Benchmarks

Run on the target GPU after correctness passes:

python grouped-moe-gemv/benchmarks/benchmark.py --backend source

The harness compares the complete W4A4 producer+grouped region against the native W4A16 grouped path, a per-route grouped loop, and the legacy per-pair quantize+GEMV loop. It also reports W4A4 kernel-only latency so an upstream fused FP4 producer can be evaluated without charging a duplicate producer.