search

jy0205 / Pyramid-Flow

Code of Pyramidal Flow Matching for Efficient Video Generative Modeling
https://pyramid-flow.github.io/
MIT License
721 stars 56 forks source link

OOM Issue #31

Open GSK666 opened 2 hours ago

GSK666 commented 2 hours ago

What a great job. When i use A800 80G with default parameters to infer the 768P video, I find that the GPU memory increases first and then decreases. The step inference is fine, but an OOM error is reported during the final VAE decode, even if I reduce the temp parameter and set save_memory parameter is True. 0zm6n12000g3a47sl9139 0zm0512000g39w4430EBC Looking forward to your reply, Thank you.

jy0205 commented 2 hours ago

Thanks for your interests. Have you added the model.vae.enable_tiling()? If not, the save_memory=True does not work.

GSK666 commented 1 hour ago

Thank you so much. model.vae.enable_tiling() This configuration solved my problem. Does this 768P model support i2v generation with dynamic resolution?

feifeiobama commented 40 minutes ago

Glad to hear you have resolved this issue. The 768p model does not support dynamic resolution because we did not use a variable-resolution training strategy. We trained the model first on 384p video and then on 768p video.