00:00
2026-09-22
rocm.blogs.amd.com
large-language-models
Zebra-HyLo: Upcycling Transformers into Long-Context Hybrid LLMs on AMD Instinctâ„¢ GPUs
AMD researchers introduced Zebra-HyLo, an upcycling recipe that converts pretrained Transformer models into long-context hybrid LLMs through post-training on roughly 10B tokens, without pretraining fr…