Input types are not broadcast compatible

[meetbryce]

I'm running into an inconsistent error when trying to run the default img2video workflow.

For almost all images it crashes with the error below, but with 2 images it worked as expected. The images are very similar and the nature of the error doesn't seem like it should have anything to do with the images.

I'll include the images and stack trace below.

I'm running on MacOS M1 Ultra using PYTORCH_ENABLE_MPS_FALLBACK=1 PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0 python3 main.py --use-pytorch-cross-attention

Workflow export is also attached for reference: ltxv-failing.json

No flash_attn import to remove
</s><s>The image shows a man in a suit sitting at a chessboard with a robot on the right side of the image. The robot is white and appears to be a robot with a futuristic design. The man is leaning over the chessboard and seems to be deep in thought, with his hand resting on his chin. The background is blurred, but it seems like the man is in a room with a large window and a view of a city skyline. The image is taken from a low angle, looking up at the robot and the chess pieces.</s>
Offloading model...
Requested to load LTXVModel
loaded completely 9.5367431640625e+25 3667.902587890625 True
  0%|                                                                                                                                                                                       | 0/25 [00:00<?, ?it/s]loc("mps_add"("(mpsFileLoc): /AppleInternal/Library/BuildRoots/4ff29661-3588-11ef-9513-e2437461156c/Library/Caches/com.apple.xbs/Sources/MetalPerformanceShadersGraph/mpsgraph/MetalPerformanceShadersGraph/Core/Files/MPSGraphUtilities.mm":233:0)): error: input types 'tensor<2x4900x2048xf32>' and 'tensor<2048xbf16>' are not broadcast compatible
LLVM ERROR: Failed to infer result type(s).
[1]    14104 abort      PYTORCH_ENABLE_MPS_FALLBACK=1 PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0 python3  
/Users/bryce/miniforge3/lib/python3.10/multiprocessing/resource_tracker.py:224: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown                           
  warnings.warn('resource_tracker: There appear to be %d '

Successful Image

Failed Image