Halo: Post-train LLMs 3x faster than TRL and Megatron

White Circle launched Halo, a post-training framework for open-source models that delivers up to 2.8x the throughput of stock TRL with lower peak memory while keeping models in native HuggingFace format. On OpenAI's gpt-oss-20b, Halo reached 2.3–2.8x TRL throughput under identical FlashAttention, Liger, fused cross-entropy and grouped-GEMM optimizations, and fine-tuning Z.ai's GLM-4.7-Flash on 177M tokens of agentic traces raised SWE-rebench-V2 from 33% to 42% at up to 1.63x TRL throughput. White Circle said every model it trains now uses Halo, which runs LoRA on a 24 GB GPU, multi-node training on B300s, and async RL, and it is partnering with the SGL project to make Halo the primary engine for rollouts.

White Circle on X: "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://t.co/3mAiUljdrN" / X White Circle on X: "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://t.co/3mAiUljdrN" 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/ha… 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/ha… Every model we train at White Circle now uses Halo. The same codebase runs LoRA on a 24 GB GPU, multi-node training on B300s, and async RL. Technical blog: Other training frameworks often require a separate implementation for each model family. In Halo, adding a new model family takes just 100 lines of connecting a wrapper instead of rewriting the model. Halo is not limited to supervised fine-tuning. It supports async RL and training with external environments. We are excited to partner with the @sgl project team to make it the primary engine for rollouts. On @OpenAI gpt-oss-20b, Halo delivered 2.3–2.8x the throughput of stock TRL with less peak memory. Both runs used the same FlashAttention, Liger, fused cross-entropy and grouped-GEMM optimizations. 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 throughput on the same model precision and data. Model and write-up:Show more Scaling an AI model shouldn't require rewriting half your codebase But that's exactly what happens when you outgrow your training framework @whitecircle just launched Halo: - Up to 2.8x faster than TRL -- Distributed training --- Native Hugging Face models ---- One stack fromShow more