原生支持HuggingFace的AI模型训练框架：Halo，吞吐量是标准TRL的2.3–2.8倍 单卡最高约24456 token/s，8卡合计19.6万 token/s，等于直接降低了定制模型的成本 如果门槛降下来，或可会促进开源模型生态，让更多人开始转向微调自己的专用模型 另外，它AI原生设计，把文档分成了human-docs和agent-docs，在镜像里内置了Claude…

THERE'S A NEW BEST FRAMEWORK FOR POST TRAINING OF OPEN SOURCE MODELS > 2.8x TRL throughput > less memory > still standard @huggingface checkpoints that’s a strong pitch 🤗

2.8x over stock trl, and to their credit the like-for-like bar is right there in the chart stock trl zero-3: 3,885 tok/s/gpu halo ep1 zero-3: 5,560 halo ep2 zero-2: 10,479 ~1.4x at matched sharding,…

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…

came across Halo and this caught my attention: take the Hugging Face model you already have and scale training without rewriting it for another framework distributed training without giving up the HF…

fine-tuning your own AI just got stupid easy up to 2.8x faster than the current standard ~100 lines to add a new model stays plain hugging face halo by @whitecircle and it's open source

here’s the training framework that keeps your models in hugging face and comes with skills for your ai agents do check it out on Github, for all the nerds who are interested in training hello halo!

I've been tiptoeing into training with Unsloth on my DGX Spark. The moment you outgrow that, the next stop is a whole different framework and your model stops being "just a Hugging Face model". Halo…

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…