Neither vLLM's nor SGLang's official MoE benchmark modeled skewed (zipf)
expert load; upstreamed an opt-in --expert-load-distribution flag to
vLLM's benchmarks/kernels/benchmark_moe.py. Measured: --tune under zipf
vs. uniform routing picks a different winning Triton config at 4 of 5
tested batch sizes. This phase landed no code in dispatch itself, only
docs -- its docs merged alongside Phase 3 in PR #3. Open PR:
vllm-project/vllm#57100. Cost: $0.77.

Full account: docs/findings/phase-2/2026-09-15-phase-2-vllm-benchmark-run.md