v5.1: revert INT8 — too lossy for 0.6B, keep LoRA merge + float32

v5: merge LoRA at startup + INT8 quantization + faster params

Skip quantization for PEFT models — fixes generate crash

Fix: quantization fallback if generate breaks

Revert to Docker/FastAPI + add dynamic INT8 quantization for faster CPU inference

Switch to Gradio SDK + ZeroGPU for free GPU inference

Deploy North Air CPU API + model

Deploy North Air CPU API + model

Deploy North Air CPU API + model

Deploy North Air CPU API + model

initial commit