The execution of ComfyUI on dgx spark is very inefficient. The GPU’s efficiency is consistently low. When using WAN 2.2 to generate videos, the execution time takes 15 to 25 minutes, and sometimes it can even cause dgx spark to crash.
Error message:
torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 269.47 GiB. GPU 0 has a total capacity of 121.69 GiB of which 87.52 GiB is free. Including non-PyTorch memory, this process has 29.90 GiB memory in use. Of the allocated memory 29.64 GiB is allocated by PyTorch, and 45.66 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management ( CUDA semantics — PyTorch 2.11 documentation )
How to fully enhance GPU performance on DPX Spark？