메뉴
Out-of-memory and resource pressure recovery without changing the user's training objective.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
MaxEvolve Triton/PyTorch to JAX/Pallas porting with GpuToJax 12-step pipeline, interpret=True validation, MAP-Elites evolution, and Colab/SSH hardware profiling.
Manage Google Colab sessions and execute code on remote Colab VMs via the colab-cli.
Use when you need to generate a synthetic dataset. You can use Afterimage Python package to create, conversational, tool-calling, multi-choice questions, or other types of structured output datasets for SFT, or preference data for DPO.
Schedule deferred or recurring work via the isanagent cron tool—add/remove jobs with correct runtime context fields.
Configure long wall clocks and use execution_run_background + job polling for ML or hours-long runs without blocking the agent loop.
Run numeric or ML code through isanagent execution tools without flooding context—artifacts, incremental runs, sandbox paths.
| name | oom-recovery-playbook |
| description | Out-of-memory and resource pressure recovery without changing the user's training objective. |
| always | false |
When a run fails with CUDA OOM, host OOM, or timeout:
per_device_train_batch_size (or equivalent) and raise gradient_accumulation_steps so effective batch size stays the same when possible.max_length / context window (changes what data the model sees).Re-run a small pilot, confirm loss/metrics move as expected, then restore full scale.