question about DCVC_RT's BDrate against VTM-17.0

Hello, i use your test suggest to calculate bdrate for DCVC_RT. But my result is far from yours, can you help me?
There are my encode commands:
DCVC_RT:

 python test_video.py 
 --model_path_i cvpr2025_image.pth.tar 
 --model_path_p cvpr2025_video.pth.tar 
--test_config dataset_config_example_yuv420.json 
--rate_num 10 --cuda 1 -w 4 --write_stream 1 --force_zero_thres 0.12 
--output_path dcvcrt_output.json 
--stream_path dcvcrt_out_bin
--force_intra_period -1 --reset_interval 64 --force_frame_num 96 
--check_existing 0 --verbose 2 verbose_json True --calc_ssim True 
--save_decoded_frame True --cuda_idx 4 5 6 7

VTM-17.0 and HM-16.25:

EncoderApp -c encoder_configuration.cfg 
-f 96 -q {qp} --IntraPeriod=-1 --InputFile={src_yuv} 
--SourceWidth={width} --SourceHeight={height} --FrameRate={frame_rate}
--Level=6.2 --InputBitDepth=8 --DecodingRefreshType=2 
-b {output.bin} -o {enc.yuv}

Then I calculate bdrate (use "ave_all_frame_bpp" and "ave_all_frame_psnr") for every seqs, then calculate their average, but the results are far from your paper's.

I also plot the RD curves for every dataset, compare my RDcurve with your paper's RDcurve figures. But they don't fit perfectly(HEVC_C is far from your curve). For example:

I notice that your paper's RDcurve figures are all for a whole dataset, not for a single seq. So I also calculate BDrate for per dataset, results follow as:

which are closer to your paper results, do you calculate bdrate in this way?
Could your answer my questions? Thanks for reading!

Start with test_video.py and dataset_config_example_yuv420.json, then compare the DCVC_RT command's frame and rate settings with the VTM-17.0/HM-16.25 EncoderApp invocation. Reproduce the per-sequence and per-dataset BDrate calculations using ave_all_frame_bpp and ave_all_frame_psnr, and compare them with the paper's curves; done means explaining the discrepancy or documenting matching settings.