Discrepancy in GSM8K evaluation results compared to Table 1

[JueVioletGrace]

Hi,

I git cloned your project and ran the ../Fast-dLLM/llada/eval_gsm8k.sh script, but I couldn't reproduce the results reported in the paper (Table 1).

Specifically, the accuracy I obtained is about 2% lower than the reported data.

I have double-checked the script multiple times, and the default parameters appear to align exactly with the experiment in Table 1. Could this discrepancy be related to hardware differences (e.g., GPU model), or is there any other specific setting I might have missed?