Added explicit memory management during the VAE decode process. < 8000 span class="f1-light color-fg-muted">#7587
Conversation
|
I can't merge this as is because it would break on non cuda devices but does it actually improve things? h = h_new does "del" the h variable. del doesn't actually delete anything, it only removes that reference to the object. |
|
@comfyanonymous Hi, thank you for your review. |
|
You are correct that non-CUDA devices cannot use torch.cuda.empty_cache() to release VRAM. However, this step is quite important for CUDA devices. I can add logic here to check if CUDA is available and apply this optimization specifically for CUDA devices. |
|
I can also show you the difference in VRAM usage during the VAE decoding stage, comparing the results with and without explicit memory management. |
If empty_cache is needed, comfy.model_management.soft_empty_cache() might be better for compatibility. |
…ache() with model_management.soft_empty_cache().
|
@chaObserv Thank you for your valuable suggestion! I have already replaced torch.cuda.empty_cache() with model_management.soft_empty_cache() to support non-CUDA devices. |
Hello, I noticed that during the VAE decode process, the feature maps as intermediate variables were not explicitly deleted, which led to unnecessary CUDA out of memory errors. I have fixed this by expli 8000 citly deleting the intermediate variables.