RuntimeError: probability tensor contains either `inf`, `nan` or element < 0

[nullnameno]

Sorry, I encountered an issue while performing video inference, which is RuntimeError: probability tensor contains either 'inf', 'nan' or element<0. I am very curious about the reason for this error.

[MAGAer13]

Did you run this demo under bfloat16?

[MAGAer13]

Since the model is training under bfloat16, while the DepthwiseConv3D only support half, so we do the conversion from bf16 to half during Conv then we convert it back which leads to the unstable results.

[LinB203]

Sorry, I encountered an issue while performing video inference, which is RuntimeError: probability tensor contains either 'inf', 'nan' or element<0. I am very curious about the reason for this error.

Same question and just use demo code.

[nullnameno]

Did you run this demo under bfloat16?

Yes, I strictly followed the example you provided in README.md for video reasoning. I don't know why the model reported such an error.

[ff1Zzd]

Same error as well. Follow the demo code strictly

[ff1Zzd]

Hi I have also tried to use the model (with video weights) to inference an image input, however I am getting an output with all special tokens. I believe it might be caused by the same bug as the one raised by nullnameno. Here is the screenshot of the output.

[sdjhshbswp]

Hi I have also tried to use the model (with video weights) to inference an image input, however I am getting an output with all special tokens. I believe it might be caused by the same bug as the one raised by nullnameno. Here is the screenshot of the output.

I meet the same problem when I run the inference_video.py. I also get the res tensor of all zero.

[MAGAer13]

See #101