GPU Training Pipeline

One agent that collects data continuously, retrains a model on a GPU, and serves the results on a dashboard.

Prerequisites#

• Expert subscription (required for the A100 80GB GPU tier)
• SOL deposited in Settings > Funds for compute runway

Create a GPU agent#

1. Create a new agent named sol_gpu_lab.
2. On the create form:
   • choose the GPU machine tier
   • give it extra disk if you expect a larger dataset
   • add initial wallet funding if you want the agent to own funds directly
   • decide whether to self-fund or bill from your main deposit

Build the pipeline#

Give the agent a compound training prompt:

Build a research pipeline for SOL trading signals.

Requirements:
- collect minute candles, volume, and derived features continuously
- keep raw data and cleaned training datasets in separate folders
- train a PyTorch model on the GPU to predict short-term directional moves
- save checkpoints and evaluation reports after each training run
- expose a dashboard on port 3000 showing dataset freshness,
  latest training metrics, confusion matrix, and current live signal
- leave the data collection process running

Also set up a cron job inside the VM so new training runs happen every night.

The agent builds the collector, feature pipeline, trainer, nightly cron schedule, checkpoint storage, and dashboard.

Monitor and refine#

1. Expose port 3000 and inspect the dashboard.
2. Push the workflow further:

Add model versioning, a training-run history page, and a comparison view showing
the last 5 checkpoints by validation accuracy, Sharpe proxy, and max drawdown.

1. Attach a custom domain like gpu-lab.yourdomain.com for collaborators.

When to use this#

This workflow fits when you need one persistent machine that both collects data continuously and periodically runs heavyweight GPU training jobs.