llama4_vision_rope: add HIP override to accept (q, k) and avoid (positions, q, k) mismatch (vllm-project#26790)

Signed-off-by: Huamin Li <[email protected]>

Author: hl475

vllm/model_executor/layers/rotary_embedding/llama4_vision_rope.py

@@ -78,3 +78,10 @@ def forward_cuda(  # type: ignore[override]
         key: torch.Tensor | None = None,
     ) -> tuple[torch.Tensor, torch.Tensor | None]:
         return self.forward_native(query, key)
+
+    def forward_hip(  # type: ignore[override]
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor | None = None,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        return self.forward_native(query, key)