[hybrid inference 🍯🐝] Wan 2.1 decode #11031
Conversation
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
|
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. |
| 512, | ||
| ) | ||
| endpoint = DECODE_ENDPOINT_WAN_2_1 | ||
| dtype = torch.float16 |
There was a problem hiding this comment.
cc @yiyixuxu Currently the endpoint is running in float16 and output seems ok (on random latent), the examples use float32 but we noticed in the original code that everything is under bfloat16 autocast context. Can we check with the authors regarding the use of float32 for VAE?
There was a problem hiding this comment.
i haven't heard back from them,
maybe let's generate some sample comparison and ask @asomoza to help decide if there is a quality difference
| @@ -176,18 +177,6 @@ def test_output_type_pt_partial_postprocess_return_type_pt(self): | |||
| f"{output_slice}", | |||
| ) | |||
|
|
|||
| def test_do_scaling_deprecation(self): | |||
There was a problem hiding this comment.
Wan VAE doesn't have a scaling_factor that we could pass, so this deprecation/removing do_scaling doesn't work, we will keep it.
|
This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread. Please note that issues that do not follow the contributing guidelines are likely to be ignored. |
What does this PR do?
Rebase after #11017
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.