Anyscale and Novasky Team Release Skyrl TX V0.1.0: Bringing TIKINSER Compatible RL Engine Learning to GPU Cluster

How can AI teams be strengthened by Tinker-style consolidation in the form of languages ​​that run their infrastructure through a single, unified engine? Anyscale and Novasky (UC Berkeley) Group tilt Skyrl tx v0.1.0 That gives developers a way to run Tinker's parallel training and lift engine directly on their hardware, while keeping a small, expensive API available.

The research team explains Skyrl tx As an integrated training and benchmarking engine the Tinker API is used and allows people to use Tinker as a service in their infrastructure. This version of V0.1.0 is the first of its series, it supports end-to-end read-only reinforcement, and it does resampling very quickly.

Tinker API in a nutshell

Tinker from Thinking Machines is a training API built around four key functions. forward_backward Make forward passes and backward passes and accumulate gradients. optim_step The model parameters are updated according to those gradients. sample Tokens are made for communication, testing or RL action. save_state You write CheckPoints to resume training.

Instead of a full version of the good order, the tinker presents these low-level primitives so that users can use their supervised or reinforced holes in the standard Python code, while the service handles the programming and distribution of the GPU.

Skyrl TX targets this api specifically and uses an open backend that users can post locally. It maintains the Tinker Programming model, while removing the need to rely solely on a managed environment.

Where skyrl tx fits inside skyrl

Skyrl is a complete learning library for styling reinforcement for major programming languages skyrl-agent For long-term agents, skyrl-train training, too skyrl-gym To find the tool use areas like statistics, coding, search and SQL.

Inside this stack, skyrl-tx marked as a test platform library that exposes a local tinker like rest api for model training. Skyrl TX therefore becomes the programming layer that connects RL Logic, environments and training code to GPU services charged by Tinker virtual concrete.

Buildings, engine construction and training

The Skyrl TX build is described as a lift engine and supports back-to-back passes. It has four main elements:

1. Call the API Server that processes incoming requests from various users.
2. The last place That tracks Metadata with models, checkpoints, requests and futures, and also works as a workflow. The current implementation uses SQLite behind the scenes which supports other SQL databases such as Postgres.
3. The machine Those schedules and batteries are applications for all users. Each motor center serves one basic model and can attach multiple lora adapters.
4. Work That retrieves the past and reverses and holds the model definitions and optimizer states. Many workers would be making powerful improvements to node sharding in future versions

What does v0.1.0 add?

The V0.1.0 Release focuses on reinforcement learning support and performance improvements. The official release highlights some major concrete changes:

• The Sampler is now very fast, because it is integrated and well spaced and tuned to the engine.
• Unique per-request Sample parameters, per-request seed and stop tokens are now supported, useful when multiple tests share the same underlying model.
• After several fixes, the RL loop now works fine with the engine.
• Gradient composite point support and small sample maturity are used.
• Postgres is now supported as a Database Backend, alongside SQLITE.

Running out of RL is the only end to the 8 H100 GPUS

The official release contains a specific code recipe to find the end of the readout to finish on a cluster with 8 H100 GPUS.

First, users wake up the Skyrl recovery site and on skyrl-tx The folder starts the engine with:

uv run --extra gpu --extra tinker -m tx.tinker.api 
  --base-model Qwen/Qwen3-4B 
  --max-lora-adapters 3 
  --max-lora-rank 1 
  --tensor-parallel-size 8 
  --train-micro-batch-size 8 > out.log

Then they put together a tinkerer's cookbook for a group of thinking machines here tinker_cookbook/recipes Run Folder:

export TINKER_API_KEY=dummy
export WANDB_API_KEY=
uv run --with wandb --with tinker rl_loop.py 
  base_url= 
  model_name="Qwen/Qwen3-4B" 
  lora_rank=1 
  max_length=1024 
  save_every=100

This generates a reward curve that ensures the RL loop is working properly with the skyrl tx backend.

Key acquisition

• Skyrl TX V0.1.0 uses a local, compatible tinker engine that combines LLM Post Training training and submission.
• The program presents tinker primitives, forward_aclacks
• The structure is divided into API server, SQL database, programming engine and front-end and back-end workers through a single basic model.
• V0.1.0 Adds end-to-end support for reinforcement learning, fast and sharded sampling, per-samples propameter, gradient tracking and postgres support.

Skyrl TX V0.1.0 is a practical step for dev teams who want Tinker-style reinforcement for their methods with their collections that don't have a fixed place for the tinker API. A design that treats the system like a lift engine and runs back to back is clean and reduces stack variability. Lora support, gradient tracking, microbetting and postgres are concrete improvements. Overall, this release changes the compatibility of the tinker in the RL of the RL working area of ​​the LLM

Look It's a waste and Official release. Feel free to take a look at ours GitHub page for tutorials, code and notebooks. Also, feel free to follow us Kind of stubborn and don't forget to join ours 100K + ML Subreddit and sign up Our newsletter. Wait! Do you telegraph? Now you can join us by telegraph.