Ignore docs/pics/ to avoid HF binary rejection

Add LoRA training assets: scripts, docs (no binaries), ui, my_dataset

Initialize 5Hz LM on Space startup (ACESTEP_INIT_LLM env, default true)

Use ACESTEP_CHECKPOINT_REPO from Space Settings (default ACE-Step/Ace-Step1.5)

Add Song Describer pipeline: prepare_song_describer.py, prepare_from_hf, preprocess/train CLI

Merge origin/main: keep Space worker resolution, return friendly result when not initialized

fix: return friendly result when service not initialized (no RuntimeError)

Upload 6 files

Create TRAINING_GUIDE.md

Create load_musiccaps_dataset.py

Update README.md

Zero GPU: worker-local handlers, init+generate in worker, avoid pickle/CUDA errors

Fix Zero GPU runtime: duration 60s, test fixes, LoKr PeftModel guard

fix: disable torch.compile on Spaces (ACESTEP_COMPILE_MODEL=0) to avoid pickle

fix: disable spaces.GPU to avoid pickle; run generation in main process on Zero

fix: reduce spaces.GPU duration to 60s for Zero free tier limit

fix: remove invalid --quantization None (unlimited tier defaults to None)

fix: re-enable spaces.GPU for Zero, disable quantization to avoid pickle

fix: disable spaces.GPU decorator to avoid model pickle (non-leaf tensor) on dedicated GPU Spaces

fix: use in-place init for nn.Parameter to avoid non-leaf tensors (HF Spaces pickle)

fix: replace Lambda with picklable Transpose12 for HF Spaces multiprocessing

fix: add --config_path acestep-v15-turbo for HF Spaces (empty checkpoints)

fix: use --server-name (hyphen) for argparse

fix: import spaces before CUDA to fix HF Spaces runtime error

fix: pin diffusers to commit with torchao logger fix

Update README.md

Add ACE-Step 1.5 Docker app

Update requirements.txt

Update README.md

Upload folder using huggingface_hub

Upload folder using huggingface_hub

initial commit