Post
4212
π Introducing Halo 1.0
Today, we are open-sourcing Halo, the training framework we use to train every model at White Circle.
It comes with:
π§ Full post-training stack: SFT, DPO/KTO/SMPO, reward modeling, GRPO, distillation
π€ Async multi-turn RL with vLLM/SGLang rollouts and sandboxed tool use
β‘ ~2.8Γ TRL throughput on 8Γ B300 (EP+FSDPv2, FA4, fp8/fp4)
π€ Dense HF models + 15 MoE families (Qwen, GLM, Mistral, DeepSeek-V4β¦)
π οΈ One halo command, prebuilt Docker images, and docs for humans and agents
π» https://github.com/whitecircle/halo
Try it and tell us what you're training
Today, we are open-sourcing Halo, the training framework we use to train every model at White Circle.
It comes with:
π§ Full post-training stack: SFT, DPO/KTO/SMPO, reward modeling, GRPO, distillation
π€ Async multi-turn RL with vLLM/SGLang rollouts and sandboxed tool use
β‘ ~2.8Γ TRL throughput on 8Γ B300 (EP+FSDPv2, FA4, fp8/fp4)
π€ Dense HF models + 15 MoE families (Qwen, GLM, Mistral, DeepSeek-V4β¦)
π οΈ One halo command, prebuilt Docker images, and docs for humans and agents
π» https://github.com/whitecircle/halo
Try it and tell us what you're training
- 1 reply
