IFM Lab Open-Sources xLLM Training Framework with Hot-Swappable Tokenizer and Data Mix

IFM Lab has open-sourced xLLM, a training framework for dense and MoE LLMs that allows tokenizer, chat templates, data mix, architecture, and training stages to be changed on the fly without rebuilding datasets, while sustaining 10050 tokens/s on H200 GPUs.

2026-09-29 ~ 2026-09-29 · 2 related posts

• xLLM open-sourced: flexible pre-training infra hits 10,050 tokens/s/GPU on H200 without dataset rebuilds — HongyiWang10 · 2026-09-29
• IFM Lab releases xLLM training library: hot-swappable tokenizers and 1k-line Jinja chat templates — HildeKuehne · 2026-09-29