[Relay] Dump all ops' outputs: VM or Graph runtime?

Hi all, For the purpose of debugging output, dump ops’ output tensors could be a very useful mechanism.

Does VM or graph runtime support this feature?

We have the debug graph executor, which has the method debug_get_output (tvm.contrib.graph_executor — tvm 0.8.dev0 documentation). This will dump all tensors from all the ops. Note that it is currently broken, but [Graph Executor Debugger] Fix parameter dump by mehrdadh · Pull Request #7903 · apache/tvm · GitHub should fix it.

One more question: For this PR,does it support BYOC self-defined fused pattern???

In BYOC, customized hardware can define its own fused pattern, however, it won’t make any change to the computation part.

So I am wondering if this graph executor debugger support it. Or, is there any de-fused pass that can revert the customized fusion patterns which also makes sense.

This PR does not dump un-fused results. As far as I know there is no de-fuse pass.