[refactor] Adopt get_parallel() everywhere and close out the parallel wrapper surface (#30492)

python/sglang/srt/managers/cache_controller.py

@@ -33,19 +33,12 @@ if TYPE_CHECKING:
     from sglang.srt.mem_cache.allocator import BaseTokenToKVPoolAllocator
     from sglang.srt.mem_cache.pool_host import HostKVCache
 
-from sglang.srt.distributed import (
-    get_pipeline_model_parallel_rank,
-    get_pipeline_model_parallel_world_size,
-    get_tensor_model_parallel_rank,
-    get_tensor_model_parallel_world_size,
-)
 from sglang.srt.layers.dp_attention import (
     get_attention_dp_rank,
-    get_attention_tp_rank,
-    get_attention_tp_size,
     is_dp_attention_enabled,
 )
 from sglang.srt.mem_cache.memory_pool import MLATokenToKVPool
+from sglang.srt.runtime_context import get_parallel
 from sglang.srt.utils import get_device_module
 
 logger = logging.getLogger(__name__)
@@ -569,16 +562,16 @@ class HiCacheController:
             storage_backend_extra_config = {}
 
         if is_dp_attention_enabled():
-            self.tp_rank = get_attention_tp_rank()
-            self.tp_size = get_attention_tp_size()
+            self.tp_rank = get_parallel().attn_tp_rank
+            self.tp_size = get_parallel().attn_tp_size
             self.dp_rank = get_attention_dp_rank()
         else:
-            self.tp_rank = get_tensor_model_parallel_rank()
-            self.tp_size = get_tensor_model_parallel_world_size()
+            self.tp_rank = get_parallel().tp_rank
+            self.tp_size = get_parallel().tp_size
             self.dp_rank = 0
 
-        self.pp_rank = get_pipeline_model_parallel_rank()
-        self.pp_size = get_pipeline_model_parallel_world_size()
+        self.pp_rank = get_parallel().pp_rank
+        self.pp_size = get_parallel().pp_size
 
         # Currently, NPUMLATokenToKVPool is the subclass of MLATokenToKVPool.
         # DeepSeekV4TokenToKVPool has compressed MLA-style rank-replicated cache