Has anyone successfully replicated the effect of the Chronos-2 model?

[gwyanCN]

We are currently attempting to reproduce the experimental results of the Chronos-2 model based on the methods described in its original paper. Due to the poor correlation of real-world data, we initially adopted a univariate training approach for the real data. Through experimental analysis, we found that our group attention module failed to undergo effective parameter updates.
Subsequently, to address the training issue of the group attention module, we tried the synthetic data methods (TCM and TSI) mentioned in the paper. We split the training process into two stages, both of which adopted a combination of real-world and synthetic data for multivariate and univariate training, respectively.
However, from the experimental results, the group attention module still could not be trained properly. By analyzing the training logs, we observed that during the gradient update process, the update gradient of group attention was several orders of magnitude smaller than that of time attention。
We would like to ask if any researchers in the community have encountered similar problems during the reproduction of Chronos-2 or related models? Any insights, experiences, or suggestions would be greatly appreciated. Thank you very much!