Questions about WanVideo Torch Compile Settings and WanVideo Block Swap nodes for optimization on 4090D

[Passenger12138]

Hello,

I am using your Wan2.2 workflow for image-to-video tasks, and I can run the workflow successfully. However, I’d like to better optimize it for my 4090D GPU with some specific acceleration settings.

Inside ComfyUI, I encountered two nodes that I don’t fully understand, and I couldn’t find much documentation about them.

1. WanVideo Torch Compile Settings

I would like to know what the following parameters mean, and in which situations I should change them for faster performance (or if they are not necessary to modify):

backend
fullgraph
mode
dynamic
dynamo cache size limit
compile transformer blocks only
dynamo recompile limit

2. WanVideo Block Swap

The parameters I am not sure about are:

blocks to swap
offload img emb
offload txt emb
use non blocking
vace blocks to swap

I tried changing blocks to swap, but it didn’t seem to affect speed much. Perhaps due to your good implementation of parallelization and model rotation.

My Question：

I want to further speed up the workflow. Could you provide guidance on:

What these parameters actually control

When it makes sense to change them

Any recommended settings for a 4090D GPU（maybe umt5 and vae load and offload casue time waste?）

Thank you for your great work on this project!

[kijai]

In most cases don't touch the torch.compile settings, defaults should be most compatible.

For block swap it's situational, you can balance how much of the model weights are on the GPU VRAM at once, higher the value less VRAM the model takes, but downside is that the inference speed decreases.

Non blocking reduces the speed loss at the cost of more RAM reserved.

Offloading the embed layers slightly reduces VRAM usage as well, but it's very minimal.

[Passenger12138]

Thanks for your reply! I have another question.

Currently, WAN2.2 video generation uses two models (a high-noise model and a low-noise model) for inference. I’d like to confirm whether there is a model loading latency between them. For example, after the first model finishes inference, does the implementation wait to load the second model onto the GPU? Is this handled sequentially or asynchronously in your current implementation?

If I want to modify the workflow to support asynchronous model loading/inference, which part of the code should I focus on?