James Brown

Now I have made some progress here. I make the dataset preparation process single threaded in ASM mode (because I cannot figure out how to sleep in assembly code in...

That being said, I think it is possible to create some better mechanism for controlling thread occupation, by first measuring the average time to perform a single loop in hash...

I found two solutions from the sources. Will try later and give feedback. One is `export HSA_ENABLE_SDMA=0`, the other is disabling power features. ```bash sudo vim /etc/default/grub # add "amdgpu.ppfeaturemask=0xffff3fff...

It works but in order to have system RAM as VRAM, not limited by the BIOS VRAM setting, further investigation is needed. When compile [force-host-alloction-APU](https://github.com/segurac/force-host-alloction-APU) with the script below and...

Try probing with `moviepy/ffmpeg/io/ffmpeg_reader.py`, I find the ffmpeg commands starting from the end of the video file, which may cause the problem: ```bash /usr/local/lib/python3.9/dist-packages/imageio_ffmpeg/binaries/ffmpeg-linux64-v4.2.2 -ss 2.060000 -i -ss 1.000000 -loglevel...

This issue happens everywhere. Many of these claimed as a fix but as we can see it is not fixed at all. Take #1323 as an example.

嘤嘤嘤

RLHF requires creating multiple models like SFT, RM, PPO-tuned model. Is it possible to improve storage and memory efficiency, reduce computation if we freeze some huge layers of the pretrained...

To address this proposal even further, I think a good reward function can self-evolve and adapt to new environments (when the data source is no longer fixed static "archives" but...

This library is "reward function guided". In order to achieve ChatGPT-level fluency and capability, you need a proper reward function. Details discussed here: #146