Phil Wang
@mahdip72 ahh nice, yea, i can just add all of it up to 3 dimensions (since there are no conv4d yet) thanks for sharing the blog post from the author,...
@mahdip72 hey, currently busy with some work, but could take a look in about 2 week's time?
thanks Altay, will take a closer look tomorrow let me get that small bug squared away first 🙏
@tridao whatever you think is the best way let's roll with that!
> My guess is that it's better to store the dtanh. Calculating the inverse with log would be a lot slower. sounds good :pray: i'll get this wrapped up today...
@tridao @Narsil ok, i've removed all the local stuff to make compiling faster, currently doing the full half hour compiling to run the entire test suite let me know if...
@tridao 
left a machine on overnight to compile, and when i woke up, it was still compiling not sure what's going on, but i'm just going to test head dimension of...
> I think when there are more than some number templates, it actually takes way way longer. e.g. I had to split to more .cu files to compile the softcap...
@tridao after we land this, i can also spend some time and introduce a pull request to automate the head dimension filtering and disable flag, just to streamline development workflow...