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 7, 19:00 UTC
[XPU]Fix for Qwen-OMNI crash (#35249)
Signed-off-by: Chendi Xue <chendi.xue@intel.com>
Chendi.Xue committed
Feb 25, 2026 at 09:31 UTC
d72b0be33cdd561e557df1ce5350a14451b9af13
1 file changed
+2
-1
vllm/_xpu_ops.py
+2
-1
@@ -105,9 +105,10 @@ class xpu_ops:
105
assert len(window_size) == 2
106
real_window_size = (window_size[0], window_size[1]) # noqa: F841
107
108
- # In encode attention, v maybe not contiguous and current
108
+ # In encode attention, k and v maybe not contiguous and current
109
# kernel can't handle it
110
if block_table is None:
111
+ k = k.contiguous()
112
v = v.contiguous()
113
return flash_attn_varlen_func(
114
out=out,