[PEFT] feat: Add support for temporary disable adapter

[HollowMan6]

What does this PR do ?

Similar to:
https://github.com/huggingface/peft/blob/261366de2e40cde64b702d6b9c527081ad850549/src/peft/mixed_model.py#L192-L201

enable_adapter_layers and disable_adapter_layers are alternatives if users want to control manually.

Can be used for volcengine/verl#4673

Changelog

• Add APIs to disable/re-enable adapters temporarily.

GitHub Actions CI

See the CI sectionin the Contributing doc for how to trigger the CI. A Nvidia developer will need to approve and trigger the CI for external contributors.

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?

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

Additional Information

• Related to # (issue)

_{✨ Presented to you with Mind Lab - A Lab for Experiential Intelligence.}

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[yaoyu-33]

@HollowMan6 lgtm, need unit tests though.

[Copilot]

Pull request overview

This PR adds support for temporarily disabling and re-enabling adapter layers in PEFT (Parameter-Efficient Fine-Tuning) implementations, providing functionality similar to HuggingFace's PEFT library. This allows users to temporarily bypass adapter computations and fall back to the base model behavior without removing or unloading the adapters.

Key changes:

• Added _adapter_enabled flag to AdapterWrapper base class with enable_adapter_layers() and disable_adapter_layers() methods
• Implemented model-level enable_adapter_layers(), disable_adapter_layers(), and disable_adapter() context manager in the PEFT base class
• Updated all adapter forward methods to check the _adapter_enabled flag and return only base output when disabled
• Refactored model walking logic into reusable _walk_model() helper method

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
src/megatron/bridge/peft/adapter_wrapper.py	Added _adapter_enabled flag and enable/disable methods to AdapterWrapper base class
src/megatron/bridge/peft/base.py	Added model-level enable/disable methods, disable_adapter context manager, and refactored model walking logic into _walk_model() helper
src/megatron/bridge/peft/lora_layers.py	Updated LoRALinear and TEFusedLoRALinear forward methods to check adapter enabled flag and updated docstring
src/megatron/bridge/peft/dora_layers.py	Updated DoRALinear forward method to check adapter enabled flag
src/megatron/bridge/peft/canonical_lora.py	Updated LoRALinearSplitQKV and LoRALinearSplitFC1UpGate forward methods to check adapter enabled flag

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 6 out of 6 changed files in this pull request and generated 2 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.