When my model contains multiple groups of outputs(RNNs), I can create output fetch and cut the necessary graph for session run on Tensorflow, but how can I fulfill such a model on TVM?
Unlike TensorFlow, in TVM/Relay you return values directly from a function instead of fetching them during a Session run. To return multiple values in Relay you can use a tuple. Check out https://docs.tvm.ai/langref/relay_adt.html#implementing-neural-nets-using-adts for an example of an RNN in Relay.