Load model in the target export precision by default in PTQ

[janekl]

What does this PR do ?

Configuring PTQ to use the same precision to load a model as export.dtype by default. This is to avoid possible OOM errors when casting weights on the model export step.

This is for use cases with less device memory available, for example, PTQ for Llama3-8b on L4 24GB GPU.

Collection: NLP

Changelog

• Add specific line by line info of high level changes in this PR.

Usage

• In order to use FP16 / BF16 weights precision enabling model.megatron_amp_O2 is necessary
• Disabling model.dist_ckpt_load_on_device avoid GPU memory spikes on model loading
• Using a lower inference.batch_size (compared to default 64) might also be needed.

Example PTQ command is:

python megatron_gpt_ptq.py \
    model.restore_from_path=Llama3-8b \
    +model.dist_ckpt_load_on_device=false \
    export.dtype=bf16 \
    inference.batch_size=16

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The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.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

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 Llama3-8b FP8 PTQ OOM #9981