Skip navigation
Introducing Halo, the best framework for post-training of open-source models.
0:00 / 0:42

Introducing Halo, the best framework for post-training of open-source models.

White Circle

@whitecircle2.5K followers

Visit on X
1,869 likesDownloadOriginal post

2,343,658 views

Introducing Halo, the best framework for post-training of open-source models. Halo delivers up to 2.8x the throughput of stock TRL with less peak memory, while models stay in their native HuggingFace format. Star us on GitHub:

56 Comments

Comments from the original post on X (Twitter)

  • White Circle's profile picture

    @whitecircle

    Every model we train at White Circle now uses Halo. The same codebase runs LoRA on a 24 GB GPU, multi-node training on B300s, and async RL. Technical blog:

    40 likes

  • White Circle's profile picture

    @whitecircle

    Other training frameworks often require a separate implementation for each model family. In Halo, adding a new model family takes just 100 lines of connecting a wrapper instead of rewriting the model.

    33 likes

  • White Circle's profile picture

    @whitecircle

    One YAML file contains the model, training method, GPU setup and checkpoint settings. One command starts training.

    29 likes

  • White Circle's profile picture

    @whitecircle

    Halo supports data, tensor, context, expert and expert-tensor parallelism. It also includes optimized attention, grouped-GEMM and fused loss kernels.

    29 likes

  • White Circle's profile picture

    @whitecircle

    Halo is not limited to supervised fine-tuning. It supports async RL and training with external environments. We are excited to partner with the @sgl_project team to make it the primary engine for rollouts.

    37 likes

  • White Circle's profile picture

    @whitecircle

    On @OpenAI gpt-oss-20b, Halo delivered 2.3–2.8x the throughput of stock TRL with less peak memory. Both runs used the same FlashAttention, Liger, fused cross-entropy and grouped-GEMM optimizations.

    30 likes

  • White Circle's profile picture

    @whitecircle

    We partnered with @liquidai to add native Halo support for LFM2.5-8B-A1B. On one B300, Halo was up to 20% faster than next best open-source framework.

    35 likes

  • White Circle's profile picture

    @whitecircle

    We also used Halo to fine-tune @Zai_org GLM-4.7-Flash on 177M tokens of agentic traces. The resulting model improved SWE-rebench-V2 by @nebiusai from 33% to 42%. Halo reached up to 1.63× TRL throughput on the same model precision and data. Model and write-up:

    30 likes

  • White Circle's profile picture

    @whitecircle

    Thanks for reading to the end! Star us on GitHub: Read more:

    35 likes

  • Daniel Smidstrup's profile picture

    @DanielSmidstrup

    Open source is the way, and it's only going to get more and more relevant in the coming time! :)

    10 likes

  • White Circle's profile picture

    @whitecircle

    True

    4 likes

  • leo's profile picture

    @leo1yu

    insane way to kick off the week -- congrats @whitecircle team

    9 likes

  • White Circle's profile picture

    @whitecircle

    Lets go!

    3 likes

  • neural nets.'s profile picture

    @cneuralnetwork

    crazy reading through the docs

    8 likes

  • Vishal Singh 🥑's profile picture

    @vishalsingh2972

    love that this is open source, congrats team ❤️

    8 likes

  • Catalin's profile picture

    @catalinmpit

    Congrats on the awesome work, team! Good to see training infra being open source 👏 Starred the repo to contribute to the growth.

    8 likes

  • White Circle's profile picture

    @whitecircle

    thanks!

    2 likes

  • Paul Mit's profile picture

    @pmitu

    great idea you've got my star, guys

    8 likes

  • giulia's profile picture

    @giutheginger

    took me one read to understand exactly who this is for. that almost never happens with infra launches. congrats team!!! starred 👀

    7 likes

  • Mac's profile picture

    @MacdevM

    Peak improvement, I will try it from first hand

    7 likes