# DSGRPO

This repository contains the DSGRPO training setup built on top of `maxrl`/`verl`.
The git repository intentionally does not track model weights, parquet datasets,
logs, or checkpoints. Prepare those files under `assets/` on each machine.

## Repository Layout

```text
DSGRPO/
  maxrl/                         # training code
  train_qwen2.5_1.5b_math.sh      # 1.5B DSGRPO Slurm job
  train_qwen2.5_7b_math.sh        # 7B DSGRPO Slurm job
  launch.sh                       # batch launch helper
  slurm_reference.sh              # local Slurm reference
  assets/                         # local-only models and parquet data, not in git
```

The training scripts infer `REPO_ROOT` from their own location, so the clone path
does not need to be `/home/guanninz/DSGRPO`.

## 1. Clone

```bash
git clone git@github.com:guanning03/DSGRPO.git
cd DSGRPO
```

## 2. Create the Python Environment

The Slurm scripts activate a conda environment named `more`. Either create that
environment or edit the scripts to use a different env name.

```bash
conda create -n more python=3.10 -y
conda activate more

pip install --upgrade pip setuptools wheel
pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu124

pip install ninja packaging psutil
MAX_JOBS=4 pip install flash-attn --no-build-isolation

pip install vllm==0.8.4 wandb math-verify huggingface_hub hf_transfer
pip install -r maxrl/requirements_sglang.txt
pip install -e maxrl
```

If `flash-attn` fails to build, run the install on a GPU node with the target CUDA
toolkit available, lower `MAX_JOBS`, or use a matching prebuilt wheel.

Optional sanity check:

```bash
python - <<'PY'
import torch, ray, vllm, verl
print("torch", torch.__version__)
print("cuda", torch.version.cuda)
print("ray", ray.__version__)
print("vllm", vllm.__version__)
print("verl import ok")
PY
```

## 3. Prepare Assets

Required for the included jobs:

```text
assets/Qwen2.5-Math-1.5B/config.json
assets/Qwen2.5-Math-1.5B/model.safetensors
assets/Qwen2.5-Math-7B/config.json
assets/Qwen2.5-Math-7B/model-00001-of-00004.safetensors
assets/dapo17k_math/train.parquet
```

Recommended validation files:

```text
assets/math500/test.parquet
assets/aime24/test.parquet
assets/aime25/test.parquet
assets/olympiadbench/test.parquet
assets/amc23/test.parquet
```

The training scripts will use every validation file that exists and will skip
missing ones. At least one validation parquet file must exist.

Fastest setup if a working machine already has the assets:

```bash
mkdir -p assets
rsync -av --info=progress2 OLD_SERVER:/home/guanninz/DSGRPO/assets/ ./assets/
```

Independent model download:

```bash
export HF_HUB_ENABLE_HF_TRANSFER=1
huggingface-cli login

huggingface-cli download Qwen/Qwen2.5-Math-1.5B \
  --local-dir assets/Qwen2.5-Math-1.5B \
  --local-dir-use-symlinks False

huggingface-cli download Qwen/Qwen2.5-Math-7B \
  --local-dir assets/Qwen2.5-Math-7B \
  --local-dir-use-symlinks False
```

Independent dataset build:

```bash
PYTHONPATH=$PWD/maxrl python maxrl/examples/data_preprocess/dapo.py \
  --local_dir assets/dapo17k_math

PYTHONPATH=$PWD/maxrl python maxrl/examples/maxrl_data_preprocess/math_500.py \
  --local_dir assets/math500

PYTHONPATH=$PWD/maxrl python maxrl/examples/maxrl_data_preprocess/aime25.py \
  --local_dir assets/aime25

PYTHONPATH=$PWD/maxrl python maxrl/examples/maxrl_data_preprocess/olympiadbench_amc23.py \
  --local_dir assets
```

For exact reproduction of the current local assets, prefer `rsync` because it
also copies `aime24`, `gsm8k`, `polaris`, `openr1`, and other optional parquets.

Asset check:

```bash
test -f assets/Qwen2.5-Math-1.5B/config.json
test -f assets/Qwen2.5-Math-7B/config.json
test -f assets/dapo17k_math/train.parquet
find assets -maxdepth 2 -name 'test.parquet' -print
```

## 4. Configure Slurm

The two training scripts currently request:

```text
partition: defq
account:   ece-research
nodes:     1
gpus:      4
cpus:      64
mem:       500G
time:      48:00:00
```

On another cluster, edit the `#SBATCH` header in:

```text
train_qwen2.5_1.5b_math.sh
train_qwen2.5_7b_math.sh
```

Adjust `--partition`, `--account`, `--gres`, `--cpus-per-task`, and `--mem` to
match that cluster. The Python arguments below the header are path-independent.

Use Slurm's dry run before submitting:

```bash
sbatch --test-only train_qwen2.5_1.5b_math.sh 101 0.05 0.05
sbatch --test-only train_qwen2.5_7b_math.sh 101 0.05 0.05
```

## 5. Submit Training

Arguments are:

```text
sbatch SCRIPT DATA_SEED CORRECT_SAMPLE_LOG_PROB_COEF INCORRECT_SAMPLE_LOG_PROB_COEF
```

Run DSGRPO smoothing with both coefficients set to `0.05`:

```bash
sbatch train_qwen2.5_1.5b_math.sh 101 0.05 0.05
sbatch train_qwen2.5_7b_math.sh 101 0.05 0.05
```

Run the full local sweep in `launch.sh`:

```bash
bash launch.sh
```

Monitor:

```bash
squeue -u $USER
tail -f logs/<jobid>_<jobname>.log
```

Outputs:

```text
logs/
checkpoints/qwen2.5-1.5b-math-dsgrpo-dapo17k/
checkpoints/qwen2.5-7b-math-dsgrpo-dapo17k/
```

## Codex Handoff Prompt

After cloning this repository on a new server, a concise prompt for another Codex:

```text
Please follow README.md to set up this DSGRPO repo on this server. Create/verify
the conda env named "more", install dependencies, prepare assets under ./assets,
run the asset checks, run sbatch --test-only on both training scripts, and tell
me the exact sbatch commands to launch 1.5B and 7B DSGRPO with coefficients
0.05 and 0.05.
```