[Single File] Add single file support for Wan T2V/I2V #10991
Conversation
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
| self.norm_added_q = RMSNorm(dim_head * heads, eps=eps) | ||
| # Wan applies qk norm across all heads | ||
| # Wan also doesn't apply a q norm | ||
| self.norm_added_q = None |
There was a problem hiding this comment.
@yiyixuxu Discovered an issue when running the slow test for the transformer. The diffusers implementation has this extra norm_added_q key which the original does not. When converting from the original checkpoint there is no weight to assign to this norm, so it remain a meta tensor, so we run into an error when setting the model to a device.
Removing this, and then adding norm_added_q to the _keys_to_ignore_on_load_unexpected in the transformer so that the warning about extra keys in the Diffusers version is suppressed.
Ideal solution is to update the weights in the model repo, but that could take time. This is a fix for the meantime
What does this PR do?
Fixes # (issue)
Before submitting
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.