Speed Question

[suimuc]

why 8H100's acceleration effect is very small, but single H100's acceleration effect is complied with the article I test with 512 * 512, 51 frames without pab（1H100）：157.61683344841003s with pab（1H100）：136.60238338s with pab（8H100）：106.87562332s

[oahzxl]

Hi, can you provide more details about how you run the code? And do your H100 have nvlink?

[suimuc]

I copy the pab_mgr.py comm.py parallel_mgr.py files and open_sora_transformer_3d.py, and then use the command :torchrun --nproc_per_node 8 scripts/inference_pab.py configs/opensora-v1-2/inference/sample.py

[suimuc]

my configuration parameter of pab is defalt

[suimuc]

I use DSP too

[suimuc]

and i wonder why your sample speed is so fast，192frames 480p noly 94s, is my H100's problem?

[ExtremeViscent]

my configuration parameter of pab is defalt

Hi, would you mind indicate which script you are using? VideoSys is currently using a builtin multi-processing launcher. Therefore elastic launch with torchrun is not recommended.