Run Experiment

Deploy and run ML experiment: $ARGUMENTS

Workflow

Step 1: Detect Environment

Read the project's AGENTS.md to determine the experiment environment:

• Local GPU: Look for local CUDA/MPS setup info
• Remote server: Look for SSH alias, conda env, code directory
• Vast.ai instance: Look for gpu: vast, vast_instance, SSH host/port, remote path, and optional auto_destroy
• Modal serverless: Look for gpu: modal, app/function name, image/dependency setup, and secrets

If no server info is found in AGENTS.md, ask the user.

Step 2: Pre-flight Check

Check GPU availability on the target machine:

Remote:

ssh <server> nvidia-smi --query-gpu=index,memory.used,memory.total --format=csv,noheader

Local:

nvidia-smi --query-gpu=index,memory.used,memory.total --format=csv,noheader
# or for Mac MPS:
python -c "import torch; print('MPS available:', torch.backends.mps.is_available())"

Free GPU = memory.used < 500 MiB.

Step 3: Sync Code (Remote Only)

Check the project's AGENTS.md for a code_sync setting. If not specified, default to rsync.

#### Option A: rsync (default)

Only sync necessary files — NOT data, checkpoints, or large files:

rsync -avz --include='*.py' --exclude='*' <local_src>/ <server>:<remote_dst>/

#### Option B: git (when code_sync: git is set in AGENTS.md)

Push local changes to remote repo, then pull on the server:

# 1. Push from local
git add -A && git commit -m "sync: experiment deployment" && git push

# 2. Pull on server
ssh <server> "cd <remote_dst> && git pull"

Benefits: version-tracked, multi-server sync with one push, no rsync include/exclude rules needed.

#### Option C: Vast.ai instance

If gpu: vast is configured, treat the Vast.ai machine as a remote server with an explicit lifecycle:

1. Verify the instance is running and reachable.
2. Sync code to the configured remote path.
3. Confirm data/checkpoints are already mounted or intentionally copied.
4. Record the instance id in the launch summary for later cleanup.

Do not silently ignore a requested Vast.ai route. If Vast.ai CLI credentials or instance metadata are missing, stop and ask the user to configure them.

Step 3.5: W&B Integration (when wandb: true in AGENTS.md)

Skip this step entirely if `wandb` is not set or is `false` in AGENTS.md.

Before deploying, ensure the experiment scripts have W&B logging:

1. Check if wandb is already in the script — look for import wandb or wandb.init. If present, skip to Step 4.

1. If not present, add W&B logging to the training script:

   import wandb
   wandb.init(project=WANDB_PROJECT, name=EXP_NAME, config={...hyperparams...})

   # Inside training loop:
   wandb.log({"train/loss": loss, "train/lr": lr, "step": step})

   # After eval:
   wandb.log({"eval/loss": eval_loss, "eval/ppl": ppl, "eval/accuracy": acc})

   # At end:
   wandb.finish()

1. Metrics to log (add whichever apply to the experiment):

• train/loss — training loss per step
• train/lr — learning rate
• eval/loss, eval/ppl, eval/accuracy — eval metrics per epoch
• gpu/memory_used — GPU memory (via torch.cuda.max_memory_allocated())
• speed/samples_per_sec — throughput
• Any custom metrics the experiment already computes

1. Verify wandb login on the target machine:

   ssh <server> "wandb status"  # should show logged in
   # If not logged in:
   ssh <server> "wandb login <WANDB_API_KEY>"

The W&B project name and API key come from AGENTS.md (see example below). The experiment name is auto-generated from the script name + timestamp.

Step 4: Deploy

#### Remote (via SSH + screen)

For each experiment, create a dedicated screen session with GPU binding:

ssh <server> "screen -dmS <exp_name> bash -c '\
  eval \"\$(<conda_path>/conda shell.bash hook)\" && \
  conda activate <env> && \
  CUDA_VISIBLE_DEVICES=<gpu_id> python <script> <args> 2>&1 | tee <log_file>'"

#### Vast.ai instance

Use the same SSH + screen pattern, but include the Vast.ai instance id, public SSH endpoint, and remote working directory in the report. If auto_destroy: true, write a cleanup command to the run notes before launch.

