[Bugfix] Restore support for larger block sizes (#11259)

Signed-off-by: Konrad Zawora <kzawora@habana.ai>

vllm/config.py

@@ -917,6 +917,10 @@ class CacheConfig:
             raise ValueError(
                 "GPU memory utilization must be less than 1.0. Got "
                 f"{self.gpu_memory_utilization}.")
+        if (current_platform.is_cuda() and self.block_size is not None
+                and self.block_size > 32):
+            raise ValueError("CUDA Paged Attention kernel only supports "
+                             f"block sizes up to 32. Got {self.block_size}.")
 
     def _verify_cache_dtype(self) -> None:
         if self.cache_dtype == "auto":