[Misc] Refactor get_kv_cache_spec into AttentionLayerBase (#26587)

Signed-off-by: NickLucche <nlucches@redhat.com>

vllm/model_executor/layers/attention_layer_base.py

@@ -5,6 +5,9 @@
 from abc import ABC, abstractmethod
 from typing import TYPE_CHECKING
 
+from vllm.config import VllmConfig
+from vllm.v1.kv_cache_interface import KVCacheSpec
+
 if TYPE_CHECKING:
     from vllm.attention.backends.abstract import AttentionBackend
 
@@ -22,3 +25,11 @@ class AttentionLayerBase(ABC):
     def get_attn_backend(self) -> type["AttentionBackend"]:
         """Get the attention backend class for this layer."""
         pass
+
+    @abstractmethod
+    def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec | None:
+        """
+        Get the KV cache spec for this layer.
+        May be None if the layer does not need KV cache.
+        """
+        pass