Kero Liang

感谢建议，这边在着手重构这个主题，pjax等功能会考虑提供支持。 CDN 加速节点，全部挂掉的可能性还是比较低的，如果遇到某个 CDN 不可用的情况，可以自行在主题设置先切换。不过，的确只依赖 CDN 的话，在本地开发时可能会有一些不便。如果目前有需要的话，可以把依赖的文件传到主题目录，然后暂时更改一下 CDN 的指向。

收到，感谢建议

感谢建议，这边在着手重构这个主题，pjax、第三方评论等功能会考虑提供支持。

> Nice work! I wonder whether it would've been easier to land this contribution in two separate PRs, one with the CUDA implementation and a follow-up with the numba one...

> @imkero @wulipc do you have any LoRA-tuned models that can be used? cc @ywang96 Sorry I don't have one currently

> @imkero Thanks for the PR! This is amazing 🚀 > Could you please resolve the merge conflict and the lint error? Sure! I will update this PR soon. TODO:...

> @imkero Just so you know: To fix the CI failure, we should move `numba` from `requirements/cuda.txt` and `requirements/rocm.txt` to `requirements/common.txt`. Thanks for your remind, I have moved it to...

@WoosukKwon I think this PR is ready for review now

I will take get_next_positions_tensor into consideration because they are reported to be time-costing as well in #17617

I have written an optimized version of `get_next_input_positions_tensor` which is zero copy and no lru cache. Will add e2e benchmarking result like #17617