run_franka_v4_finetune_qwenoft.sh

#!/bin/bash
set -eo pipefail

# Activate the starVLA conda environment.
CONDA_ENV=${CONDA_ENV:-/gpfs/wangzixuan/conda_envs/starVLA}
# shellcheck disable=SC1091
source "$(conda info --base)/etc/profile.d/conda.sh"
conda activate "${CONDA_ENV}"

export NCCL_SOCKET_IFNAME=${NCCL_SOCKET_IFNAME:-bond0}
export NCCL_IB_HCA=${NCCL_IB_HCA:-mlx5_2,mlx5_3}
export NCCL_BLOCKING_WAIT=${NCCL_BLOCKING_WAIT:-1}
export NCCL_ASYNC_ERROR_HANDLING=${NCCL_ASYNC_ERROR_HANDLING:-1}
export NCCL_TIMEOUT=${NCCL_TIMEOUT:-10000}

SCRIPT_DIR=$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)
REPO_ROOT=$(cd -- "${SCRIPT_DIR}/../../.." && pwd)
cd "${REPO_ROOT}"
export PYTHONPATH="${REPO_ROOT}:${PYTHONPATH:-}"

# v4 single-arm Franka datasets (make_sandwich + move_place_gaizi, co-trained).
config_yaml=${config_yaml:-"${REPO_ROOT}/examples/Franka/train_files/starvla_cotrain_franka_v4.yaml"}
franka_data_root=${franka_data_root:-/gpfs/wangzixuan/Octopus/workspace/data/lerobot/v4}
data_mix=${data_mix:-real_franka_v4}
# Leave empty to train from Qwen3-VL base only (action head from scratch).
pretrained_checkpoint=${pretrained_checkpoint:-}
base_vlm=${base_vlm:-/gpfs/wangzixuan/Octopus/workspace/MODEL/Qwen3-VL-4B-Instruct}
# reload_modules=${reload_modules:-'qwen_vl_interface'}

run_root_dir=${run_root_dir:-"${REPO_ROOT}/results/Checkpoints"}
run_id=${run_id:-0622_franka_v4_qwen3OFT_sandwich}
num_processes=${num_processes:-8}
per_device_batch_size=${per_device_batch_size:-16}
max_train_steps=${max_train_steps:-100000}
save_interval=${save_interval:-10000}
logging_frequency=${logging_frequency:-100}
eval_interval=${eval_interval:-1000}

if [[ ! -f "${config_yaml}" ]]; then
  echo "Config yaml not found: ${config_yaml}" >&2
  exit 1
fi

if [[ -n "${pretrained_checkpoint}" && ! -f "${pretrained_checkpoint}" ]]; then
  echo "Pretrained checkpoint not found: ${pretrained_checkpoint}" >&2
  exit 1
fi

output_dir="${run_root_dir}/${run_id}"
mkdir -p "${output_dir}"
cp "$0" "${output_dir}/"

echo "============================================"
echo "Conda env             : ${CONDA_ENV}"
echo "Config yaml           : ${config_yaml}"
echo "Base VLM              : ${base_vlm}"
echo "Pretrained checkpoint : ${pretrained_checkpoint:-<none> (Qwen3-VL base only)}"
echo "Franka data root      : ${franka_data_root}"
echo "Data mix              : ${data_mix}"
echo "Batch size / process  : ${per_device_batch_size}"
echo "Num processes         : ${num_processes}"
echo "Max train steps       : ${max_train_steps}"
echo "Run ID                : ${run_id}"
echo "============================================"

launch_args=(
  --config_file "${REPO_ROOT}/starVLA/config/deepseeds/deepspeed_zero2.yaml"
  --num_processes "${num_processes}"
  "${REPO_ROOT}/starVLA/training/train_starvla.py"
  --config_yaml "${config_yaml}"
  --framework.name QwenOFT
  --framework.qwenvl.base_vlm "${base_vlm}"
  --datasets.vla_data.data_root_dir "${franka_data_root}"
  --datasets.vla_data.data_mix "${data_mix}"
  --datasets.vla_data.per_device_batch_size "${per_device_batch_size}"
  --datasets.vla_data.sequential_step_sampling false
  --trainer.max_train_steps "${max_train_steps}"
  --trainer.save_interval "${save_interval}"
  --trainer.logging_frequency "${logging_frequency}"
  --trainer.eval_interval "${eval_interval}"
  # --trainer.reload_modules "${reload_modules}"
  --run_root_dir "${run_root_dir}"
  --run_id "${run_id}"
  --wandb_project roboclaw
  --wandb_entity zwanggk
)

if [[ -n "${pretrained_checkpoint}" ]]; then
  launch_args+=(--trainer.pretrained_checkpoint "${pretrained_checkpoint}")
fi

accelerate launch "${launch_args[@]}"