Results are different between multi-gpu inference and single gpu inference

Hi Author, @IceClear @numz
Thanks for the nice work.
I have tested the model with multi-gpu inference and single gpu inference, but I got the different result with same input video and same settings, and I really don't know why.
Here are some result and following cmd (tested with 1xA100 80G or 2xA100 80G):

**single-gpu inference cmd and results:**
CUDA_VISIBLE_DEVICES=3 torchrun --nproc-per-node=1 projects/inference_seedvr_7b.py --video_path PATH --output_dir PATH --seed 42 --res_h 720 --res_w 1280 --sp_size 1

https://github.com/user-attachments/assets/20e42e51-e81a-4106-a5f0-574a01c19aca

**multi-gpu inference cmd and results:**
CUDA_VISIBLE_DEVICES=2,3 torchrun --nproc-per-node=2 projects/inference_seedvr_7b.py --video_path PATH --output_dir PATH --seed 42 --res_h 720 --res_w 1280 --sp_size 2

https://github.com/user-attachments/assets/945ffd50-a263-47ff-9a7c-dbf832a39151

It seems the multi-gpu inference result is normal?
Looking forward to the reply, many thanks!