| # OPD-only ChartQA：8×H GPU 部署 |
|
|
| 在现有 [`config/config_opd_only_7b_chartqa.yaml`](../config/config_opd_only_7b_chartqa.yaml) 中直接修改： |
|
|
| ```yaml |
| paths: |
| # 唯一的项目内路径基目录；project://... 会由 loader 拼接到这里 |
| project_root: /workspace/agentic-rl-main |
| |
| model: |
| # SFT 完成后的本地完整 Hugging Face checkpoint（必须含 config.json 与模型权重） |
| pretrained_model_path: /mnt/models/chartqa-sft/final_checkpoint |
| # 本地冻结 7B teacher |
| teacher_model_path: /mnt/models/llava-onevision-qwen2-7b-ov |
| teacher_device_map: same |
| |
| training: |
| stage: opd_only |
| num_gpus: 8 |
| num_client: 8 |
| dyme_args: |
| output_dir: /mnt/experiments/opd-only-chartqa-8h |
| per_device_train_batch_size: 2 |
| gradient_accumulation_steps: 16 |
| |
| dataset: |
| train_dataset: project://data/chartqa/train_new_prerefine_vf_full_real_deplot_fp32_qwen25.json |
| ``` |
|
|
| 项目内的 `project://outputs/...`、`project://data/...` 会自动使用该基目录，无需逐项修改。若显存不足，设 `per_device_train_batch_size: 1`、`gradient_accumulation_steps: 32`。 |
|
|
| ## 8 卡启动 |
|
|
| ```bash |
| cd /workspace/agentic-rl-main |
| mkdir -p logs |
| |
| CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \ |
| accelerate launch \ |
| --config_file default_config_8gpu.yaml \ |
| --num_processes 8 \ |
| --main_process_port 29531 \ |
| main.py \ |
| --config config/config_opd_only_7b_chartqa.yaml \ |
| --mode rl \ |
| 2>&1 | tee logs/opd_only_chartqa_8h.log |
| ``` |
|
|
| 不要用 `DYME_*` 环境变量覆盖训练配置。 |
|
|
| ## 运行后检查 |
|
|
| ```bash |
| rg -n "training_stage=opd_only|opd_route_count|grpo_route_count|sft_route_count|loss/opsd|teacher_trajectory|teacher_probe|visual" \ |
| logs/opd_only_chartqa_8h.log |
| ``` |
|
|
| 应确认： |
|
|
| - `training_stage=opd_only` |
| - `opd_route_count` 覆盖全部 completion |
| - `grpo_route_count=0`、`sft_route_count=0` |
| - `loss/opsd` 有限且非零 |
| - teacher probe、trajectory、visual checker、refiner 都有输出 |
|
|
| 运行快照位于 `training.dyme_args.output_dir`，包括 `resolved_config.yaml`、`run_env.json` 与 `logs/`。 |
|
|
