Fix chord parser: preserve <unk> separator from T5 decode

Disable TF32 + cudnn nondeterminism (torch 2.8 default TF32 shifts T5 output)

Unpin pandas (1.4.1 lacks the kwarg gradio 6 queueing analytics expects)

Defensively skip malformed chord items (torch 2.8 shifts T5 sampling)

Restore original audio-stack pins (only modernize torch family); revert defensive split

Use rsplit(' at ',1) for chord parse (model now sometimes emits multiple 'at')

Pin librosa<0.10 (pad_center 'size' was made keyword-only in 0.10)

Pin librosa==0.10.2 (audioldm uses pre-keyword pad_center signature)

Import spaces first (CUDA must not be initialized before)

ZeroGPU: torch 2.8, unpin most deps (keep transformers/protobuf/diffusers pins)

upgradio 6.14.0

Update app.py

Update app.py

upgradio 6.13.0

huggingface_hu >=0.33.5

upgradio 6.6.0

Update app.py

use community sd2-1

Update README.md

Update requirements.txt

Update README.md

Update README.md

python 3.10

Update README.md

Update requirements.txt

Update requirements.txt

upgradio 6.5.1

upgradio 5.45.0

Update app.py

disable caching examples

upgradio 5.44.1

upgradio 5.44.0

fix typo

Upgrade hf-xet

upgradio 5.43.1

upgradio 5.39.0

upgradio 5.38.2

upgradio 5.38.0

upgradio 5.34.2

Update requirements.txt

Update requirements.txt

upgrade pytorch 2.6.0

Update requirements.txt

Update requirements.txt

Update requirements.txt

Update requirements.txt

upgrade hf packages

upgrade pytorch

upgradio to 5.34.0

Update README.md