I wanted to share some feedback regarding stability during local execution on a Windows 11 system with an RTX 3090
While the initial video generation often completes successfully, subsequent runs frequently trigger GPU driver crashes (TDR / Black Screen), especially during the VAE decoding phase.
For comparison, running local WanGP workflows in alternative frontends like Maestro works completely fine and remains 100% stable across multiple consecutive generations on the exact same hardware. This indicates that the GPU and drivers are healthy, but there seems to be a memory management issue or VRAM leak in this specific setup where PyTorch/CUDA resources aren't fully cleared after a run.
A fix or improvement for VRAM garbage collection and VAE offloading between generations would make a huge difference in keeping the environment stable.
Thanks for looking into this!
I wanted to share some feedback regarding stability during local execution on a Windows 11 system with an RTX 3090
While the initial video generation often completes successfully, subsequent runs frequently trigger GPU driver crashes (TDR / Black Screen), especially during the VAE decoding phase.
For comparison, running local WanGP workflows in alternative frontends like Maestro works completely fine and remains 100% stable across multiple consecutive generations on the exact same hardware. This indicates that the GPU and drivers are healthy, but there seems to be a memory management issue or VRAM leak in this specific setup where PyTorch/CUDA resources aren't fully cleared after a run.
A fix or improvement for VRAM garbage collection and VAE offloading between generations would make a huge difference in keeping the environment stable.
Thanks for looking into this!