Tyler Griggs, Philipp Moritz, Eric Tang, and the SkyRL Team

🗓️ Posted: February 13, 2025

<aside>

We are excited to announce that SkyRL now implements the Tinker API — a simple, powerful training API introduced by Thinking Machines Lab.

Today, any training script written in the Tinker API can run locally on your own GPUs using SkyRL's high-performance backends, with zero code changes.

Get started now with the Quickstart.

</aside>

Why Tinker?

Today, post-training LLMs in open source is fragmented. When a new training method is published, the code only runs on the author’s framework of choice. Porting it to your own preferred stack takes days or weeks of engineering, if it happens at all. The open-source post-training community is leaving one of its greatest assets on the table: sharing.

We believe post-training can be greatly benefited by an open training standard, akin to the OpenAI API’s role in bringing wide interoperability to inference. After extensively using the Tinker API in the SkyRL tx project, we have found it to be the most compelling opportunity for a widely-adopted training API. Why?

API at the right level. Tinker's API surface is small and simple:

• forward_backward() — run a forward and backward pass
• optim_step() — apply an optimizer step
• sample() — generate rollouts from the current model
• save_checkpoint() / load_checkpoint() — persist and restore training state

That's it. We've found this simple surface very effective at separating infrastructure logic from the algorithm and dataflow of post-training.

Below the API: a Tinker backend handles infrastructure concerns like modal sharding, gradient management, weight sync primitives, and model onloading & offloading.

Above the API: researchers have full flexibility to compose Tinker’s low-level primitives to implement arbitrary post-training algorithms. Whether it’s SFT, RL, DPO, or a new research idea, a user need only focus on the algorithm and the backend remains agnostic.

This well-separated contract between the research and infrastructure is what makes Tinker both simple and powerful to build on.

Tinker on your own GPUs