https://github.com/woct0rdho/transformers5-qwen3.5-recipe An update to my LoRA over GGUF series: Now we can train Qwen3.8-Flash-Next (125B-A6B + 51B engram) in 40 GiB VRAM, with no CPU offloading, with engram on disk that does not reduce training speed. On Strix Halo it trains context chunk size 204
