Comment by wmf
Ultimately all of them except Gemini.
Ultimately all of them except Gemini.
Wrong.
It is the same PyTorch whether it runs on an AMD or an NVIDIA GPU.
The exact same PyTorch, actually.
Are you're trying to suggest that the machine code that runs on the GPU is the one that is different?
If you knew a bit more, you would know that this is the case even between different generations of GPUs of the same vendor; making that argument completely absurd.
Wrong.
Show me one single CUDA kernel on Llama's source code.
(and that's a really easy one, if one knows a bit about it)