diff --git a/vllm/model_executor/layers/rotary_embedding/ernie45_vl_rope.py b/vllm/model_executor/layers/rotary_embedding/ernie45_vl_rope.py index 2eda63a34a..ec0c5a3286 100644 --- a/vllm/model_executor/layers/rotary_embedding/ernie45_vl_rope.py +++ b/vllm/model_executor/layers/rotary_embedding/ernie45_vl_rope.py @@ -80,3 +80,14 @@ class Ernie4_5_VLRotaryEmbedding(MRotaryEmbedding): key: torch.Tensor | None = None, ) -> tuple[torch.Tensor, torch.Tensor | None]: return self.forward_native(positions, query, key) + + def forward_xpu( # type: ignore[override] + self, + positions: torch.Tensor, + query: torch.Tensor, + key: torch.Tensor | None = None, + ) -> tuple[torch.Tensor, torch.Tensor | None]: + # No fused XPU kernel for this 3D t/h/w rope; base + # MRotaryEmbedding.forward_xpu forwards an extra `offsets` arg that + # this class's forward_cuda override doesn't accept. Use native path. + return self.forward_native(positions, query, key)