Tomorrowdawn
Tomorrowdawn
> [出错页面](https://yuudawnlight.com/mathematics-docs-1) > 在本地的publish markdown上正常渲染. > [这是本地截图](https://i.853tv.cn/imgs/2021/02/4640bf65f9a6c8b9.png) > 非常感谢您的软件,但是能否请您修复/指明该错误? 现在是使用了先渲染后上传的方式. 若不渲染而改用Simple MathJax插件进行渲染(已设置Custom mathjax config),则无法进行任何渲染.
Hi now flashinfer supports custom mask, which is a great work! But how about the positional embedding? I found #69 introducing q_position and kv_position in C++ kernels, but I didn't...
2024-10-01 仍然遇到此问题 datasets下降版本至2.14.6可暂时解决此问题, 但是出现了trust_remote_code关键字不兼容(似乎新版本加入) 随着datasets的不断更新, 可能分歧会越来越大, 希望能尽快解决, 也希望以上信息能有所帮助.
Supplementary: It works with DynamicCache.  So it must be something wrong with SinkCache and relevant control code.
Supplementary: Well that's a bug hiding in rotary embedding(which is not here so I didn't find it...). Re-applying rotary embedding at each step is neccessary so I modify the whole...
@yubofredwang I have the same issue. Could you please point it out? Current implementation doesn't support FP8(like finegrained_fp8) and simple type casting is not enough because the scale of activation...
Sorry, I don't understand what you mean. Storing the mapping directly is equivalent to remove the `i for i in range(len(used_tokens))` term. So, it uses same memory. However, storing the...