Aaron Shi comments

Results 52 comments of


                                            Aaron Shi

Automated submodule update: kineto

Fix has landed in Kineto, will update this submodule now.

[WIP] Raise RuntimeError when models are missing get_module() implementation

This change requires Issue #567 to be fixed.

[WIP] Raise RuntimeError when models are missing get_module() implementation

@xuzhao9 - I'm wondering if we should skip tests that don't have this implemented, so we can enforce this?

Module view did not show up

Hi, I believe the module view is built on the older stack trace information. To enable it, you may need to collect your profiles using ``` from torch._C._profiler import _ExperimentalConfig...

Roctracer: Device ID is off by 2, manually start from 0

Closing this as we may have a better fix via: https://github.com/pytorch/kineto/pull/926. prototyping it.

Is it able to profile for Gloo backend and distributed CPU settings?

What issues did you run into when profiling a multi-CPU host?

datapipe profiler showing too much information

There is no functionality for this now, let's continue the discussion about this internally (on your post). Thank you.

[WIP] Raise RuntimeError when models are missing get_module() implementation

Closing old prs

Step time in Overview and in Trace don't match

There may be a bug causing a difference between the Tensorboard plugin and trace annotation. Could you please provide a screenshot of what you are seeing in both places?

add simple test for nccl metadata

I guess USE_DISTRIBUTED and USE_C10D are checking different things, would like to get some eyes from distributed team on whether this is okay