Eshika Rana

LLM training changes constantly. The problem is that every change can mean expensive rework. xLLM is built around a simple idea: keep the training pipeline flexible without sacrificing throughput. 6,295 tokens/sec/GPU on K2-Horizon-MoVA-36B-A4B and 10,050 on Llama 3 8B with H20

2.57% is a pretty telling number. PPTBench tested 18,000 reconstructions across 500 tasks and 36 configurations, checking whether agents could actually reconstruct scientific diagram slides correctly. Writing code is only part of visual coding. The ability to inspect and self-