Yueming Hao

optimization for zero_grad

8

### Description The original `zero_grad` calls `grad.zero_()` for every eligible tensor. There will be a lot of tiny cuda kernels in this way. I add a temporary list to save...

搜狗输入法linux 2.3.2.07 无法在qq微信中使用

3

ubuntu 20.04 使用install_2.8.22.sh 安装当前repo下的包，qq和微信是readme中推荐的版本。 使用rime输入法可以正常在qq和微信中输入中文。使用新版本的搜狗输入法则无法输入中文。 没有找到给搜狗输入法反馈的地方。

在页面边缘选词浮动框位置问题

1

在页面边缘选词，浮动框会自动添加滚动条，建议检测划词位置，比如单词在屏幕右下角，那释义的浮动框可以单词左上角弹出。

扇贝背单词界面的一些问题

2

问题1：在扇贝背单词的界面，需要点两次发音才能发音。 问题2：在背单词时，如果使用了快捷键，比如D键翻到下一个单词，则上一次划词翻译弹出来的窗口还在，而且有时需要点击空白处两次，窗口才会消失。

add TorchExpert

1

Add an argument `--torchexpert` to enable TorchExpert analysis for profiling. Add submodule [TorchExpert](https://github.com/FindHao/TorchExpert). TorchExpert is an analysis tool for profiling results of PyTorch Profiler.

I found that in theme pelican-bootstrap3, there's no support of plugin pin_to_top which is really useful. So I add this feature for this theme.

Performance Issue in nmk_kernel

2

### 🚀 The feature Hi, when analyzing the profiling results of [detectron2_fasterrcnn_r_101_fpn](https://github.com/pytorch/benchmark) in inference mode, I found some inefficiencies. More details described in the following box. ### Motivation, pitch The...

topbar不显示

1

开机后topbar自动透明，pkill后重新打开正常。 更新6.19版本后出现该问题。

fix run_6.7b.sh for single gpu

2

The `train.py` will search for `run_6.7b.sh` by default for 6.7B size, but no files are named `run_6.7b.sh` in step1 and step3. So I copied them from the existing `run_6.7b_lora.sh` and...

Replace the frequent permutation generations with a pre-allocated permutation table

9

# What does this PR do? This PR fixes the performance issue in model big_bird. In model bigbird, the random masks are generated each time, and it calls a huge...