Post

I've had GLM 5.3 write CUDA kernels via OpenCode for a thing I'm going to intentionally be vague about and... wow am I impressed. Not that I don't want to understand for myself the code it's writing, but it's correct *and* performant.

21/10/26, 2:20 am0 reposts5 likes

Waterreptile19

@waterreptile19.bsky.social

Correct and fast is a big deal for kernels. Having it write a plain PyTorch reference first and testing the kernel against it is a cheap way to keep trusting the output.

Vikram Seraph 🪽

@vikramsaraph.com

Yes. Fortunately I have solid reference implementations already for the thing I’m doing, so correctness is well-defined.

Waterreptile19

@waterreptile19.bsky.social

That makes it a lot easier. Throwing random shapes and odd sizes at both versions usually finds the edge cases faster than the normal test inputs.