更新时间:2026-11-06 GMT+08:00
分享

starVLA模型训推命令参考

本节汇总了starVLA模型的4个任务的相关训练(全参数微调)和推理部署的启动命令。

  • conda环境:starvla
  • 训练启动命令中可自主配置的参数:
    • num_processes:卡数
    • framework.qwenvl.base_vlm:基模型路径
    • datasets.vla_data.data_root_dir:数据根目录
    • datasets.vla_data.lerobot_version:默认v2.1格式,可取值v3.0
    • datasets.vla_data.data_mix:注册的nakename
    • datasets.vla_data.per_device_batch_size:批次大小
    • trainer.max_train_steps:训练的最大步数
    • trainer.save_interval:checkpoint保存频率
    • run_root_dir:公共根目录
    • run_id:单次执行子目录

训练启动命令

  • QwenOFT任务
    cd /opt/cloud/starvla_ascend_notebook/starVLA && conda activate starvla && export PYTHONPATH="$(pwd):$(pwd)/migrate_to_cloudrobo:${PYTHONPATH:-}" && export WANDB_MODE=disabled && accelerate launch --config_file ./starVLA/config/deepseeds/deepspeed_zero2.yaml --num_processes 2 starVLA/training/train_starvla.py --config_yaml ./migrate_to_cloudrobo/train_libero_cloudrobo.yaml --framework.name QwenOFT --framework.qwenvl.base_vlm ./playground/Pretrained_models/Qwen3-VL-4B-Instruct --datasets.vla_data.data_root_dir ./playground/Dataset/ --datasets.vla_data.lerobot_version v3.0  --datasets.vla_data.data_mix libero --datasets.vla_data.per_device_batch_size 2 --trainer.freeze_modules '' --datasets.vla_data.video_backend torchvision_av --trainer.max_train_steps 1000 --trainer.save_interval 500 --trainer.logging_frequency 100 --trainer.eval_interval 100 --run_root_dir ./playground/Checkpoints --run_id test --wandb_project starVLA_Libero --wandb_entity npu_adapt
  • QwenGR00T任务
    cd /opt/cloud/starvla_ascend_notebook/starVLA && conda activate starvla && export PYTHONPATH="$(pwd):$(pwd)/migrate_to_cloudrobo:${PYTHONPATH:-}" && export WANDB_MODE=disabled && accelerate launch --config_file ./starVLA/config/deepseeds/deepspeed_zero2.yaml --num_processes 2 starVLA/training/train_starvla.py --config_yaml ./migrate_to_cloudrobo/train_libero_cloudrobo.yaml --framework.name QwenGR00T --framework.qwenvl.base_vlm ./playground/Pretrained_models/Qwen3-VL-4B-Instruct --datasets.vla_data.data_root_dir ./playground/Dataset/ --datasets.vla_data.lerobot_version v3.0  --datasets.vla_data.data_mix libero --datasets.vla_data.per_device_batch_size 2 --trainer.freeze_modules '' --datasets.vla_data.video_backend torchvision_av --trainer.max_train_steps 1000 --trainer.save_interval 500 --trainer.logging_frequency 100 --trainer.eval_interval 100 --run_root_dir ./playground/Checkpoints --run_id test --wandb_project starVLA_Libero --wandb_entity npu_adapt
  • QwenFast任务
    cd /opt/cloud/starvla_ascend_notebook/starVLA && conda activate starvla && export PYTHONPATH="$(pwd):$(pwd)/migrate_to_cloudrobo:${PYTHONPATH:-}" && export WANDB_MODE=disabled && accelerate launch --config_file ./starVLA/config/deepseeds/deepspeed_zero2.yaml --num_processes 3 starVLA/training/train_starvla.py --config_yaml ./migrate_to_cloudrobo/train_libero_cloudrobo.yaml --framework.name QwenFast --framework.qwenvl.base_vlm ./playground/Pretrained_models/Qwen3-VL-4B-Instruct-Action --datasets.vla_data.data_root_dir ./playground/Dataset/ --datasets.vla_data.lerobot_version v3.0  --datasets.vla_data.data_mix libero --datasets.vla_data.per_device_batch_size 2 --trainer.freeze_modules '' --datasets.vla_data.video_backend torchvision_av --trainer.max_train_steps 1000 --trainer.save_interval 500 --trainer.logging_frequency 100 --trainer.eval_interval 100 --run_root_dir ./playground/Checkpoints --run_id test --wandb_project starVLA_Libero --wandb_entity npu_adapt
  • QwenPI任务
    cd /opt/cloud/starvla_ascend_notebook/starVLA && conda activate starvla && export PYTHONPATH="$(pwd):$(pwd)/migrate_to_cloudrobo:${PYTHONPATH:-}" && export WANDB_MODE=disabled && accelerate launch --config_file ./starVLA/config/deepseeds/deepspeed_zero2.yaml --num_processes 3 starVLA/training/train_starvla.py --config_yaml ./migrate_to_cloudrobo/train_libero_cloudrobo.yaml --framework.name QwenPI_v3 --framework.qwenvl.base_vlm ./playground/Pretrained_models/Qwen3-VL-4B-Instruct --datasets.vla_data.data_root_dir ./playground/Dataset/ --datasets.vla_data.lerobot_version v3.0  --datasets.vla_data.data_mix libero --datasets.vla_data.per_device_batch_size 2 --trainer.freeze_modules '' --datasets.vla_data.video_backend torchvision_av --trainer.max_train_steps 1000 --trainer.save_interval 500 --trainer.logging_frequency 100 --trainer.eval_interval 100 --run_root_dir ./playground/Checkpoints --run_id test --wandb_project starVLA_Libero --wandb_entity npu_adapt

推理部署启动命令

QwenOFT/QwenGR00T/QwenPI三个任务使用如下推理部署启动命令:

cd /opt/cloud/starvla_ascend_notebook/starVLA && conda activate starvla && export PYTHONPATH="$(pwd):$(pwd)/migrate_to_cloudrobo:${PYTHONPATH:-}" && python migrate_to_cloudrobo/deployment/server_starvla.py --ckpt_path ./playground/Checkpoints/test/pytorch_model.pt --base_vlm ./playground/Checkpoints/test/Qwen3-VL-4B-Instruct --port 8000

QwenFast任务使用如下推理部署启动命令:

cd /opt/cloud/starvla_ascend_notebook/starVLA && conda activate starvla && export PYTHONPATH="$(pwd):$(pwd)/migrate_to_cloudrobo:${PYTHONPATH:-}" && python migrate_to_cloudrobo/deployment/server_starvla.py --ckpt_path ./playground/Checkpoints/test/pytorch_model.pt --base_vlm ./playground/Checkpoints/test/Qwen3-VL-4B-Instruct-Action --port 8000

相关文档