cd rewards_services/api_services/editreward_scorer_service
python -m venv .venv
source .venv/bin/activate
pip install --upgrade pip wheel setuptools   # `wheel` is required by flash-attn's source build
# Pick the PyTorch wheel matching your driver. PyPI's default now ships cu130, which
# needs driver >=580 — if `nvidia-smi` reports CUDA 12.x, use the cu128 index instead.
# See https://pytorch.org/get-started/locally/ for other CUDA targets.
pip install --index-url https://download.pytorch.org/whl/cu128 torch torchvision torchaudio
pip install -r requirements.txt
pip install flash-attn --no-build-isolation

git clone https://github.com/TIGER-AI-Lab/EditReward.git
huggingface-cli download TIGER-Lab/EditReward-MiMo-VL-7B-SFT-2508 \
  --local-dir EditReward/EditReward-MiMo-VL-7B-SFT-2508

export EDITREWARD_CUDA_DEVICES=0     # 1 worker is usually enough; bump to "0,1" + WORKERS=2 if reward is your bottleneck
export EDITREWARD_WORKERS=1
export EDITREWARD_PORT=18088
# export EDITREWARD_HOST=0.0.0.0   # only if the trainer runs on a different machine; defaults to 127.0.0.1
bash run.sh

export EDITREWARD_REPO_DIR=/path/to/EditReward
export EDITREWARD_CHECKPOINT_PATH=/path/to/EditReward-MiMo-VL-7B-SFT-2508

conda activate unirl
cd /path/to/PrompRL-opensource
python -m unimodel.qwenkontext.qwenkontext_inference
# produces outputs/pretrain/qwenkontext/ — use this as MODEL_NAME_OR_PATH below.

# Alternative: Qwen + FLUX.2-klein-base-4B
python -m unimodel.qwenklein.qwenklein_inference
# produces outputs/pretrain/qwenklein/ — point MODEL_NAME_OR_PATH at this and
# launch with scripts/train/edit_grpo_klein.sh instead of edit_grpo.sh.

export MODEL_NAME_OR_PATH=outputs/pretrain/qwenkontext   # from step 2
# Optional. Defaults to the PromptRL OmniEdit 50k parquet on Hugging Face.
export PROMPTS_FILE=https://huggingface.co/wangfuyun/PrompRL/resolve/main/data/omni_edit_train_50k.parquet
export EDITREWARD_URL=http://127.0.0.1:18088/
export CUDA_VISIBLE_DEVICES=1,2,3,4,5,6,7
export NPROC_PER_NODE=7
export NUM_SDE=2
export RUN_NAME=qwenkontext-editreward
export NUM_GENERATIONS=8          
export NUM_SKIP_REFINEMENT=2      # 6 refined + 2 original (partial-refinement preserved)
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
bash scripts/train/edit_grpo.sh

export MODEL_NAME_OR_PATH=outputs/pretrain/qwenklein   # from step 2
export RUN_NAME=qwenklein-editreward
export PROMPTRL_EDIT_NUM_INFERENCE_STEPS=6
export PROMPTRL_EDIT_GUIDANCE_SCALE=4
export PROMPTRL_DIFFUSION_LOSS_BATCH_SIZE=4
export PROMPTRL_EDIT_HEIGHT=512
export PROMPTRL_EDIT_WIDTH=512
export WANDB_PROJECT=PrompRL-klein 
export REPORT_TO=wandb
bash scripts/train/edit_grpo_klein.sh

export NUM_GENERATIONS=8
export NUM_SKIP_REFINEMENT=2
export NUM_INFERENCE_STEPS=8      # total FLUX denoising steps per edit (SDE prefix + ODE tail)
export NUM_SDE=2                  # first N steps run SDE w/ logprob; remaining steps run pure ODE
export SDE_NOISE_SCALE=0.8        # multiplier on per-step SDE noise std (FlowGRPO default 0.8)
export PER_DEVICE_TRAIN_BATCH_SIZE=1
export DIT_LEARNING_RATE=2e-7
export LLM_LEARNING_RATE=3e-7
export DIT_BETA1=0.9              # joint RL: try 0.0 or 0.5 to react faster to LLM prompt shift
export DIT_BETA2=0.999
export LLM_BETA1=0.9
export LLM_BETA2=0.999
export BETA=1e-2
export IMAGE_COLUMN=image
export PROMPT_COLUMN=prompt
export REPORT_TO=wandb

cd rewards_services/api_services/geneval_scorer_service
conda create -n geneval python=3.11 -y
conda activate geneval

# CUDA toolchain for mmcv/mmdet source builds
export CUDA_HOME=/usr/local/cuda
export PATH=$CUDA_HOME/bin:$PATH
export LD_LIBRARY_PATH=$CUDA_HOME/lib64:$LD_LIBRARY_PATH

# Mask2Former checkpoint used as the object detector
mkdir -p object_models
wget -O object_models/mask2former_swin-s-p4-w7-224_lsj_8x2_50e_coco.pth \
  https://download.openmmlab.com/mmdetection/v2.0/mask2former/mask2former_swin-s-p4-w7-224_lsj_8x2_50e_coco/mask2former_swin-s-p4-w7-224_lsj_8x2_50e_coco_20220504_001756-743b7d99.pth

# Pinned deps from the GenEval reference setup
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu121
pip install open-clip-torch==2.26.1 clip-benchmark einops lightning tomli platformdirs
pip install -U openmim
pip install "diffusers[torch]" transformers
pip install --upgrade setuptools==60.2.0

# mmcv / mmdetection from source (adjust -arch=sm_XX to your GPU)
git clone https://github.com/open-mmlab/mmcv.git
( cd mmcv && git checkout 1.x && MMCV_WITH_OPS=1 MMCV_CUDA_ARGS="-arch=sm_90" pip install -v -e . )
git clone https://github.com/open-mmlab/mmdetection.git
( cd mmdetection && git checkout 2.x && pip install -v -e . )

pip install gunicorn flask "numpy<2"

export GENEVAL_PORT=18085
# export GENEVAL_HOST=0.0.0.0   # only if the trainer runs on a different machine
bash run.sh

conda activate unirl
cd /path/to/PrompRL-opensource
python -m unimodel.qwenflux.qwenflux_inference   # outputs/pretrain/qwenflux/
# Or use the Klein checkpoint built earlier for the edit task; the path "qwenklein"
# is detected automatically.

export MODEL_NAME_OR_PATH=outputs/pretrain/qwenflux
export GENEVAL_URL=http://127.0.0.1:18085/
export RUN_NAME=qwenflux-geneval-prompt
bash scripts/train/geneval_grpo_flux.sh

export JOINT=true
export RUN_NAME=qwenflux-geneval-joint
bash scripts/train/geneval_grpo_flux.sh

export MODEL_NAME_OR_PATH=outputs/pretrain/qwenklein
export RUN_NAME=qwenklein-geneval-prompt   # or qwenklein-geneval-joint with JOINT=true
bash scripts/train/geneval_grpo_klein.sh

export NUM_GENERATIONS=8
export NUM_INFERENCE_STEPS=20      # total denoising steps
export NUM_SDE=10                  # flux.dev: first 10 with SDE; klein default: 8
export GUIDANCE_SCALE=3.5          # klein: 4.0
export GENEVAL_ONLY_STRICT=False   # set True for binary GenEval reward only
export LLM_LEARNING_RATE=1e-6      # prompt-only LR; joint mode also reads DIT_LEARNING_RATE
export DIT_LEARNING_RATE=2e-7