One pipeline, your GPUs

TT orchestrates training locally. Run GPU processes on your own machine or an existing EC2 instance in your AWS account. Dataset preparation, scoring, run history, and reports stay local.

tt validate tunedtensor.json
tt pipeline run --spec tunedtensor.json
tt runs list
tt runs report <run-id>

These commands need no Tuned Tensor account or access token.

Use your AWS GPU

Add GPU connection settings to local-runner.json beside your spec:

{
  "gpu": {
    "provider": "aws",
    "profile": "research",
    "region": "eu-west-1",
    "instanceId": "i-0123456789abcdef0",
    "user": "ubuntu",
    "identityFile": "~/.ssh/research-gpu.pem"
  }
}

The AWS profile resolves your instance; SSH provides machine access. TT requires a running instance, a verified SSH host key, and noninteractive authentication. You provide capacity and manage the instance lifecycle. AWS charges your account directly. The instance stays running afterward.

Adapter training and CUDA evaluation use the remote GPU. Foundation pretraining, fine-tuning, RL, and evaluation do too; tokenization stays local. Keep pipeline targets local. No Step Functions pipeline is involved.

tt models prefetch tunedtensor.json --config local-runner.json
tt doctor tunedtensor.json --config local-runner.json
tt pipeline run --spec tunedtensor.json --config local-runner.json

See the AWS GPU setup guide for dependencies, private networking, cancellation, and checkpoint recovery. Keep your laptop running for orchestration. GPU work has a configurable remote timeout; this is not a detached scheduler.

Dashboard and historical runs

Hosted training submission and estimates are retired. Existing reports remain available in the dashboard. Publish local run evidence when you want to share it there.

tt auth login
tt publish <local-run-id>
tt cloud runs list
tt cloud runs report <run-id>
tt cloud runs cancel <older-active-run-id>