[vLLM IR] add `import_ir_kernels()` to support OOT platforms (#38807)

Signed-off-by: Icey <1790571317@qq.com>

vllm/config/kernel.py

@@ -68,10 +68,13 @@ class IrOpPriorityConfig:
     def set_priority(self):
         """
         Context manager to set the IR op priority for all op members.
-        It also imports vllm.kernels to ensure all implementations are made available.
+        It also imports IR kernel implementations for the current platform
+        to ensure all implementations are made available.
         """
-        import vllm.kernels  # noqa: F401, registers IR op implementations
         from vllm.ir.op import IrOp
+        from vllm.platforms import current_platform
+
+        current_platform.import_ir_kernels()
 
         with contextlib.ExitStack() as stack:
             for field in fields(self):

vllm/platforms/interface.py

@@ -207,6 +207,15 @@ class Platform:
         """
         return cls.simple_compile_backend
 
+    @classmethod
+    def import_ir_kernels(cls) -> None:
+        """
+        The default implementation imports ``vllm.kernels``, which registers
+        the built-in IR op implementations. Out-of-tree (OOT) platforms should
+        override this method to import their own kernel modules.
+        """
+        import vllm.kernels  # noqa: F401
+
     @classmethod
     def device_id_to_physical_device_id(cls, device_id: int):
         # Treat empty device control env var as unset. This is a valid