Spaces:
Running
Running
Commit History
-
fix understand_audio: clone tensors for inference mode, working on GPU (52s)
4b2f4ad -
add understand_audio (LM reverse), demucs-infer fix, commit refs, dtype fixes
6bfdc38 -
major update: PyTorch inference, Gradio 6, session isolation, /understand captioning
- generate_audio(): full PyTorch inference pipeline (no ace-server needed) - tiled_vae_decode(): memory-bounded VAE decoding - Gradio 6 migration (gr.update -> component constructors) - Session isolation (random suffix on LoRA names) - /understand captioning before training (falls back to librosa) - Active tab detection pattern from rvc-beatrice - Gradio API fix (tempfile for adapter download) - dtype fix for mixed precision inferenceff239f5
-
SDPA first on Blackwell, FA2 only for Ampere/Hopper, txt caption support
04ccf32 -
add GPU/CUDA auto-detect, mixed precision, flash_attn, txt caption parser
917e4ed -
save only on cancel/finish, max epochs 1000, default 3
d42aa91 -
add mid/sas analysis modes (Demucs + ensemble), auto-select by dataset size
b38d0b1 -
add auto-captioning (BPM/key/signature via librosa), add librosa+mutagen deps
1d42836 -
add XL variant mappings to _VARIANT_DIR
3c5ce9c -
fix: adapter saved to clean dir, LM dropdown no 'Default', on-demand download
e62602f -
Side-Step training engine, tested locally on CPU
- train_engine.py: 1347-line standalone LoRA training (ported from Side-Step) - Adafactor optimizer, 2-pass sequential preprocessing, FA->SDPA->eager fallback - Generator-based progress (yield per epoch), cancel button, checkpoint saves - Pin transformers<4.58.0 (meta tensor fix), skip device_map on CPU - Tested locally: preprocessing 174s + training 51s (2 epochs, 10s audio)a07b39d