whitecircle/halo
Halo is a White Circle framework for training Hugging Face LLMs and multimodal models from one GPU to multi-node runs.
The repo says Halo supports pre-training, SFT, preference optimization, distillation, and asynchronous multi-turn RL on the same distributed stack. It keeps checkpoints in standard Hugging Face/SafeTensors form, with `from_pretrained` still working. White Circle claims up to about 2.8x stock TRL throughput on 8x B300 in its published benchmarks, with lower peak memory in the ZeRO-3 comparison. The first public release is listed as Halo 1.0.0, dated 2026-08-20. GitHub · LLM repos' note
The repo says Halo supports pre-training, SFT, preference optimization, distillation, and asynchronous multi-turn RL on the same distributed stack. It keeps checkpoints in standard Hugging Face/SafeTensors form, with `from_pretrained` still working. White Circle claims up to about 2.8x stock TRL throughput on 8x B300 in its published benchmarks, with lower peak memory in the ZeRO-3 comparison. The first public release is listed as Halo 1.0.0, dated 2026-08-20. GitHub · LLM repos' note
score 5