Training on Baseten

Train models on managed GPUs with Truss Train or Loops and deploy any checkpoint to production inference.

Baseten trains models on managed GPUs and deploys the resulting checkpoints to production inference on the same platform. There are two ways to train: Truss Train, where you bring your own container and training code, and Loops, a Tinker-compatible SDK for LoRA fine-tuning and RL on a curated set of base models.

Choose your path

• Coming from Tinker: Loops is a Tinker-compatible backend for your existing scripts. import tinker works with one install change, and the Tinker compatibility page lists exactly what differs. Start with the Loops overview.
• Already training elsewhere: Your Axolotl config, TRL script, or custom loop runs unchanged in a container. Baseten provisions the GPUs, syncs checkpoints as your job saves them, and deploys any checkpoint as a production endpoint. That’s Truss Train, documented in this section.

If you match both or neither, the paths differ in who drives. With Truss Train, you hand Baseten a program: a batch job that runs to completion on hardware you declare. With Loops, your program calls Baseten: each training step is an API call to a live trainer, and a paired sampler serves the latest weights throughout. Loops covers LoRA fine-tuning and RL on a curated model list and is in early access; request access for your workspace. Truss Train runs any training code and is available to every workspace today.

	Truss Train	Loops
Training code	Any container image	Tinker-compatible Python (import tinker)
Models	Any	Supported base models only
Hardware	You declare GPUs in a Truss config	Baseten picks GPUs for the base model
Inference path	Deploy any synced checkpoint with one CLI command	Sampler serves new weights live during training; deploy checkpoints when ready
Lifecycle	Job runs to completion and exits	Session stays live until you run truss loops deactivate
Availability	All workspaces	Early access
Documentation	This section	Loops docs

The rest of this page covers Truss Train.

How Truss Train works

Baseten stores your checkpoints while the job runs and deploys any of them as a production endpoint. You don’t download weights, re-upload them, or manage separate serving infrastructure. The core workflow is two commands:

# Train your model
truss train push config.py

# Deploy from the checkpoint
truss train deploy_checkpoints --job-id <job_id>

From job submission to a served model:

1. Define your job: Declare compute, container image, runtime, and checkpointing in a Python config file.
2. Submit it: truss train push packages your code and starts the job on H100 or H200 GPUs, single-node or multi-node.
3. Watch checkpoints sync: Baseten stores each checkpoint your job saves.
4. Deploy a checkpoint: truss train deploy_checkpoints turns any synced checkpoint into a production endpoint.

Supported frameworks

Truss Train is framework-agnostic: if it runs in a container, it runs here.

Framework	Best for	Example
Axolotl	Configuration-driven fine-tuning with LoRA/QLoRA	oss-gpt-20b-axolotl
TRL	SFT, DPO, and GRPO with Hugging Face	oss-gpt-20b-lora-trl
TRL	LoRA DPO fine-tuning	qwen3-8b-lora-dpo-trl
VeRL	Reinforcement learning with custom rewards	qwen3-8b-lora-verl
MS-Swift	Long-context and multilingual training	qwen3-30b-mswift-multinode

Browse the ML Cookbook for more examples including multi-node training with FSDP and DeepSpeed.

Key features

Checkpoint management

Checkpoints sync automatically to Baseten storage during training. You can:

• Deploy any checkpoint as a production endpoint with truss train deploy_checkpoints.
• Download checkpoints for local evaluation and analysis.
• Resume from any checkpoint if a job fails or you want to train further.

Learn more about checkpointing.

BDN weight and data loading

Load model weights and training data through Baseten Delivery Network (BDN). Mount weights from Hugging Face, S3, GCS, R2, or any HTTPS URL directly into your training container with no download code needed. BDN mirrors weights before compute is provisioned, then caches them for faster mounting on subsequent jobs.

See storage and data ingestion for setup details.

Persistent caching

Speed up training iterations by caching models, datasets, and preprocessed data between jobs. The cache persists across training runs, so you don’t re-download 70B models every time.

See the training cache guide for configuration options.

Multi-node training

Scale training across multiple GPU nodes with InfiniBand networking. Baseten handles node orchestration, communication setup, and environment variables. You set node_count in your configuration.

Learn more about multi-node training.

Remote access

Connect to running training containers to debug, inspect state, and iterate without resubmitting. Baseten offers two options:

• SSH: Connect from any OpenSSH client for terminal sessions and file transfer with scp or sftp.
• VS Code & Cursor: Connect from VS Code or Cursor Remote Tunnels for a full IDE experience.

See the Remote access overview to choose between them.

Next steps

Get started

Run your first training job and deploy the result.

Loops

Tinker-compatible SFT and async RL with checkpoint deploys to inference.

ML Cookbook

Production-ready examples for frameworks and models.

Reference

• CLI reference
• SDK reference
• API reference