[qeternity]

PyTorch is only part of it. There is still a huge amount of CUDA that isn’t just wrapped by PyTorch and isn’t easily portable.

[svara]

... but not in deep learning or am I missing something important here?

[qeternity]

Yes, absolutely in deep learning. Custom fused CUDA kernels everywhere.

[Scene_Cast2]

Yep. MoE, FlashAttention, or sparse retrieval architectures for example.