[XPU] Route mm_prefix models to Triton attention backend (#47688)
Signed-off-by: zhenwei-intel <zhenwei.liu@intel.com>
liuzhenwei committed
Jul 6, 2026 at 16:52 UTC
736f1a590714d4f1286fcc764439ad39cca2e8f5
1 file changed
+9
vllm/platforms/xpu.py
+9
@@ -148,6 +148,15 @@ class XPUPlatform(Platform):
148
if selected_backend == AttentionBackendEnum.TRITON_ATTN:
149
logger.info_once("Using Triton backend.")
150
return AttentionBackendEnum.TRITON_ATTN.get_path()
151
+ elif attn_selector_config.use_mm_prefix:
152
+ # Flash Attention on XPU has no FA4 kernel, so it cannot apply the
153
+ # multimodal prefix-LM bidirectional mask. Fall back to Triton
154
+ # Attention, which supports mm_prefix.
155
+ logger.warning_once(
156
+ "Flash Attention on XPU does not support multimodal prefix-LM "
157
+ "attention. Falling back to Triton Attention backend."
158
+ )
159
+ return AttentionBackendEnum.TRITON_ATTN.get_path()
160
elif dtype == torch.float32:
161
logger.warning_once(
162
"Flash Attention on XPU does not support float32 dtype. "