Episode 33: Tri Dao, Stanford: On FlashAttention and sparsity, quantization, and efficient inference
Generally Intelligent
00:00
The Importance of Sparsity in GPUs
In optimization, sparsity has been an important idea. We wanted to make them actually hardware efficient so that we can train with sparsity. That has been difficult for a lot of people simply because the hardware that we use GPUs are very much block oriented devices. So it wasn't in terms of expressiveness, we had to sacrifice it a little bit to gain in hardware efficiency. And after that paper, we thought a little bit more and then we found this nice parameterization, carmonark, that we didn't have to sacrifice expressiveness. It can represent anything that perhaps butterfly matrices can represent at the same time that remains hardware efficient. But that could work pretty well for training things
Transcript
Play full episode
