Remove Nemotron from Space backend due to unresolvable Mamba dependencies

Fallback to slow Mamba path for Nemotron

Support Nemotron and Gemma base models in ZeroGPU backend

Use merged model directly — no PEFT/LoRA needed at runtime

Fix generate crash: extract input_ids tensor from BatchEncoding

CPU inference: remove @spaces.GPU, fail-safe LoRA loading

Split model loading: base at startup, LoRA inside @spaces.GPU

Dual backend: transformers+ZeroGPU on Spaces, llama.cpp locally

Update Readme and GPU usage

Update Readme and GPU usage

Update Readme and GPU usage

Updates for GPU usage

Initial deploy to Hugging Face Space