Binary working-set optimizer

Put heat where memory is.

Profile routed experts, set a GPU memory cap, and solve the integral placement that maximizes activation-weighted latency savings.

Use measured data in production. Shares must sum to 1.00. This model excludes transfer, contention, batching, NUMA, cache, and multi-expert overlap.

Load the bundled synthetic profile or paste measured JSON.
—GPU memory used
—activation coverage
—expected saving
—CPU activation tail

GPU working set

CPU capacity tail

Integral optimum
—Maximize summed share × (CPU ms − GPU ms), subject to footprint ≤ budget.
Boundary
Synthetic until replacedThe sample proves the optimizer lifecycle, not KTransformers or hardware benchmark performance.
Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.