diff --git a/vllm/config/compilation.py b/vllm/config/compilation.py
index 657c430049f8..17a063d0c233 100644
--- a/vllm/config/compilation.py
+++ b/vllm/config/compilation.py
@@ -498,6 +498,16 @@ def __post_init__(self) -> None:
         if isinstance(self.pass_config, dict):
             self.pass_config = PassConfig(**self.pass_config)
 
+        if (
+            is_torch_equal_or_newer("2.9.0.dev")
+            and "combo_kernels" not in self.inductor_compile_config
+            and "benchmark_combo_kernel" not in self.inductor_compile_config
+        ):
+            # use horizontal fusion, which is useful for fusing qk-norm and
+            # qk-rope when query and key have different shapes.
+            self.inductor_compile_config["combo_kernels"] = True
+            self.inductor_compile_config["benchmark_combo_kernel"] = True
+
         # migrate the deprecated flags
         if not self.use_cudagraph:
             logger.warning(