Blackroom
A Transformer bottlenecks differently at every input shape. We time the candidates at runtime and cache the winner per shape: ~10x geomean speedup, 13/13 correct, on a laptop GPU.
A Transformer bottlenecks differently at every input shape. We time the candidates at runtime and cache the winner per shape: ~10x geomean speedup, 13/13 correct, on a laptop GPU.
Log in or sign up for Devpost to join the conversation.