search

rhymes-ai / Allegro

Allegro is a powerful text-to-video model that generates high-quality videos up to 6 seconds at 15 FPS and 720p resolution from simple text input.
https://rhymes.ai/
Apache License 2.0
291 stars 14 forks source link

Error while running: Must be CUDA tensor #15

Closed ajikmr closed 23 hours ago

ajikmr commented 23 hours ago

Hello.

Thanks a lot for sharing code and model.

I am getting error while running the sample case.

File "/**/Allegro/allegro/pipelines/pipeline_allegro.py", line 702, in call ) = self.encode_prompt( File "/***/Allegro/allegro/pipelines/pipeline_allegro.py", line 240, in encode_prompt prompt_embeds = self.text_encoder(text_input_ids, attention_mask=prompt_attention_mask)

(****

File "/****/python3.10/site-packages/torch/autograd/function.py", line 574, in apply return super().apply(args, kwargs) # type: ignore[misc] File "/*****/python3.10/site-packages/apex/normalization/fused_layer_norm.py", line 75, in forward output, invvar = fused_layer_norm_cuda.rms_forward_affine( RuntimeError: input must be a CUDA tensor

I checked both input in are being passed as cuda tensor in line 240 of pipeline_allegro.py

OS: Ubuntu 22.04 Python 3.10.9

ajikmr commented 23 hours ago

Error is not cropping up with a fresh conda env. I guess some previously installed module was causing error.