Drop gpu_lock: Gradio's default concurrency_limit=1 already serializes process_fn

replace torchaudio.load() with sf.read()

no flex_attention

sidestepping dynamo

move extractor to GPU

kwargs fix for HF version

added torchvision

Defer GPU move to inside process_fn (ZeroGPU can't touch CUDA from a background thread)

Fall back to sdpa when flash-attn isn't installed

Fix ZeroGPU import order: spaces must be imported before torch

initial deploy

initial commit