HeartMuLa Studio by @ambsd is a great UI for using the HeartMuLa song generation AI model
Apparently the minimum VRAM requirement is 8GB, and the app automatically uses as much VRAM as the system has in order to generate the song as quickly as possible.
Here's an example:
I asked him more about how the VRAM optimization works, here's what he said:
On your 20GB A4500, it detects plenty of VRAM so it uses "no swapping" mode:
- Loads BOTH models simultaneously (HeartMuLa 4-bit ~3GB + HeartCodec ~6GB + KV cache ~5GB = ~14GB)
- Faster because no model swapping needed
- Uses 14.47GB - this is expected!
On an 8GB card, it would automatically switch to "sequential offload" mode:
- Loads ONLY HeartMuLa during generation (~3GB + ~4GB KV cache = ~7GB)
- After generation, UNLOADS HeartMuLa from VRAM
- Then loads HeartCodec for decoding (~6GB)
- Peak VRAM: ~7GB (fits in 8GB!)
- Trade-off: ~70s extra time for model swapping
TL;DR:
- ✅ YES it will work on 8GB - uses different mode automatically
- Your 20GB card = fast mode (both models loaded)
- 8GB card = swap mode (one model at a time)
The system adapts based on available VRAM:
- 20GB+ → Full precision, no swap
- 14-20GB → 4-bit, no swap (your card)
- 10-14GB → 4-bit + swap
- 8-10GB → 4-bit + swap (with warning)
Your 14.47GB usage is correct for a 20GB card. On 8GB it would only use ~7GB peak!
HeartMuLa Studio by @ambsd is a great UI for using the HeartMuLa song generation AI model
Apparently the minimum VRAM requirement is 8GB, and the app automatically uses as much VRAM as the system has in order to generate the song as quickly as possible.
Here's an example:
I asked him more about how the VRAM optimization works, here's what he said: