Ivan Komarov

> your current version and 'past_key_values" version are mathematically different, right? I think they are pretty much identical up to renaming (not intentionally, it's just that it's hard to implement...

I guess it's best for me to close this PR, given that the upstream version has diverged from mine (after integrating FlashAttention) and that I never intended this to be...

Whoops, sorry, I didn't notice you were referring to Andrej, not me. But if you don't mind me answering anyway: *mathematically*, the version in the main branch of this repo...

> So I'm left wondering how they get around this problem once the input exceeds the max window and tokens are effectively shifted one position between steps This is a...

> Perhaps that makes sense? This sounds promising, but I am little unclear on how exactly this will work. As far as I understand, once you compute an embedding for...

Oooooh, now I see, thanks! This is really clever and makes perfect sense. I bet there are a lot of real-life scenarios where you have a model with absolute position...

@luigiScarso Hi, this is awesome! We have a [similar project](https://github.com/busytex/busytex) with slightly broader scope: we also compile other TeX binaries, such as pdfTeX and XeTeX, into a single busytex-style binary....

@karashiiro I took a stab at debugging this, and I think I know the reason why it exits. The immediate problem is that `getc()` prints `Invalid parameter passed to C...

@karashiiro > zig build does allow passing custom libc overrides via --libc I think `--libc` changes be the C library (e.g., `ucrt`, which contains `getc()` and `getwc()`), not the C++...

> what's the actual root cause (why zero bytes are being input to a UTF-8 stream in the first place) Whoa, I think I found the root cause. It's [this...