Skip to content

[Debug] add ONLY_CALC_MISMATCH_RATIO and support cuda_device#1847

Merged
hhaAndroid merged 1 commit into
InternLM:mainfrom
hhaAndroid:cuda_device
May 27, 2026
Merged

[Debug] add ONLY_CALC_MISMATCH_RATIO and support cuda_device#1847
hhaAndroid merged 1 commit into
InternLM:mainfrom
hhaAndroid:cuda_device

Conversation

@hhaAndroid
Copy link
Copy Markdown
Collaborator

@hhaAndroid hhaAndroid commented May 27, 2026

新特性

  • 支持 ONLY_CALC_MISMATCH_RATIO 可以仅仅跑 RL 的 mismatch。不需要真的训练,适用于 debug + 少卡情况。比如 qwen35b 之前需要 8 卡,现在只需要 2 卡即可跑
  • 支持 CUDA_VISIBLE_DEVICES 方便指定任意卡数 debug
  • 支持 TRAIN_BATCH_SIZE,方便想快速 debug 或者在卡数减少情况下 debug 而无需修改配置

以上新特性都是为了更方便的 debug。用法如下:

export ONLY_CALC_MISMATCH_RATIO=1
export TRAIN_BATCH_SIZE=8
export CUDA_VISIBLE_DEVICES=2,3

bash examples/v1/scripts/run_rl.sh examples/v1/config/rl_qwen3p5_vl_35B_grpo_mixdata.py "lmdeploy" $QWEN3P5_VL_MODEL_PATH $META_DATA_PATH

可以进一步配合如下参数,实现更高效 debug

export DEBUG_ROLLOUT_DIR='work_dirs_rl/debug_rollout'
export DEBUG_TRAIN=False # True
export DEBUG_ROLLOUT=False  # True

@hhaAndroid hhaAndroid changed the title [Feature] add ONLY_CALC_MISMATCH_RATIO and support cuda_device [Debug] add ONLY_CALC_MISMATCH_RATIO and support cuda_device May 27, 2026
@hhaAndroid hhaAndroid merged commit 5d7b104 into InternLM:main May 27, 2026
6 of 7 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants