For #125323
* Fixes typing for python < 3.10
* Fixes#129390
For #124688
* Improved attribution by registering `register_hook` and `post_accumulate_grad_hook` on params.
* Fixed pre-mature per module bw peak state initialization for AC.
* This improves per-module stats, global `peak_mem` was already accurate and remains unaffected.
For #128508
* When AC is applied to a `mod (nn.Module)` the backward order of execution is `pre-bw -> pre-fw -> post-fw -> post-bw`. Since the `ModTracker` maintains the `parents` attribute as set, the `post-fw` during backward was prematurely removing it from parents.
* With the fix we now maintain a per-module counter and only remove a module from `parents` when its counter goes to 0.
* Added tests to ensure this.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/129400
Approved by: https://github.com/awgu, https://github.com/huydhn
This is an extension of [ModuleTracker](https://github.com/pytorch/pytorch/blob/main/torch/utils/module_tracker.py) with added features and bug fixes.
1. Allows installing user-defined hooks to be called in pre-fw, post-fw, pre-bw and post-bw hooks of the ``ModTracker``.
2. Adds a function ``get_known_fqn`` that retrieves the fqn of the module as tracked by the ``ModTracker``.
3. Only registers the multi-grad hooks if we are in the forward pass. This is important because, a module's pre-fw and post-fw hooks get called in the backward during AC and we do not want to register multi-grad hooks in this case.
4. Sets the kwarg ``always_call=True`` for post-fw hooks, so that they are called post AC.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/128508
Approved by: https://github.com/wanchaol
This is an extension of [ModuleTracker](https://github.com/pytorch/pytorch/blob/main/torch/utils/module_tracker.py) with added features and bug fixes.
1. Allows installing user-defined hooks to be called in pre-fw, post-fw, pre-bw and post-bw hooks of the ``ModTracker``.
2. Adds a function ``get_known_fqn`` that retrieves the fqn of the module as tracked by the ``ModTracker``.
3. Only registers the multi-grad hooks if we are in the forward pass. This is important because, a module's pre-fw and post-fw hooks get called in the backward during AC and we do not want to register multi-grad hooks in this case.
4. Sets the kwarg ``always_call=True`` for post-fw hooks, so that they are called post AC.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/128508
Approved by: https://github.com/wanchaol
