A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[XPU] Fix circular import error. (#24927)
Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
Kunshang Ji committed
Sep 16, 2025 at 11:35 UTC
5206ab20ba7477e5457c2e64469590d548fa15e6
1 file changed
+5
-1
vllm/platforms/xpu.py
+5
-1
@@ -9,7 +9,6 @@ import torch
9
import vllm.envs as envs
10
from vllm.logger import init_logger
11
from vllm.utils import DEFAULT_MAX_NUM_BATCHED_TOKENS
12
-from vllm.v1.attention.backends.utils import set_kv_cache_layout
12
13
from .interface import DeviceCapability, Platform, PlatformEnum, _Backend
14
@@ -164,11 +163,16 @@ class XPUPlatform(Platform):
163
vllm_config.scheduler_config.max_num_batched_tokens = max(
164
vllm_config.scheduler_config.max_model_len,
165
DEFAULT_MAX_NUM_BATCHED_TOKENS)
166
+ from vllm.v1.attention.backends.utils import set_kv_cache_layout
167
168
set_kv_cache_layout("NHD")
169
logger.info("Setting VLLM_KV_CACHE_LAYOUT to 'NHD' for XPU; "
170
"only NHD layout is supported by XPU attention kernels.")
171
172
+ @classmethod
173
+ def support_hybrid_kv_cache(cls) -> bool:
174
+ return True
175
+
176
@classmethod
177
def is_pin_memory_available(cls):
178
return True