Sync-R1 Core Training Logic

This repository snapshot keeps only the current paper-aligned GRPO training path for Sync-R1.

Included Code

The GitHub-facing version intentionally keeps a single public entrypoint and its direct dependencies:

• train_grpo_paper.py: training launcher
• grpo_paper.py: paper-closer GRPO implementation
• ref_model.py: reference-side rollout helper
• pdata.py, utils.py, clip_eval.py, glm_api.py: runtime helpers
• models/, training/, llava/: model and training modules
• configs/: config files
• requirements.txt: minimal dependency list

Older draft variants are intentionally excluded to avoid ambiguity for users.

Paths To Fill In

The config files use placeholder paths:

• path/to/show-o-512x512
• path/to/show-o
• path/to/magvitv2
• path/to/phi-1_5
• path/to/checkpoints

Update them in:

• configs/showo_demo_512x512.yaml
• configs/showo_demo.yaml

The training command also expects:

• --data_root path/to/unictokens_data
• --pre_trained_ckpt_name path/to/second_stage_checkpoint_dir

External Assets Not Included

This trimmed repo does not include:

• training data
• pretrained checkpoints
• CLIP weights
• facenet weights
• generated images and logs

You need to provide them locally before training.

Launch

Run from Sync-R1/. The script initializes torch.distributed, so torchrun is the recommended launcher.

3 GPUs

torchrun --nproc_per_node=3 train_grpo_paper.py \
  --num_gpus 3 \
  --config_file configs/showo_demo_512x512.yaml \
  --data_root path/to/unictokens_data \
  --pre_trained_ckpt_name path/to/second_stage_checkpoint_dir \
  --concept adrien_brody \
  --save_dir ./tmp_result_accelerate/ \
  --epoch_to_load 15 \
  --batch_num 10 \
  --batch_size 1 \
  --num_gen 9 \
  --llm glm \
  --accelerate True \
  --semantic True

1 GPU

torchrun --nproc_per_node=1 train_grpo_paper.py \
  --num_gpus 1 \
  --config_file configs/showo_demo_512x512.yaml \
  --data_root path/to/unictokens_data \
  --pre_trained_ckpt_name path/to/second_stage_checkpoint_dir \
  --concept adrien_brody \
  --save_dir ./tmp_result_accelerate/ \
  --epoch_to_load 15 \
  --batch_num 10 \
  --batch_size 1 \
  --num_gen 3 \
  --llm glm \
  --accelerate True \
  --semantic True

Runtime Notes

• --num_gen is the rollout group size for a single prompt.
• The current training loop assumes batch_size=1 and multiple rollouts per prompt.
• --num_gpus should match --nproc_per_node.

Optional Environment Variables

If you use the LLM-based scoring paths, configure credentials through environment variables instead of hardcoding them:

• ZAI_API_KEY
• VERTEXAI_PROJECT
• VERTEXAI_LOCATION
• GOOGLE_APPLICATION_CREDENTIALS

Scope

This trimmed repo is meant for:

• reading the current GRPO training logic
• reproducing the paper-closer implementation
• auditing or modifying the Sync-R1 training path

It is not a plug-and-play full training package until you attach the required local datasets, checkpoints, and external model assets.