[CI] Revert PRs 34818 and 33600 (#34979)

vllm/v1/attention/backend.py

@@ -4,7 +4,7 @@
 from abc import ABC, abstractmethod
 from dataclasses import dataclass, replace
 from enum import Enum
-from typing import TYPE_CHECKING, Any, ClassVar, Generic, Protocol, TypeVar
+from typing import TYPE_CHECKING, Any, ClassVar, Generic, Protocol, TypeVar, get_args
 
 import numpy as np
 import torch
@@ -144,9 +144,15 @@ class AttentionBackend(ABC):
 
     @classmethod
     def supports_block_size(cls, block_size: int | None) -> bool:
+        from vllm.config.cache import BlockSize
+
         if block_size is None:
             return True
 
+        valid_sizes = get_args(BlockSize)
+        if block_size not in valid_sizes:
+            return False
+
         supported_kernel_block_sizes = cls.get_supported_kernel_block_sizes()
         if not supported_kernel_block_sizes:
             return True
@@ -161,17 +167,6 @@ class AttentionBackend(ABC):
                 return True
         return False
 
-    @classmethod
-    def get_preferred_block_size(cls, default_block_size: int = 16) -> int:
-        supported_sizes = cls.get_supported_kernel_block_sizes()
-        if not supported_sizes:
-            return default_block_size
-
-        if cls.supports_block_size(default_block_size):
-            return default_block_size
-
-        return min(s.base if isinstance(s, MultipleOf) else s for s in supported_sizes)
-
     @classmethod
     def is_mla(cls) -> bool:
         return False