Add feature: priority

Signed-off-by: Jing Wang <jingwang96@qq.com>

vllm_ascend/device_allocator/camem.py

@@ -62,7 +62,10 @@ try:
             python_create_and_map_offload as python_create_and_map,python_unmap_and_release_offload as python_unmap_and_release,
             python_get_mem_info_offload as python_get_mem_info,
             python_try_lock_gpu_offload as python_try_lock_gpu,
-            python_unlock_gpu_offload as python_unlock_gpu
+            python_unlock_gpu_offload as python_unlock_gpu,
+            python_start_wait_offload as python_start_wait,
+            python_cancel_wait_offload as python_cancel_wait,
+            python_has_higher_priority_waiter_offload as python_has_higher_priority_waiter
         )
     else:
         from vllm_ascend.vllm_ascend_C import (  # type: ignore # noqa: F401
@@ -73,6 +76,9 @@ try:
         python_get_mem_info = None
         python_try_lock_gpu = None
         python_unlock_gpu = None
+        python_start_wait = None
+        python_cancel_wait = None
+        python_has_higher_priority_waiter = None
 
     lib_name = find_loaded_library("vllm_ascend_C")
     camem_available = True
@@ -84,6 +90,9 @@ except ImportError as e:
     python_get_mem_info = None
     python_try_lock_gpu = None
     python_unlock_gpu = None
+    python_start_wait = None
+    python_cancel_wait = None
+    python_has_higher_priority_waiter = None
     lib_name = None
     libcudart = None
 
@@ -306,15 +315,37 @@ class CaMemAllocator:
             return False, False
 
     def _vnpu_lock_gpu(self) -> bool:
+        is_waiting = False
         while True:
             success, _ = self.vnpu_try_lock_gpu()
             if success:
+                if is_waiting:
+                    self.vnpu_cancel_wait()
                 return True
+            else:
+                if not is_waiting:
+                    self.vnpu_start_wait()
+                    is_waiting = True
+                self.vnpu_unlock_gpu(keep_wait=True)
+
             time.sleep(0.001)
 
-    def vnpu_unlock_gpu(self):
+    def vnpu_unlock_gpu(self, keep_wait: bool = False):
         if python_unlock_gpu:
-            python_unlock_gpu()
+            python_unlock_gpu(keep_wait)
+
+    def vnpu_start_wait(self) -> None:
+        if python_start_wait:
+            python_start_wait()
+
+    def vnpu_cancel_wait(self) -> None:
+        if python_cancel_wait:
+            python_cancel_wait()
+
+    def vnpu_has_higher_priority_waiter(self) -> bool:
+        if python_has_higher_priority_waiter:
+            return python_has_higher_priority_waiter()
+        return False
 
     def get_pool_mem_info(self) -> tuple[int, int]:
         """