diff --git a/vllm/model_executor/layers/rotary_embedding/xdrope.py b/vllm/model_executor/layers/rotary_embedding/xdrope.py index dab7aad975..a799b10a19 100644 --- a/vllm/model_executor/layers/rotary_embedding/xdrope.py +++ b/vllm/model_executor/layers/rotary_embedding/xdrope.py @@ -137,6 +137,18 @@ class XDRotaryEmbedding(DynamicNTKAlphaRotaryEmbedding): key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape) return query, key + def forward_xpu( + self, + positions: torch.Tensor, + query: torch.Tensor, + key: torch.Tensor | None = None, + offsets: torch.Tensor | None = None, + ) -> tuple[torch.Tensor, torch.Tensor | None]: + # No fused XPU kernel for XD-sectioned RoPE; the base + # RotaryEmbedding.forward_xpu calls the generic C++ kernel, which + # rejects the 2-D [4, num_tokens] xdrope positions. Use the native path. + return self.forward_native(positions, query, key, offsets) + @staticmethod def get_next_input_positions( context_len: int,