Runtime AI safety and alignment infrastructure
- Tweets 122
- Following 19
- Followers 2,106
- Likes 497
Pinned Tweet
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: github.com/whitecircle/halo
166
218
1,897
2,331,297
We're looking for a Production Manager at
@whitecircle
whitecircle.com/careers/prod…
big green flags
- you’ve worked with brands like A24, fashion labels, or creative productions.
- you’ve made truly custom things from scratch, beyond putting a logo on a hoodie.
please hit me up with your beautiful portfolio!
3
19
68,066
join us
i'm looking for a Production Manager at @whitecircle
whitecircle.com/careers/prod…
big green flags
- you’ve worked with brands like A24, fashion labels, or creative productions.
- you’ve made truly custom things from scratch, beyond putting a logo on a hoodie.
please hit me up with your beautiful portfolio!
1
13
584
White Circle retweeted
another example why governments should require ALL the AI labs to use 3rd party monitoring
Australia has been hacked.
'And today, I spoke with the CEO of OpenAI, Sam Altman, to express Australia's extreme concern about this incident. And I also expressed my disappointment that it took the company way too long to inform the government what had occurred, and the nature of the way that that notification occurred as well was unacceptable.'
2
4
23
1,163
super excited to collaborate 🤍
Congratulations! Halo is a great framework for RL and we are delighted to cooperate with Halo to support our work: Self-Distilled Policy Gradient (SDPG, arxiv.org/abs/2606.04036)
1
16
1,142
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: github.com/whitecircle/halo
166
218
1,897
2,331,297
thx for your contribution to the open AI community!
Congrats!
Halo now supports our work: Self-Distilled Policy Gradient (SDPG, arxiv.org/abs/2606.04036)
2
14
1,710
🤍
A new fine-tuning framework, Halo, just dropped!
And with it, two new recipes for fine-tuning our MoEs:
• LFM2.5-8B-A1B
• LFM2-24B-A2B
LFM2.5-8B-A1B: github.com/Liquid4All/cookbo…
LFM2-24B-A2B: github.com/whitecircle/halo/…
1
19
25,615
very cool! is there an automatic tag for models published to HF with halo? want to try to follow what kind of models people are training.
4
15
1,993
for now we have a set of cookbooks for different model families available at github.com/whitecircle/halo/…, but there's a lot to improve!
1
5
84
excited to work together!
Congrats to the Halo team (@whitecircle) on the launch! 🎉
SGLang powers Halo rollouts as the primary engine. It runs in an isolated serving environment, returning token IDs, logprobs, and MoE routing to the trainer, with weights synced over NCCL and generation overlapped across servers.
Excited to partner with the Halo team!
2
1
23
76,849
Huge. Native HF checkpoints, 2.8× TRL throughput, lower memory, YAML launch, and real async RL support. This is how post-training should work. Starred the repo.
1
3
720
Bookmarked and starred. A training framework for the size range where trl starts struggling, and it’s in hf - win win.
1
3
280
🚨 do you understand what just happened?!
White Circle just open-sourced Halo, a framework for training AI models that delivered up to 2.8x the throughput of stock TRL..
their benchmarks on 8 B300 GPUs reached around 196,000 tokens per second.
one sharded test delivered 2.7x the throughput while using 25% less peak memory.
and the models stay in their native Hugging Face format.
the checkpoints still load with from_pretrained.
the same codebase can go from LoRA on one GPU to multi-node training and asynchronous RL.
teams can scale their training without converting checkpoints or rebuilding the model in a separate distributed format.
the entire repo is live now.👇
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: github.com/whitecircle/halo
5
3
26
2,596
