How do "tactics" work?

Are there any resources to learn what these tactics are? I understand on a high level they are optimizing the network, but I’d like to learn a bit about what specifically they do, maybe some of the math behind them, how many there are, etc. Does anyone know any good resources for these? Im sure there are some on NVIDIAs docs, but I couldn’t find any decent ones.

Hi @duner, the specific tactic configurations are internal to TensorRT. For each layer in the network, there are a number of potential tactics that could be chosen to execute that layer. Each tactic has different kernel configurations, methods, ect. that they use. The fastest one that runs in the allowable workspace memory gets selected to execute the layer.

Is there any way to see what the actual tactics are? Im sure they work fine and everything, im just curious how specifically they work…