3 ms·
That simply isn't true in practice. Maybe for inference, but even then you're running up against common CUDA kernels such as FlashAttention which will be far f
by throwaway314155 2y ago
That simply isn't true in practice. Maybe for inference, but even then you're running up against common CUDA kernels such as FlashAttention which will be far from plug and play with PyTorch.