Stas Bekman

That sounds like a wise plan of action to me, @tohtana - thank you for working on this. You will make many users very happy!

I was only able to test a few cases, I will continue testing/studying your work tomorrow, @tohtana - thank you for working on it.

Apologies that tomorrow turned into 2 weeks, @tohtana - I was trying to find a big enough slot to give it a serious attention and just wasn't manifesting. I was...

Amazing work, @tohtana! I'm looking forward to trying it out Here is a quick feedback: Could we please flip `disable` to `enabled` so that the logic is consistent with other...

tried it out and the compiled engine doesn't seem to forward some (all?) custom methods to the unwrapped model, e.g. it's failing: ``` [28:7]: File "/data/env/lib/repos/retro-llama/tr043-dawn-llama-3/DeepSpeed/deepspeed/runtime/engine.py", line 468, in __getattr__...

I hacked around it via `model.module.method...` and then I get many warnings and errors with the `inductor` backend and then it fails. I have attached the log. This is just...

If I disable the ds profiler than it runs despite the compilation errors/warnings - same log as in the previous comment, other than the last traceback where it crashes.

I'm also observing a very strange behavior of performance cycling: the tflops go like this per iteration: 196, 196, 192, 196, 196, 192, 196, 196, 192, - 2 fast one...

That's totally understandable, Masahiro. Tunji made that clear when he tagged me. If it's too early to provide feedback please ping me when you're ready for it. ---------------- `disable` vs...

- please note that it's `enabled` that DS uses everywhere else and not `enable` - wrt other options I'd say - use the minimal amount of options - 1. let's...