PKU-YuanGroup / Open-Sora-Plan

This project aim to reproduce Sora (Open AI T2V model), we wish the open source community contribute to this project.
MIT License
11.49k stars 1.02k forks source link

Inference speed questions about osp1.3. #493

Open Edwardmark opened 1 week ago

Edwardmark commented 1 week ago

I test osp 1.3 using the config below, but found that the inference speed is slower than 1.2 with same 93x720p 24fps. Do I miss something? CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 torchrun --nnodes=1 --nproc_per_node 8 --master_port 29514 \ -m opensora.sample.sample \ --model_path .../any93x640x640 \ --version v1_3 \ --num_frames 93 \ --height 720 \ --width 1280 \ --cache_dir "../cache_dir" \ --text_encoder_name_1 .../mt5-xxl/ \ --text_prompt "examples/sora.txt" \ --ae WFVAEModel_D8_4x8x8 \ --ae_path .../OpenSora-v1.3/vae \ --save_img_path "./train_1_3_nomotion_fps18" \ --fps 24 \ --guidance_scale 7.5 \ --num_sampling_steps 100 \ --max_sequence_length 512 \ --sample_method EulerAncestralDiscrete \ --seed 1234 \ --num_samples_per_prompt 1 \ --rescale_betas_zero_snr \ --prediction_type "v_prediction"

LinB203 commented 1 week ago

Looks like everything is fine, could you share your gpu as well as the generation speed? e.g, 2.7s/step.

Edwardmark commented 6 days ago

Looks like everything is fine, could you share your gpu as well as the generation speed? e.g, 2.7s/step.

I forget to add --sp in the command, now the speed is fater then 1.2, which is 5.2s/step, 100 steps cost about 8min 40s. In v1.2, it cost 12min. It is faster, but not faster 3x~4x as expected. @LinB203