Reported slowness due in variable expansion

[connor4312]

Isidor pointed out a tweet that mentions

S Code’s debugger would lock up whenever I stepped past a line like: S=tensor.untyped_storage()... the debugger window that automatically displays local variables started a dump of over a billion elements from the debug server before it would process any other commands.

I have not attempted to repro myself, but there are a few ways in the debug protocol that cases like this can be handled:

• Set lazy: true on the VariablePresentationHints of potentially expensive variables so that they aren't expanded without user action
• Use indexedVariables to allow the client to request a small number of a variable's contained data and page through it

Even if it's unavoidable that a giant variable's data is loaded into the debug adapter, it should use these mechanisms to avoid overloading clients and slowing down interaction on the protocol.

Let me know if there's anything I can help with from the VS Code side :)

[rchiodo]

I think we likely need to specialize for tensors. We've had other issues with the repr for them too:
#1525

[xuanhua]

Hi, glad to see this thread, on debugging large language model, I also face the same issue. Each tensor contains huge amount of data. Each step could cost lot of time on transfering these variables' values. On network of limited speed, this is especially painful for remote debugging LLM.