Record the estimated hourly cost, expected run duration, and cleanup owner. If the command fails to start or the instance becomes unreachable, do not relaunch blindly; capture logs and ask for a rescue / second opinion before spending more GPU time.

#### Modal (serverless)

If gpu: modal is configured, deploy through Modal instead of SSH:

modal run <module_or_app>.py -- <args>

Before launch, verify required secrets, volumes, image dependencies, and output persistence. If Modal is requested but the project lacks Modal configuration, stop and ask the user to configure it rather than falling back to local execution.

Record the Modal app/function name, GPU type, timeout, mounted volumes, and where results will be stored. If Modal reports an image, secret, or volume error, preserve the exact error and run a configuration fix before retrying.

#### Local

# Linux with CUDA
CUDA_VISIBLE_DEVICES=<gpu_id> python <script> <args> 2>&1 | tee <log_file>

# Mac with MPS (PyTorch uses MPS automatically)
python <script> <args> 2>&1 | tee <log_file>

For local long-running jobs, use run_in_background: true to keep the conversation responsive.

Step 5: Verify Launch

Remote:

ssh <server> "screen -ls"

Local: Check process is running and GPU is allocated.

Step 6: Feishu Notification (if configured)

After deployment is verified, check ~/.codex/feishu.json:

• Send experiment_done notification: which experiments launched, which GPUs, estimated time
• If config absent or mode "off": skip entirely (no-op)

Step 7: Auto-Destroy Vast.ai Instance (when gpu: vast and auto_destroy: true)

Only run this after the experiment has completed and results/logs/checkpoints have been copied or otherwise persisted.

1. Verify the target process has exited.
2. Copy result files and logs to the configured durable location.
3. Ask for confirmation unless AGENTS.md explicitly says auto_destroy: true.
4. Destroy only the recorded instance id for this run.

If any artifact copy fails, do not destroy the instance.

Key Rules

• ALWAYS check GPU availability first — never blindly assign GPUs
• Each experiment gets its own screen session + GPU (remote) or background process (local)
• Use tee to save logs for later inspection
• Run deployment commands with run_in_background: true to keep conversation responsive
• Report back: which GPU, which screen/process, what command, estimated time
• If multiple experiments, launch them in parallel on different GPUs

AGENTS.md Example

Users should add their server info to their project's AGENTS.md:

## Remote Server
- SSH: `ssh my-gpu-server`
- GPU: 4x A100 (80GB each)
- Conda: `eval "$(/opt/conda/bin/conda shell.bash hook)" && conda activate research`
- Code dir: `/home/user/experiments/`
- code_sync: rsync          # default. Or set to "git" for git push/pull workflow
- wandb: false              # set to "true" to auto-add W&B logging to experiment scripts
- wandb_project: my-project # W&B project name (required if wandb: true)
- wandb_entity: my-team     # W&B team/user (optional, uses default if omitted)

## Vast.ai
- gpu: vast
- vast_instance: 123456
- SSH: `ssh -p 12345 [email protected]`
- Code dir: `/workspace/experiments/`
- auto_destroy: false

## Modal
- gpu: modal
- modal_app: `train.py`
- modal_secrets: `wandb-secret`
- modal_volume: `experiment-results`

## Local Environment
- Mac MPS / Linux CUDA
- Conda env: `ml` (Python 3.10 + PyTorch)

W&B setup: Run wandb login on your server once (or set WANDB_API_KEY env var). The skill reads project/entity from AGENTS.md and adds wandb.init() + wandb.log() to your training scripts automatically. Dashboard: https://wandb.ai/<entity>/<project>.