Conversation
Signed-off-by: Eric Kerfoot <[email protected]>
Signed-off-by: Eric Kerfoot <[email protected]>
Signed-off-by: eric <[email protected]>
Signed-off-by: Eric Kerfoot <[email protected]>
|
Add @binliunls as he was doing something similar. Thanks in advance. |
Signed-off-by: Eric Kerfoot <[email protected]>
Signed-off-by: Eric Kerfoot <[email protected]>
Signed-off-by: Eric Kerfoot <[email protected]>
Signed-off-by: Eric Kerfoot <[email protected]>
|
Ok, I am glad to review this PR. |
Signed-off-by: Eric Kerfoot <[email protected]>
|
There's something really odd about multiprocess behaviour in Darwin and in one the Nvidia images used for testing (nvcr.io/nvidia/pytorch:22.07-py3). I tried reproducing with Pytorch 1.13 but the test passes locally on Ubuntu 20.04. I think I need to disable multiprocess tracing test and state it's unreliable. |
Signed-off-by: Eric Kerfoot <[email protected]>
|
/build |
wyli
left a comment
There was a problem hiding this comment.
Thanks, it looks good to me.
|
Hi @binliunls we merged this PR but any feedback you have on this would still be great. I had avoided doing profiling of CPU, GPU, or memory since there's the Pytorch profiler for that and TensorRT. |
Description
This adds a time profile class to measure how much recorded time various components of workflows consume. There is a tutorial notebook here that will be PR'd to the tutorials repo once this is accepted.
Status
Ready
Types of changes
./runtests.sh -f -u --net --coverage../runtests.sh --quick --unittests --disttests.make htmlcommand in thedocs/folder.