feat: Add safe_serialization option to the dcp checkpoint conversion script

[giordanobsf]

What does this PR do ?

Currently, the script converts dcp checkpoints to .bin for the HF checkpoint. This PR allows users to also convert checkpoints to .safetensors.

Issues

Closes #491

Usage

• You can potentially add a usage example below

uv run python examples/converters/convert_dcp_to_hf.py \
    --config results/grpo/step_170/config.yaml \
    --dcp-ckpt-path results/grpo/step_170/policy/weights/ \
    --hf-ckpt-path results/grpo/hf \
    --safe-serialization true

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.

[terrykong]

thanks for the change. could you add a unit test for this?

as a heads up for where we are heading, we will eventually integrate the native pytorch DCP checkpointing of HF so that this script wont' be needed in the future. cc @ffrujeri

It's a little longer horizon, so I'm okay with this change as long as it's tested

[euronymous-aithal]

@ffrujeri @joyang-nv is this still relevant with Automodel ?

[ffrujeri]

@ffrujeri @joyang-nv is this still relevant with Automodel ?

With the Automodel integrations, the users will be able to save directly a checkpoint in .safetensors format. But if a previous checkpoint was created on the .dcp format and the user wants to convert to safetensors and then start training from there, I think this conversion script could be useful? Did I understand this PR correctly?

[euronymous-aithal]

@ffrujeri i think so. @joyang-nv what is your feedback. how do we plan to transition folks from earlier DTensor path to Automodel path ?

[ashors1]

gentle bump here @joyang-nv

[joyang-nv]

Deeply sorry for so late reply.

[joyang-nv]

Here is my feedbacks.
@giordanobsf , deeply sorry from me on so late reply on this PR.
Can you rebase your PR and could help to test:

1. Convert checkpoint from v1 policy worker?
2. Convert checkpoint from v2 policy worker?
   I knew with v2 HF converter might have issue. It will be great if your PR could work in both paths.

[dhineshkumar-r]

Reviving this thread.
@ffrujeri / @terrykong - What's the recommendation in March 2026?
I need this capability as well.