Thomas Viehmann

We should not add alternative ways to do things much, it is good when there is only one way to achieve things. What's bad with using `thunder.last_traces`?

To my mind, this seems to be fundamentally "memory use" and not "compute perf" if the batch size needed to be lowered.

Thank you for the report, @tfogal If args is empty, this will give an empty list: https://github.com/Lightning-AI/lightning-thunder/blob/50f587d6ff6c17a8f8392c57a0e6a73b0fe298fb/thunder/transforms/cudagraph.py#L36 and so unpacking fails. We should handle empty args by special casing it...

The other trouble we'll run into is that the autograd passes delete the marks of static inputs. This needs fixing and is on our list, but it is quite a...

@tfogal with #1324 fixing the immediate issue, I would close this or assign it back to you for update. Sorry for taking long.

This is another, more prototypical, example the commented line does not work, but it should with a flag (or maybe even by default, so only transform for execution does it...

I'm glad you bring this up! Let's do a technical session on this. As a preview, my thoughts here are that we should make a connection, but a different one:...

To my mind, we conceptually need tags on proxies because properties like "STATIC_MEMORY_LOCATION" or "DONT_SAVE_FOR_BACKWARD" are properties of the proxies. We might have things in the future that tag proxies...

Thinking about @IvanYashchuk 's comment more, I'm warming to the idea that maybe putting a rematerialization tag on the bsym and inheriting it to the subsymbols would be better than...

For the transform compatibility: - Given that the typical use would be custom kernels, I would not overly worry about autocast for at first. - for backward with `op.register_autograd` we...