Propagate tensor shapes through the graph using operation semantics

Evaluate operations with all-constant inputs at compile time

Remove operations not reachable from any output

Fuse compatible operation sequences into single optimized kernels

Convert float32 weights and activations to int8 for faster inference

Convert tensor layouts from NCHW to NHWC for optimal GPU execution

Analyze tensor lifetimes and assign memory offsets to minimize peak usage