lora merge as it was really tricky to get it to work of https://huggingface.co/152334H/miqu-1-70b-hermes2.5-qlora.
Base Model: Miqu 70B (Mistral AI Leak) Dequantized by 152234h Finetune also by 152234h
Outputs seem good, but the prompting is still a bit buggy, not sure if that's an error on my part.
For me it wouldn't generate text until I activated flash attention 2 in Oogabooga. You need around 130 GB vram, 2 a100 80 or h100 work, as does 6 3090 or 4090.