[EPD] Fix HTTP dispatch lock blocking cross-request encoder batching (#31275)

Co-authored-by: 晟海 <huangtingwei.htw@antgroup.com>

python/sglang/srt/disaggregation/encode_server.py

@@ -3880,9 +3880,6 @@ async def handle_encode_request(request: dict):
             encoder.background_tasks.add(task)
             task.add_done_callback(encoder.background_tasks.discard)
 
-        # broadcast request, lock together with rank0 await so NCCL
-        # launch order matches the ZMQ dispatch order rank>0 sees.
-        async with encoder.encode_dispatch_lock:
         request.update({"enter_time": time.time()})
         modality = Modality.from_str(request["modality"])
         if time_stats_json:
@@ -3912,6 +3909,9 @@ async def handle_encode_request(request: dict):
                     },
                 )
         else:
+            # Lock direct dispatch together with rank0 await so its NCCL launch
+            # order matches the ZMQ dispatch order rank>0 sees.
+            async with encoder.encode_dispatch_lock:
                 for socket in send_sockets:
                     sock_send(socket, wrap_as_pickle(request))
                 nbytes, embedding_len, embedding_dim, error_msg, error_code = (