Clarification about operation producing multiple outputs

[jerryyiransun]

according to #8, there exists some operations that could produce multiple outputs and one of the examples is shown in mlsys-2026-13.json. I wasn't able to find an output in this specific benchmark that had a length more than 1. I would appreciate if we can get some clarification on this. Thanks!

@yarongmu-google

[Richard1688Sun]

@yarongmu-google additionally, could a subgraph produce more than 1 tensor output? If it does, how will the Solution specify that at the end of a subgraph, multiple tensors output write back to slow memory?

[yarongmu-google]

Thanks for the questions.

@jerryyiransun : I can see multiple cases where multiple outputs are present in 13 - can you please doublecheck?

@Richard1688Sun: yes, a subgraph could produce more than 1 tensor output. Since these are subgraph outputs, per the spec, the hardware will automatically evict all to the slow memory, unless you explicitly retain any of them.

[yarongmu-google]

I will resolve this for now. Please open a new bug, reference this one, if the above doesn't make sense.