[Hybrid]: Decouple Kernel Block Size from KV Page Size (#24486)

Signed-off-by: lizhiyuan <uniartisan2017@gmail.com>
Signed-off-by: Zhiyuan Li <uniartisan2017@gmail.com>

vllm/v1/attention/backends/rocm_aiter_fa.py

@@ -3,7 +3,7 @@
 """Attention layer with AiterFlashAttention."""
 
 from dataclasses import dataclass
-from typing import Optional
+from typing import Optional, Union
 
 import torch
 
@@ -12,6 +12,7 @@ from vllm.attention.backends.abstract import (
     AttentionImpl,
     AttentionMetadata,
     AttentionType,
+    MultipleOf,
 )
 from vllm.config import VllmConfig
 from vllm.logger import init_logger
@@ -359,6 +360,10 @@ class AiterFlashAttentionBackend(AttentionBackend):
     def get_supported_head_sizes(cls) -> list[int]:
         return [64, 128, 256]
 
+    @staticmethod
+    def get_supported_kernel_block_size() -> list[Union[int, MultipleOf]]:
+        return [MultipleOf(16)]
+
     @classmethod
     def validate_head_size(cls, head_size: int) -> None:
         supported_head_sizes = cls.get_supported_head_sizes()