Fix HunyuanVideo produces NaN on PyTorch<2.5 #10482
Merged
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
NaN
tracked todiffusers/src/diffusers/models/transformers/transformer_hunyuan_video.py
Lines 117 to 135 in 811560b
Specifically, some elements of
encoder_hidden_states
.The dimensions of query, key, value and mask are large which suggests versions <2.5 used 32-bit indexing, this tracks with #10314 if ROCm versions are still using 32-bit indexing, this may also close that issue, awaiting confirmation from user.
Tested on CUDA 2.4.1
output.mp4
Code
There's also a small performance increase
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.
@sayakpaul @a-r-r-o-w