support profiling and hooks from pl17 from pl 1.6

[paantya]

🚀 Feature

support profiling and hooks from pl17 in pl 1.6, in particular hooks/steps for creating a batch (fetch_batch ) and transferring it to the device - I want to be able to get time for these steps when profiling.

Motivation

due to the fact that in PL 1.7 there is no support for torch 1.8.2, I want PL 1.6 to have useful functions that are in PL 1.7

Submission

Get the opportunity to profile and measure the time at the stage of receiving the batch and transferring the batch to the device.

Alternatives

for good, I want it to be without alternatives, and it was simply implemented as it is in 1.7, tk. we support 1.7 but there are conflicts with the torch version -- we want to keep the lightning and torch versions fully compatible.

Additional context

a question about profiling where the missing steps in 1.6 are indicated
https://pytorch-lightning.slack.com/archives/CRBLFHY79/p1659645752234179

support request sticking out 1.8.2 in pl 1.7
#14086

---

Rays of love for lightning ⚡

cc @Borda @carmocca @kaushikb11 @ninginthecloud @rohitgr7 @nbcsm @guotuofeng

[paantya]

while there is a suspended state with the version, I will request what we need

[awaelchli]

Hello!

I don't exactly understand what this means:

support profiling and hooks from pl17 in pl 1.6,

But here is a PR that will add profiling for the batch transfer hooks.

To be clear, there won't be any 1.6.x releases anymore. Any new additions or fixes will land in 1.8dev or 1.7.x depending on whether it is a feature or a fix.

[paantya]

@awaelchli looks great!
will he add to PL 1.6?

FYI: this is not about the new, but about what is already in 1.7

It was previously discussed that if something is needed critical, then it can be added in 1.6: #14086 (comment)

I asked if there's an updated LTS schedule for PyTorch in their developer Slack and got no answer. I would respect our support window. If there's enterprise interest in an LTS PL version, then they should contact us to discuss it. A solution could be to include critical bug fixes into the 1.6.x branch, but not support PyTorch 1.8 endlessly across versions.

[carmocca]

Hi, with something critical I meant serious bugs that block a large percentage of users or security patches. We would not backport features into previous releases. Note that this was a suggested potential solution, but there's been no decision on whether we would do this.

Does your company require using PyTorch 1.8? Is there a specific reason why you cannot upgrade?

[rohitgr7]

If I am not wrong! PL==1.7 might not break if you install PT==1.8 again manually