
+        npu_id: int,
+        disaggregation_mode: DisaggregationMode,
+        disaggregation_decode_enable_fake_auto: bool,
     ):
         if import_error is not None:
             logger.warning(
@@ -32,6 +36,9 @@ 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")
@@ -67,6 +74,12 @@ 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

@@ -52,6 +52,9 @@ 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