Skip to content

[Feat] support action forward dynamics finetuning - #283

Merged
rickzw merged 3 commits into
NVIDIA:mainfrom
WeiChengTseng:cosmos3-action-fd
Jul 31, 2026
Merged

[Feat] support action forward dynamics finetuning#283
rickzw merged 3 commits into
NVIDIA:mainfrom
WeiChengTseng:cosmos3-action-fd

Conversation

@WeiChengTseng

Copy link
Copy Markdown
Contributor

Summary

  • Add a launch script for Cosmos3-DROID forward-dynamics post-training.
  • Add structured TOML configuration with bf16, distributed sharding, activation checkpointing, and a 20K-step schedule.

Testing

  • Not run (configuration and launcher changes only).

@WeiChengTseng

Copy link
Copy Markdown
Contributor Author

Update Summary

Addressed the review comments in the latest commit:

  • Fixed cookbook-relative paths and anchored them to the launcher location, so the script works regardless of the caller's working directory.
  • Removed the dependency on the missing _sft_launcher_common.sh; the launcher is now self-contained with input validation, logging, overrides, and single-/multi-node torchrun setup.
  • Added action_fd_droid_posttrain.md with prerequisites, dataset layout, setup instructions, config validation, training commands, and output details.
  • Added framework-root resolution and model-config validation so framework-relative assets resolve correctly.

Comment thread cookbooks/cosmos3/generator/action/finetune/action_fd_droid_posttrain.md Outdated
@rickzw
rickzw merged commit 404b9bf into NVIDIA:main Jul 31, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants