Practical overview
When it fits
Use Orbit when you need an adapter-based reinforcement-learning post-training workflow and can provide the specified CUDA environment, model checkpoints, training data and backend checkouts.
Who it is for
- Researchers and engineers implementing reinforcement-learning post-training for large language models.
- Teams with CUDA server infrastructure and experience managing model data and distributed-training checkpoints.
Documented uses
- Adapt a language model using reinforcement-learning post-training while keeping its base at low precision.
- Choose and launch documented BF16 or FP8 OFT example recipes.
- Prepare the framework's Python and CUDA environment with its pinned training and inference backends.
Requirements & limits
The repository is public and licensed under Apache-2.0.
- The public launcher path specifies Python 3.12, CUDA 13.2 and PyTorch 2.11.
- The documented setup expects sibling Megatron-Bridge, Megatron-LM and SGLang checkouts at pinned revisions.
- Training requires a prompt JSONL file, a Hugging Face checkpoint and a Megatron checkpoint root; evaluation JSONL is conditional on evaluation being enabled.
- The project documents INT4, NVFP4, FP8 and BF16 support, with BF16 LoRA and OFT adapters.
- The full source build is estimated at around one to two hours on a CUDA 13.2 and B200 machine.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up Orbit's documented environment
Prepare the Python environment and build the documented CUDA stack from the Orbit checkout.
Before you start
- Python 3.12
- CUDA 13.2 and a CUDA-capable environment
- Sibling checkouts of Orbit, Megatron-Bridge, Megatron-LM under Megatron-Bridge/3rdparty, and SGLang at the pinned backend revisions
- uv
Steps
Change to the Orbit checkout.
Pin Python 3.12 with uv.
Source env.sh to configure detected CUDA, GPU architecture and Python settings.
Run uv sync --extra allinone to build the documented stack from source.
Expected outcome
An environment prepared with Orbit and its backend stack for use with the project launchers.
Guide limitations
- The documented public launcher path targets Python 3.12, CUDA 13.2 and PyTorch 2.11.
- The source build is estimated at around one to two hours on a CUDA 13.2 and B200 machine.
- Orbit currently points to local backend paths and expects the sibling repositories at pinned revisions.
Launch the Qwen3-4B BF16 OFT example
Use the documented high-precision example launcher with supplied training and checkpoint paths.
Before you start
- A prepared Orbit environment and required backend checkouts
- Training prompt JSONL file
- Hugging Face checkpoint directory
- Megatron distributed checkpoint root
- CUDA-capable server environment
Steps
Set TRAIN_JSONL, HF_CKPT and MEGATRON_LOAD to the required data and checkpoint paths.
Set ORBIT_VENV and CUDA_HOME where needed for the environment.
Run bash examples/high_precision/run-qwen3-4b-instruct-2507-bf16-math-oft.sh from the Orbit checkout.
Provide TEST_JSONL if evaluation is enabled; the documented DISABLE_EVAL=1 setting skips evaluation.
Expected outcome
The Qwen3-4B BF16 OFT example launcher is invoked using the configured inputs.
Guide limitations
- The guide documents the launcher and its required inputs, not a guaranteed successful training result.
- GPU model and other site-specific settings may need adjustment for the local environment.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 14 October 2026
- Original website title
- Sphere-AI-Lab/orbit · Official homepage title · retained in its source language