Zhangir Azerbayev

For a long time I've been wanting to write a detailed documentation page using Sphinx. I am swamped for the next week or so, but could start working on it...

I am also having this issue with zero stage 1. Here is the log of my first few steps of training: ``` [2023-07-25 16:20:45,998] [INFO] [checkpointing.py:529:forward] Activation Checkpointing Information [2023-07-25...

Note that If I set gradient clipping to some very low value (in the following log, 1e-7) the loss jump at step 1 still occurs, but the loss doesn't change...

upgrading `gdown` to 4.7.3 fixed this for me

Currently, we require users to build the package from source from within this repository. Once we get the compiled version of the repl working, the python package should be installable...

Thanks for the reviews. I see that since I opened this PR, the bug that prevented the repl from compiling was fixed. Therefore I'm going to convert this PR back...

That is basically what I'm doing in pySagredo. However doing it this way squares runtime. I haven't run into any scalability issues yet, so this issue isn't high priority. However,...

@EdAyers Would you be interested in helping us with UI and databases?

Here's a rough suggestion. Let me know if you want it to be shorter/longer or if there are parts that are unclear. The mathlib docs is testing a beta feature...