xLLM promises flexible LLM training without giving up throughput

IFM introduced xLLM for pre-training and fine-tuning dense and mixture-of-experts LLMs. It reports 6,295 tokens per second per GPU on K2-Horizon-MoVA-36B-A4B and 10,050 on Llama3-8B, both on H200s. IFM says xLLM also ships with K2 Horizon checkpoints, training logs and recipes.