Some way for tlparse to automatically upload into gist for easy sharing

[zou3519]

We ask open source people for tlparses. They are commonly confused about what artifact to actually send us (the zip file, but I receive inductor output_code quite often). We should provide some way for the user to upload the tlparse artifacts to e.g. gist and so a pytorch developer can download things from there. (suggested by @oulgen)

[ezyang]

Why do you not just ask them for the raw logs? We can always run tlparse locally. The raw log is preferred. Do you want them to host a small github pages that has the rendered website on it?

[oulgen]

Why do you not just ask them for the raw logs? We can always run tlparse locally. The raw log is preferred. Do you want them to host a small github pages that has the rendered website on it?

The problem we want to solve is when a user wants to share the logs with us, they are not always able to identify exactly what to share, and also it requires us to educate each user. This has caused quite a few back and forths in the past. Similar to internal XYZ rage tools, we were thinking we can automate this process by a single command rather than asking every user to understand what to share.

[ezyang]

But there's only one log, the one that PyTorch generated into the TORCH_TRACE directory?

If you want a rage experience, that's a bit different, and we would need to be proactively tracking previous users sessions so we can ask them which one to send to us. But tlparse data is generally a bit sensitive and so it's better to force the user to think about what exactly it is they are sharing with us.

[zou3519]

@ezyang I didn't know that TORCH_TRACE just generated a single log file that can then be passed back to us. My original point was that tlparse generated a directory with multiple files and users sometimes just copy-paste the inductor output code (instead of sending over a zip of the directory).

I think the TORCH_TRACE is what I was looking for then. Would it be fair to say we should ask users to give us the TORCH_TRACE output (if they're OK with it) and then we can directly run tlparse on it to inspect it?

Also there is zero marketing about TORCH_TRACE

[ezyang]

Yes, for sharing I would recommend having them just send us the raw TORCH_TRACE.

Re marketing, I guess we can spray it everywhere in the docs lol