[davidatbu]

Good point. But the overall point about Mojo availing a different level of abstraction as compared to Python still stands: I imagine that no amount of magic/operator-fusion/etc in `torch.compile()` would let one get reasonable performance for an implementation of, say, flash-attn. One would have to use CUDA/Triton/Mojo/etc.

[boroboro4]

But python is already operating fully on different level of abstraction - you mention triton yourself, and there is new python cuda api too (the one similar to triton). More to this - flash attention 4 is actually written in python.

Somehow python managed to be both high level and low level language for GPUs…

[davidatbu]

IIUC, triton uses Python syntax, but it has a separate compiler (which is kinda what Mojo is doing, except Mojo's syntax is a superset of Python's, instead of a subset, like Triton). I think it's fair to describe it as a different language (otherwise, we'd also have to describe Mojo also as "Python"). Triton's website and repo describes itself as "the Triton language and compiler" (as opposed to, I dunno, "Write GPU kernels in Python").

Also, flash attention is at v3-beta right now? [0] And it requires one of CUDA/Triton/ROCm?

[0] https://github.com/Dao-AILab/flash-attention

But maybe I'm out of the loop? Where do you see that flash attention 4 is written in Python?