Thomas Viehmann

You mean something along the lines of Lightning-AI/lit-thunder-LEGACY#1779 ?

Thinking about this more, it's not so clear that it is a reasonable implementation though, because the JITed things will need backward hook calls generated for them because we don't...

Thank you for the PR @kiya00 ! I can see how allowing the user to query for this is good, but I'm a bit hesitant about this approach. Could the...

I'm not quite understanding this. To my mind, a hacky way to get the auto symbols in the trace would be `{bsym.sym for bsym in lt.bound_symbols if bsym.sym.meta.__code__ in thunder.torch.meta_adaptor.__code__.co_consts}`...

>If we think that eagerly collecting statistics is burdensome, then we could add some architecture to support the lazy collection of statistics. This lazy collection could work by reviewing the...

We will have a discussion on this tomorrow (Aug 21) in the Thunder Tech sessions. Please ping me if you want an invite. > how do checks in the prologue...

Outcomes of the design discussion: (https://docs.google.com/presentation/d/1JhmJqikAneYDv4S-g9rcAOS9xy35FztydE-PzxFhi3w/edit?usp=sharing) - tag proxies as static memory location - tag operators as safe or not - bubble up constraints like strides? (maybe the dynamic shape...

The goal of the renaming has been to use more user variable names, not to get consecutive generic names. Is that what you want? (Also, with any bit of luck,...

I htink what happens often is that it tries to rename something to an internal name ("tos" etc.) and then finds that that is taken already for whatever reason. With...

The error sounds like either device or dtype don't match between the two. Maybe you could insert a debug print to check what's going on. Note that this is not...