support exlusion of params when using low bit optim

[nighting0le01]

This PR allows, exclusion and inclusion of params layerwise when using low bit optimizers. this will allow for improving stability by running certain layers with 32 bit adam. https://huggingface.co/docs/bitsandbytes/main/en/optimizers

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/1225

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[facebook-github-bot]

Hi @nighting0le01!

Thank you for your pull request and welcome to our community.

Action Required

In order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you.

Process

In order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it.

If you have received this in error or have any questions, please contact us at [email protected]. Thanks!

[facebook-github-bot]

Thank you for signing our Contributor License Agreement. We can now accept your code for this (and any) Meta Open Source project. Thanks!

[gau-nernst]

Thank you for the PR! Left some small comments.

[gau-nernst]

I think you can hash tensor directly (it will use object id internally). PyTorch optimizer already hashes tensors when it uses params as keys in self.state.

[nighting0le01]

@gau-nernst shall i also add configurable min_8bit_size like https://github.com/bitsandbytes-foundation/bitsandbytes/blob/9568735b21b9325e4789d6a5004517f2287f47c8/bitsandbytes/optim/optimizer.py#L603

over here :

ao/torchao/prototype/low_bit_optim/adam.py

Line 61 in 71a442a

if p.numel() >= 4096 and p.numel() % self.block_size == 0:

[gau-nernst]

@nighting0le01 Adding something like min_8bit_size should be good. Though personally I don't know if having it is useful in any ways (does anyone use it / does adjust it help with stability?) If you still want to add it, maybe we can call it more generic, like min_size_for_low_bit, since we also have 4-bit and FP8.

Do you mind rebase/merge from main and make sure the tests pass?

[nighting0le01]

hi @gau-nernst !

1. yes i have rebased and confirmed all test cases pass.
2. min_size_for_low_bit. why i propose this is to allow running gradient exploding or unstable layers in 32bit precision. similar motivation to https://huggingface.co/docs/bitsandbytes/main/en/optimizers#optimize-unstable-parameters
3. i can push it in another PR if you suggest

[gau-nernst]

1. There are conflicts in your branch, hence I can't run the CI. Do you mind double-check? (from Github UI it shows test_low_bit_optim.py and adam.py have conflicts) Seem like you rebase from an outdated main? The diff for this PR looks kinda strange (there are changes in unwanted places)
2. From what I understand min_size_for_low_bit (or the original min_8bit_size) is to skip small params that don't contribute much memory savings if we use low-bit optim state for them (e.g. biases, norm params). How would it improve exploding gradients or instability? Usually instability appears in embedding layer or LM head I think (correct me if I'm wrong), which are large params but receive (somewhat) sparse gradients. In other words, how does increasing (or decreasing) the threshold help to improve stability? With this PR, the users can already select which specific params they want to keep optim state in original precision.