perf: drop bnb 4-bit and torch.compile for faster ZeroGPU inference

fix: move demo.load inside Blocks context

add page-load warmup to pre-load models and burn torch.compile on first visit

add torch.compile with reduce-overhead mode

add streaming + sdpa attention for faster generation UX

add bitsandbytes to runtime install so 4-bit quantization actually applies

fix inference: use return_dict=True and unpack inputs for generate

bypass adapter_config.json by downloading weights-only snapshot

load fine-tune directly as merged model, drop peft dependency

add 4-bit quantization to bring model size back to ~3GB each

fix dtype deprecation warning

programmatic install as workaround for requirements.txt not being picked up

fix requirements: remove spaces (pre-installed), pin transformers>=4.51 for Gemma 4

force rebuild

First Commit

initial commit