tiny-grpo

A minimal, hackable implementation of Group Relative Policy Optimization (GRPO).

Goal: Implementation of GRPO for training a local llama-3.2-3b model using RL. Focus on understanding GRPO algorithm. Run everything locally with a single RTX A6000 node and Axolotl Docker image on RunPod.

This project is inspired by and builds upon open-thought/tiny-grpo.

Updates

[2025-07-19] (in-progress) Supporting GPG: Group Policy Gradient. Some configurations can be found in verl/gpg. The main task here is to update reward modeling, grouping advantages, remove KL divergence.

• Following GPG/open-r1/src/open_r1/gpg_trainer.py, we implemented inverse_alpha, avoiding divide by zero when n_valid_samples = 0.

• (in-progress) Resampling based on ratio of valid samples.

[2025-07-16] Upgrading transformers 4.48.1 -> 4.53.2. Starting from transformers>=4.50, the library modularized model support (see huggingface/transformers (release v4.50.0)). Switched to using AutoModelForCausalLM.

• To load LLaMA models, you must explicitly install the llama extra. Perform sanity check with python -c "from transformers.models.llama import LlamaForCausalLM; print(LlamaForCausalLM)"

• See lamng3/tiny-grpo (issue #1) for more details.

[2025-07-15] Supporting DAPO, following this huggingface/trl#3130 (comment).

• Token-level Loss is already implemented as masked_mean.
• Clip-Higher is implemented following huggingface/trl#3118 (comment). DAPO recommends using clip_eps_low = 0.2 and clip_eps_high = 0.28.
• Dynamic Sampling (in progress) is skipped in huggingface/trl#3130 (comment) because of inefficiency. The configurations can be found in verl/dapo.
• Overlong Filtering is skipped (see the reason in huggingface/trl#3130 (comment) and verl/dapo).
• Soft Overlong Punishment is implemented following huggingface/trl#3130 (comment), with L_cache = 256 for L_max = 1024, inspired from verl/dapo, where L_cache is overlong_buffer.

[2025-07-12] Supporting Dr.GRPO, with modifications in calculating {masked_mean with constant generation max tokens (512 from oat/oat/args.py)} and {group_advantage without std bias}, following understand-r1-zero/train_zero_math.py.

Setup

1. Spin up a RunPod instance

choose 1 RTX A6000 ($0.49/hr) on Axolotl Docker image.

2. Create conda env

conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/main
conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/r
conda create --name grpo python=3.12 -y
source ~/.bashrc # or ~/.zshrc if you're using zsh
conda init
conda activate grpo

3. Install dependencies

cd tiny-grpo
pip install -r requirements.txt
pip install hf_transfer
pip install flash-attn --no-build-isolation

4. HuggingFace and WandB login

huggingface-cli login
wandb login

5. Play with the source in train.py

python train.py

6. Pushing code to GitHub

<!-- generate SSH key (if not yet) -->
ssh-keygen -t ed25519 -C "your_email@example.com"

<!-- add public key to github -->
cat ~/.ssh/id_ed25519.pub

<!-- change repo remote to SSH -->
git remote set-url origin git@github.com:<username>/<repo_name>.git

<!-- test connection -->
ssh -T git@github.com

7. Transfer file from local computer to RunPod instance

scp -i ~/.ssh/id_ed25519 -P <port> <local_file_path> root@<host_name>:<remote_destination_path>

Training Results

The run stopped at 9K steps due to insufficient storage memory. To prevent this, consider doubling the storage capacity or offloading checkpoints to temporary storage.

References

• GPG: A Simple and Strong Reinforcement Learning Baseline for Model Reasoning
• DAPO: An Open-Source LLM Reinforcement Learning System at Scale
• Understanding R1-Zero-Like Training: A Critical Perspective
• DeepSeek-R1 Tech Report
• DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models