hicache storage backend mooncake support ascend hixl (#20016)

python/sglang/srt/mem_cache/storage/mooncake_store/mooncake_store.py

@@ -497,7 +497,8 @@ class MooncakeStore(HiCacheStorage, MooncakeBaseStore):
             "page_first",
             "page_first_direct",
             "page_head",
-        ], "mooncake store storage backend only support page first or page first direct layout"
+            "page_first_kv_spilt",
+        ], "mooncake store storage backend only support page first, page first direct, page head and  page_first_kv_spilt layout"
         buffer = self.mem_pool_host.kv_buffer
         try:
             super().register_buffer(buffer)

python/sglang/srt/model_executor/model_runner.py

@@ -449,12 +449,12 @@ class ModelRunner(ModelRunnerKVCacheMixin):
         if self.device == "cpu":
             self.init_threads_binding()
 
-        # Initialize MooncakeTransferEngine
-        self.init_shared_mooncake_transfer_engine()
-
         # Get available memory before model loading
         pre_model_load_memory = self.init_torch_distributed()
 
+        # Initialize MooncakeTransferEngine
+        self.init_shared_mooncake_transfer_engine()
+
         # Init forward stream for overlap schedule
         self.forward_stream = torch.get_device_module(self.device).Stream()