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 13, 09:30 UTC
[XPU]Fix crash due to removed VLLM_USE_V1 attribute (#28520)
Signed-off-by: chaojun-zhang <chaojun.zhang@intel.com>
Chaojun Zhang committed
Nov 12, 2025 at 18:20 UTC
a4730c1b4fa297848aeb8b81a36c5158b3845e9c
1 file changed
+3
-2
vllm/platforms/xpu.py
+3
-2
@@ -65,7 +65,6 @@ class XPUPlatform(Platform):
65
66
if use_sparse:
67
raise NotImplementedError("Sparse Attention is not supported on XPU.")
68
- use_v1 = envs.VLLM_USE_V1
68
if not use_v1:
69
raise ValueError("XPU backend only supports V1.")
70
if selected_backend == AttentionBackendEnum.TRITON_ATTN:
@@ -115,7 +114,9 @@ class XPUPlatform(Platform):
114
@classmethod
115
def get_vit_attn_backend(
116
cls, head_size: int, dtype: torch.dtype
118
- ) -> AttentionBackendEnum:
117
+ ) -> "AttentionBackendEnum":
118
+ from vllm.attention.backends.registry import AttentionBackendEnum
119
+
120
return AttentionBackendEnum.FLASH_ATTN
121
122
@classmethod