Add option for LoRA with Transformer Engine op fuser

[timmoon10]

What does this PR do ?

This PR adds a variant of the LoRALinear module using Transformer Engine's operation fuser. It also adds logic in the existing LoRALinear module so that it can call the fused variant. This is an experimental feature.

Collection: NLP

Changelog

• Add TEFusedLoRALinear module that implements LoRA with TE op fuser.
• Add logic in LoRALinear module to internally call TEFusedLoRALinear.

Usage

Set use_transformer_engine_op_fuser=True in the LoRA model config (see #13776).

from nemo.collections.llm import Llama2Config70B
from nemo.collections.llm.api import _setup
from nemo.collections.llm.gpt.model import LlamaModel
from nemo.collections.llm.peft.lora import LoRA

config = llm.Llama2Config70B(..., use_transformer_engine_op_fuser=True)
model = LlamaModel(config, ...)
peft = LoRA(...)
_setup(model=model, model_transform=peft, ...)

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The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.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

• Depends on Transformer Engine bugfixes: [PyTorch] Register weight and bias params in linear op NVIDIA/TransformerEngine#2027, [PyTorch] Avoid registering FP8 scale update in ops without backward pass NVIDIA/TransformerEngine#2063
• Benefits from Transformer Engine optimizations: Fuse linear+scale+add NVIDIA/TransformerEngine#2042, Dropout with 8-bit RNG NVIDIA/TransformerEngine#2014

[cuichenx]

Thanks for the PR! Do you have any data on the performance benefits of fused lora?

[timmoon10]

Currently I don't see a significant difference between the fused and unfused impl (174 ms/step with a small Llama model on B200). We skip one add (fused into dgrad GEMM), but the dropout implementation is quite unoptimized. We should see speedup as more fusions are supported in TE.

[gautham-kollu]

Need to revert the manifest.json before merging