Training - Veri Documentation

Training on Veri has two paths:

Custom Training

Bring your own training code, framework, entrypoint, and dependencies. Veri provisions the GPU and runs it.

Managed Training

Use Veri-managed GRPO, SFT, or DPO loops. You provide data, model, rewards or preferences, and hyperparameters.

Which should I use?

Path	Use it when
Custom Training	You already have a `torchrun`, OpenRLHF, slime, verl, or custom Python training loop.
Managed Training	You want Veri to own the trainer and orchestration for GRPO, SFT, or DPO.

Both paths run on Veri-provisioned GPUs, stream logs and metrics, and capture the trained checkpoint into your account. Managed jobs also save intermediate checkpoints while training runs, so an interrupted job still leaves usable weights.