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 10, 14:00 UTC
[XPU]fallback to TRITON_ATTN on xpu when use float32 dtype (#31762)
Signed-off-by: sihao.li <sihao.li@intel.com>
sihao_li committed
Jan 7, 2026 at 16:10 UTC
59fe6f298e16ee8d2f54e2567b76516807a4733b
1 file changed
+7
vllm/platforms/xpu.py
+7
@@ -52,11 +52,18 @@ class XPUPlatform(Platform):
52
"only NHD layout is supported by XPU attention kernels."
53
)
54
55
+ dtype = attn_selector_config.dtype
56
if attn_selector_config.use_sparse:
57
raise NotImplementedError("Sparse Attention is not supported on XPU.")
58
if selected_backend == AttentionBackendEnum.TRITON_ATTN:
59
logger.info_once("Using Triton backend.")
60
return AttentionBackendEnum.TRITON_ATTN.get_path()
61
+ elif dtype == torch.float32:
62
+ logger.warning_once(
63
+ "Flash Attention on XPU does not support float32 dtype. "
64
+ "Falling back to Triton Attention backend."
65
+ )
66
+ return AttentionBackendEnum.TRITON_ATTN.get_path()
67
elif selected_backend == AttentionBackendEnum.FLASH_ATTN:
68
logger.info_once("Using Flash Attention backend.")
69
return AttentionBackendEnum.FLASH_ATTN.get_path()