Paul Fultz II
Paul Fultz II
We still need to update our quantizer so we dont insert zero scales in the first place and we insert a small value instead.
@causten Why was this merged? There was no unit tests added for this.
> does every single matcher go through find_matches_for? Yes. > This is a good way to globally enable/disable match timing. But maybe we also want to enable/disable timing programmatically for...
> Clang tidy thinks things are too complex @pfultz2 Yea I need to do some refactoring here.
> Which libc++ on which system is failing? I am using LLVM 20 and 21 locally and so far the tests passed. I am using apple clang version 16, but...
> Please have a look. I obviously have no idea what I am doing. I dont really have a context of what you are trying to do here, but it...
> We did things wrong and doing it right somehow doesn't really work although it should. Sure, but #7658 doesn't change that. > So cleaning that up first could help...
Closing as #4427 and #4428 has the fixes with unit tests.
The problem is that we would now output fp16 instead of int8. We should try to re-enable this matcher. Of course, there is accuracy loss from quantization, but we would...
I dont see a reason to use the bf16 header since we already use the compiler type directly(which enables us to use vector types). The fp8 header will be useful...