Benchmark time on H100
yossigandelsman opened this issue · 2 comments
Hi,
I loved your work!
I am trying to repeat the experiments you had on 8xH100s.
When I run your codebase with ./scripts/opensora/sample_pab.sh (after following the installation instructions), and change the number of frames in configs/opensora/sample_pab.yaml to 192, it takes around 9 seconds to sample a video.
In your blog, you show that it takes around 4 seconds, so I am not sure what is the source for the difference in performance.
Will you be able to provide the exact parameters and script to reach the same performance as you reported?
Thank you very much,
Yossi
we use torch2.3 and cuda12.1, and skip the first cold start iter.
oh sorry for the mistake. the caption for the video in the blog is wrong. it should be 96 frames. you can refer to speedup section for correct speed.