Trace the router
Record selected routed experts, weights, shared experts, layer, generation step, and model provenance.
Project journal · Phase 01 · In progress
See the route.Measure the work.
MoEscope follows Mixture-of-Experts inference from a token’s routing decision to the GPU workload it creates—then makes the evidence inspectable.
Fig. 01 / expert loadTrace ready
256routed experts
6selected / token
2shared experts
01 / Research notes
The work, documented while it is still being worked out.
02 / The method
From model semantics to hardware evidence, one trace at a time.
Describe the work
Turn assignments into expert histograms, grouped matrix shapes, tile counts, padding, and theoretical waves.
Test the evidence
Replay the same workload against a correct reference and measured GPU backends without hiding losses.