Operating AutoDL Training

Use this skill for remote training operations on an AutoDL Linux server. It is designed for high-frequency workflows around "start training, watch progress, inspect resources, read logs, diagnose failures, and decide what to do next" while keeping execution constrained to one configured project directory.

What This Skill Does

• - Starts a configured training command in the target project directory over SSH.
• Activates the remote Python environment with Conda or virtualenv fallbacks.
• Checks whether training is still running by combining process, GPU, and log freshness signals.
• Summarizes GPU, CPU, memory, and disk pressure instead of dumping raw command output.
• Reads recent logs and extracts likely metrics such as epoch, step, loss, lr, grad_norm, val_loss, accuracy, mAP, and F1.
• Detects common training failures such as CUDA OOM, NCCL errors, NaN, disk full, timeout, and segmentation faults.
• Produces a human-readable training summary and recommends whether to continue, tune, or resume from a checkpoint.

Required Inputs

Collect or confirm these values before running any script:

• - host: AutoDL server hostname or IP.
• INLINECODE10: SSH port, usually 22.
• INLINECODE12: Remote Linux username.
• INLINECODE13: Absolute project directory on the remote server, for example /root/autodl-tmp/your-project.
• One environment option: env_name, env_activate, or venv_path.
• INLINECODE18: The training launch command, such as python train.py, python -m torch.distributed.run ..., or bash scripts/train.sh.
• Optional password mode: provide AUTOCLAW_TRAIN_SSH_PASSWORD as an environment variable or local .env file when SSH key login is not available.

Prefer a config file at config.example.json copied to a real file such as config.json, or environment variables based on .env.example.

Safety Rules

• - Only operate inside the configured project_path.
• Do not invent missing SSH credentials or secrets.
• Do not write plaintext passwords into files.
• Prefer SSH keys or environment variables.
• Refuse obviously destructive launch commands such as rm -rf, reboot, shutdown, mkfs, or fork bombs.
• Do not kill unrelated processes or run global destructive recovery commands.

Workflow

1. Confirm Configuration

Read config.example.json and references/usage.md to understand the expected fields. Ask the user for any missing values instead of guessing.

2. Start Or Resume Training

Run scripts/remote_train.py to start a background job or build a resume command:

CODEBLOCK0

Use this when the user asks to launch training, re-launch after interruption, or resume from a checkpoint.

3. Check Live Status

Run scripts/check_status.py when the user asks whether training is still running:

CODEBLOCK1

This script combines process matching, nvidia-smi, and recent log updates to classify the run as running, stopped, failed, or unknown.

4. Inspect Resource Pressure

Run scripts/monitor_resources.py to summarize GPU/CPU/memory/disk usage:

CODEBLOCK2

Use the human-readable bottleneck assessment in the output instead of pasting raw command output unless the user asks for raw data.

5. Read Logs And Summaries

Run scripts/summarize_log.py in one of these modes:

CODEBLOCK3

Use read for recent excerpts and metrics, detect-failure for exception diagnosis, and summarize for a concise human-facing assessment with next steps.

Script Map

• - scripts/remote_train.py: start training, optional resume templating, structured launch result.
• INLINECODE47: process/GPU/log-based training status.
• INLINECODE48: GPU/CPU/memory/disk summary and bottleneck hints.
• INLINECODE49: read logs, detect failures, summarize convergence and next actions.
• INLINECODE50: shared config loading, SSH execution, safe path checks, remote helpers.
• INLINECODE51: reusable log parsing, failure detection, trend analysis, recommendation logic.

References

• - Read references/usage.md for setup steps, example configs, and example commands.
• Read references/troubleshooting.md when SSH, environment activation, logs, or training recovery fail.

Agent Guidance

• - Start with the least invasive action that answers the user’s request.
• When the user asks a yes/no status question, prefer scripts/check_status.py before reading a long log.
• When the user asks why training stopped, run scripts/check_status.py and then scripts/summarize_log.py --action detect-failure.
• When the user asks whether to continue training, run scripts/summarize_log.py --action summarize and include the recommendations from the script in the final response.
• When a checkpoint path is provided, prefer scripts/remote_train.py --resume-from ... so the resume command is explicit and auditable.