White Circle 开源分布式后训练框架 Halo，面向已超出 HuggingFace TRL 能力、但无需完整 Megatron 栈的模型。在 8× B300 上，Halo 训练吞吐量最高达原生 TRL 的约 2.8 倍，双方均启用 ZeRO-3 分片时峰值内存还低 25%。模型保持原生 HuggingFace 格式，可直接跑多轮 RL，从单 GPU 扩展到集群。
White Circle just unveiled Halo, an open-source, distributed post-training framework for models that have outgrown Huggingface TRL but don't need the full Megatron stack.
On 8× B300, Halo delivers up to ~2.8× the training throughput of stock Huggingface Transformers Reinforcement Learning (2.7× at 25% less peak memory when both sides shard ZeRO-3), with larger margins over the other frameworks benchmarked.
you can run multi-turn RL directly on Hugging Face models, from one GPU to a cluster.
Rather than reimplementing model architectures, Halo wraps existing Transformers models with expert, context, tensor and expert-tensor parallelism.
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: https://github.com/whitecircle/halo在 X 查看被引用的帖子
来源：Rohan Paul · x.com