[XPU] Fix circular import error. (vllm-project#24927)

Signed-off-by: Kunshang Ji <[email protected]>

Author: jikunshang

vllm/platforms/xpu.py

@@ -9,7 +9,6 @@
 import vllm.envs as envs
 from vllm.logger import init_logger
 from vllm.utils import DEFAULT_MAX_NUM_BATCHED_TOKENS
-from vllm.v1.attention.backends.utils import set_kv_cache_layout
 
 from .interface import DeviceCapability, Platform, PlatformEnum, _Backend
 
@@ -164,11 +163,16 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
             vllm_config.scheduler_config.max_num_batched_tokens = max(
                 vllm_config.scheduler_config.max_model_len,
                 DEFAULT_MAX_NUM_BATCHED_TOKENS)
+        from vllm.v1.attention.backends.utils import set_kv_cache_layout
 
         set_kv_cache_layout("NHD")
         logger.info("Setting VLLM_KV_CACHE_LAYOUT to 'NHD' for XPU; "
                     "only NHD layout is supported by XPU attention kernels.")
 
+    @classmethod
+    def support_hybrid_kv_cache(cls) -> bool:
+        return True
+
     @classmethod
     def is_pin_memory_available(cls):
         return True