| # Benchmarks | |
| Run on the target GPU after correctness passes: | |
| ```bash | |
| 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. | |