Real production sizing scripts don't loop — they vectorize. This track starts from the four broadcasting primitives every numpy expression decomposes into (bias-add, batched dot, outer product, axis reduce) and then layers them up into the kind of multi-config memory and FLOPs sweeps that capacity-planning teams actually ship. By the end you should be able to take any scalar napkin formula and turn it into a one-line numpy expression that runs across thousands of configs at once.
12 problems · suggested order