Practical overview
When it fits
Use Halo when you need a training framework that brings several model-training methods and distributed execution options into a shared PyTorch-based stack, while retaining Hugging Face model and checkpoint workflows.
Who it is for
- Researchers and engineering teams training or fine-tuning language and multimodal models.
- Practitioners building distributed training runs across GPUs or nodes.
- Teams developing asynchronous multi-turn reinforcement-learning workflows with vLLM or SGLang inference servers.
Documented uses
- Run supervised fine-tuning, including a documented single-GPU LoRA example.
- Configure distributed SFT with Expert Parallelism or FSDP2.
- Train models through pre-training, preference optimization or distillation using shared infrastructure.
- Set up asynchronous multi-turn RL rollouts against vLLM or SGLang servers.
Requirements & limits
The repository describes Halo as open-source. Its README says the documented Blackwell and Hopper Docker images can be pulled anonymously without an account.
- The documented quick-start images target Blackwell (B200/B300) and Hopper (H100/H200) architectures; the README also describes local image builds.
- Parallelism choices include EP, CP, TP and ETP alongside FSDP2, but the README says certain combinations are rejected at configuration time.
- The RL workflow uses Ray actors for environment rollouts and vLLM or SGLang servers for inference, separate from the Transformers training side.
- Halo adds parallelism to existing models rather than requiring a separate distributed model implementation, according to the README.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Prepare a Halo training container
Pull the documented Docker image for a supported GPU architecture, or build the image locally.
Before you start
- Docker for pulling or building an image.
- For the documented image choices: Blackwell (B200/B300) or Hopper (H100/H200) GPU architecture.
Steps
For Blackwell, pull
public.ecr.aws/whitecircle/halo:blackwell; for Hopper, pullpublic.ecr.aws/whitecircle/halo:hopper.Alternatively, build locally with the command below or the command below, matching the target architecture.
make build-blackwellmake build-hopper
Expected outcome
A Halo container image with the documented PyTorch, Flash Attention and DeepEP V2 components.
Guide limitations
- The images are architecture-specific; the README says there is no
latesttag. - The README documents image preparation, not a complete host setup or a verified training run.
Launch a documented SFT configuration
Start a supervised fine-tuning run from a Halo example YAML configuration.
Before you start
- A prepared Halo environment.
- A Halo SFT example configuration from the repository.
- For the single-GPU LoRA example, the README identifies
examples/sft/qwen3/qwen3-4b-ultrachat-lora.yaml; the multi-GPU examples specify their launcher and GPU count.
Steps
Choose an SFT YAML example suited to the intended setup.
Run
halo launch sftfollowed by the configuration path.For the documented eight-GPU Expert Parallelism example, use
halo launch sft examples/sft/qwen3_5/qwen3.5-35b-a3b-ultrachat-ep.yaml -n 8.
Expected outcome
Halo selects a launcher such as `python`, `torchrun` or `accelerate launch` according to the supplied arguments and configuration.
Guide limitations
- The example commands describe launch patterns; they do not establish that an arbitrary model, configuration or machine is compatible.
- The README documents
--dry-runfor printing the command without running it.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 18 October 2026
- Original website title
- whitecircle/halo · Official homepage title · retained in its source language