Gamahea/ACE-Step-Custom

Force traditional loading: low_cpu_mem_usage=False + explicit device move

Use official ACE-Step demo approach: dtype as string (bfloat16)

Use device_map to force CUDA placement and prevent meta device

Fix: Force CPU device context during model init to prevent meta tensor operations

Fix: Remove device_map to prevent meta tensor initialization - ACE-Step custom model needs direct device placement

Fix: Implement lazy model loading for ZeroGPU compatibility

Fix: Add device_map to prevent meta tensor errors on ZeroGPU

Fix: Add device_map to prevent meta tensor errors on ZeroGPU

Try _fast_init=False to prevent meta device initialization

Fix dtype parameter: use dtype instead of torch_dtype, remove device_map

Fix meta tensor: add device_map=None and use torch_dtype parameter

Fix meta tensor error: disable low_cpu_mem_usage for model loading

Add vector-quantize-pytorch dependency for ACE-Step model

Fix DiT initialization: pass model name only, not full path

Fix TypeError: pass Path object to ensure_main_model, not string

Fix syntax error: close unterminated docstring in _load_models

Use official ACE-Step model downloader for automatic model downloads

Monkey patch get_api_info to prevent JSON schema errors

Move endpoint override before launch

Fix fastapi_app attribute name and move endpoint override after launch

Add /info endpoint override and pin Gradio version

Add loguru dependency and fix Gradio 4.0.0 compatibility

Fix engine initialization and audio I/O components

Deploy ACE-Step Custom Edition with bug fixes