WAN Video And Phantom 14B in ComfyUI Issues

I have recently been working with WAN 2.2 for AI video generation in comfyui.

One of the workflows that I experimented with was the Phantom Wan 14B Workflow that is included in comfyui as a default template.

However, I had two issues with this workflow.

Firstly, when the “base_precision” is set to the default of “fp16_fast” I get the following error:

WanVideoModelLoader

torch.backends.cuda.matmul.allow_fp16_accumulation is not available in this version of torch, requires torch 2.7.0.dev2025 02 26 nightly minimum currently

This can be easily fixed by setting the base_precision to “fp16”

However, I then ran into another issue, which took several hours to solve. This issue was coming from the WanVideo Sampler:

WanVideoSampler

must be called with a dataclass type or instance

After updating comfyui UI, Cuda, and verifying that my pytorch versions were compatible, I was eventually (With the help of chatGPT) able to figure out that the issue was with the “WanVideo Torch Compile Settings” node included in the default workflow.

When I bypassed this, the Phantom 14B Workflow worked fine.

I am making this post in the hopes that this might help save someone else some time, I couldn’t find any users with a similar issue online.