Refactor Quantizer for reusing in QAT

[kevalmorabia97]

What does this PR do ?

This PR refactors the Quantizer class used for PTQ so it can be reused for QAT as well (follow-up PR)

Note that functionally wise it is same as before

Collection: N/A

Changelog

• N/A

Usage

• N/A

GitHub Actions CI

The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.

The GitHub Actions CI will run automatically when the "Run CICD" label is added to the PR.
To re-run CI remove and add the label again.
To run CI on an untrusted fork, a NeMo user with write access must first click "Approve and run".

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

• New Feature
• Bugfix
• Documentation
• Refactoring existing feature code

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

• Related to # (issue)

[kevalmorabia97]

Can we remove this disable of activation checkpointing as well? For the SFT model there is no accuracy regression even if we enable activation checkpointing plus that reduces GPU memory requirement

[janekl]

I'm not sure regarding this, see also related examples in #9276 (comment). Maybe you could ask internally?

[janekl]

So in the end it turns out that it's OK to remove setting these to None?:

model_cfg.activations_checkpoint_method = None
model_cfg.activations_checkpoint_granularity = None

I hope it will not blow up in some large tests

[kevalmorabia97]

Users always have the option to disable it in the yaml if they want to, but we should not restrict them to not be able to use it as it allows reducing GPU memory requirement

[kevalmorabia97]

I will set them as None in the yaml file so there is no difference than previous version

[janekl]

Could you demostrate/link how do you want to use that refactored Quantizer?

[janekl]

Note that changes will need to be unstreamed to another project as well here https://github.com/NVIDIA/NeMo-Framework-Launcher/blob/main/launcher_scripts/conf/ptq/model/quantization.yaml. I think we'll need to chop this slightly

[kevalmorabia97]

Could you demostrate/link how do you want to use that refactored Quantizer?

Its used in #9276

[janekl]

We'll need to do follow up MRs to:

• https://github.com/NVIDIA/NeMo-Framework-Launcher/blob/main/launcher_scripts/conf/ptq/model/quantization.yaml
• https://github.com/NVIDIA/TensorRT-Model-Optimizer/blob/main/llm_ptq/nemo_ptq.py
  (I know it's duplication but it is what it is)

[kevalmorabia97]

@janekl should I rename alpha to sq_alpha so it is clear alpha is used only for SQ?

[janekl]

@janekl should I rename alpha to sq_alpha so it is clear alpha is used only for SQ?

OK, good idea (same applies to Launcher ofc.)