Different numbers of outputs result in inconsistent running results (original: 不同个数的输出导致运行结果不一致)

Description

Auto-translated

(Part of the linear attention code)

hidden_states = query_KV / query_Z
return hidden_states, query_KV, query_Z

and

hidden_states = query_KV / query_Z
return hidden_states

When converting to TensorRT from ONNX, the results are different between the two above methods. The first is correct, while the second results in NaN; it seems that outputting the intermediate states affects the correctness of the results. How can this issue be resolved?

Further testing shows: This issue occurs with TensorRT 10.7, but TensorRT 10.6 works correctly.

Original

（部分linear attetion代码）
hidden_states = query_KV / query_Z
return hidden_states, query_KV, query_Z

和

hidden_states = query_KV / query_Z
return hidden_states

上面两者方式onnx转tensorrt时，两者结果不一样，前者是正确的，后者会出现nan;
这相当于输出了中间状态会导致结果的正确性，该怎么解决这种问题哇？

后面测试：tensorrt 10.7会出现这种问题，10.6是正确的