Config Options for Reproducing Table 3 (DiT-XL/2)

[applezoos]

Thank you so much for making the research code publicly available. I’m excited to test DuCa, but I’m a bit unsure about the names of the config options in the code that correspond to those mentioned in the paper.

• I’m currently trying to reproduce the results from Table 3 of the paper. Could you kindly let me know which options I should use in sample_ddp.py to perform S-A score and V-Caching?

• From my understanding of both the paper and the code, R=0.93 corresponds to "--fresh-ratio 0.07" and N=3 corresponds to "--fresh-threshold 3". Could you please confirm if this is correct?

If there are any additional settings I need to configure in the default config in order to reproduce Table 3, I would greatly appreciate it if you could let me know.

[Shenyi-Z]

When reproducing the results in Table 3, you should set --cache-type to 'kv-norm', and keep the other parameters as described in the paper. Your understanding is correct: R=93% corresponds to --fresh-ratio 0.07, and similarly, R=95% corresponds to 0.05. The value of N corresponds to --fresh-threshold.
I’ve been busy with work recently — I hope this hasn’t significantly affected your research progress.