-
Notifications
You must be signed in to change notification settings - Fork 72
Open
Description
I just built the master branch on NVIDIA Jetson Thor (Blackwell, CUDA 13.0, CUDA Arch 11.0), but I get the following issue with the version set specified below:
torch=2.9.0a0+50eac811a6.nv25.9
torchvision=0.24.0a0+98f8b375
torchcodec=0.9.0a0
Test code:
rom torchcodec.decoders import set_cuda_backend, VideoDecoder
video_file = "video.mp4"
with set_cuda_backend("beta"): # Use the BETA backend, it's faster!
decoder = VideoDecoder(video_file, device="cuda")
Error:
Traceback (most recent call last):
File "/root/test_codec.py", line 6, in <module>
decoder = VideoDecoder(video_file, device="cuda")
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/torchcodec/decoders/_video_decoder.py", line 154, in __init__
core.add_video_stream(
File "/usr/local/lib/python3.12/dist-packages/torch/_ops.py", line 840, in __call__
return self._op(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: Unsupported device: cuda (device type: cuda, variant: beta)
As it is already specified with RTX 5090 here #1006 (comment), I guess it is related to Blackwell architecture. Is there any workaround to this?
Metadata
Metadata
Assignees
Labels
No labels