feat: support convert_lora_to_hf without merge

[RayenTian]

Summary

• New --adapter-only flag: exports only the LoRA adapter weights in HuggingFace PEFT format (no base model required at export time), enabling workflows like vLLM's split base+adapter serving
• Bug fix in merge_lora_to_hf: adapter weights are now loaded via dist_checkpointing.load with a filtered sharded state dict (via apply_peft_adapter_filter_to_state_dict), fixing a KeyError that occurred when using the previous _load_model_weights_from_checkpoint path which silently drops adapter
  parameters
• Functional test: test_converter_roundtrip.py now covers the adapter-only export path and verifies that re-merging the exported PEFT adapter via merge_lora_to_hf produces weights identical to the direct merge path
• Docs: updated checkpointing.md and sft.md with both export modes (Option A: merged, Option B: adapter-only)

What does this PR do ?

Add a one line overview of what this PR aims to accomplish.

Issues

closes #2190

Usage

• You can potentially add a usage example below

# Add a code snippet demonstrating how to use this

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 run the unit tests and functional tests locally? Visit our Testing Guide for how to run tests
• Did you add or update any necessary documentation? Visit our Document Development Guide for how to write, build and test the docs.

Additional Information

• ...

[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.

[RayenTian]

/ok to test eeef66c

[RayenTian]

/ok to test 9de1880

[RayenTian]

/ok to test 64a8679

[RayenTian]

/ok to test ed1bbdd

[RayenTian]

/ok to test f24be7a

[yuki-97]

thanks for fixing and adding the support! lgtm

[yuki-97]

@terrykong could you take a review as well?

[RayenTian]

/ok to test 5396d85

[terrykong]

Nice work on this PR! The refactoring to extract _build_megatron_model_with_lora as a shared context manager is clean, the bug fix for adapter loading (using dist_checkpointing.load with a filtered state dict via apply_peft_adapter_filter_to_state_dict) correctly addresses the KeyError from the old _load_model_weights_from_checkpoint path, and the docs clearly present both export modes with examples. Nice work addressing the community request (#2190).

Two minor suggestions below — neither is a blocker.

Generated by Claude Code

[terrykong]

lgtm docstring was just a nit, btw looks like sphinx build failure

[RayenTian]

/ok to test 2db3e7d