How to fine-tune Wan 2.1 with Context Parallelism?

[vviper25]

I am trying to fine-tune the Wan 2.1 model and would like to leverage the Context Parallelism (CP) feature to manage memory and scale the training. I saw in the main README that CP support is listed as a key feature.

I have looked through the examples/training directory and the documentation, but I couldn't find a specific example or launch script demonstrating how to fine-tune the Wan model with Context Parallelism enabled.

Could you please provide some guidance or a minimal example on how to properly configure a training job for Wan 2.1 with Context Parallelism?

[a-r-r-o-w]

When launching the script, you can specify --cp_degree 2 if you want to use CP across 2 GPUs

[yihao-meng]

When launching the script, you can specify --cp_degree 2 if you want to use CP across 2 GPUs

@a-r-r-o-w Hi, thank you for your amazing work! If we want to use context parallel now, should we use the main branch or the default released v0.2.0 branch in the readme?

[a-r-r-o-w]

CP changes only exist in the main branch and hasn't been part of a release yet (I haven't had the time to continue working on the project). So please use the main branch for now and feel free to let me know if you face any problems :)