Qwen3.6-27B · vLLM · NVFP4 · conc 1

# conc-1 base (no spec). Same recipe as the published conc-32 base — only --max-num-seqs differs.
scripts/bench-vllm-serving.sh unsloth/Qwen3.6-27B-NVFP4 65536 1 500 600 256 \
  --trust-remote-code --dtype bfloat16
# 22/500 prompts (hit 600 s cap; dense 27B is slow single-stream), 0 errors. ready after 396 s.

conc-1 base (no-spec) point of the Qwen3.6-27B NVFP4 sweep — matched baseline for the MTP conc-1 row (EXPERIMENTS.md #4/#14). Same Unsloth NVFP4 recipe as the published conc-32 base; only --max-num-seqs changes.

• Result (conc 1): prefill 1.81 / decode 9.33 tok/s; 22/500 prompts (hit the 600 s cap), 0 errors; peak mem 108.5 GB.
• Dense, so slow single-stream — the point of the curve. At conc-1 this dense 27B decodes 9.33 tok/s, ~8× slower than the MoE 35B-A3B NVFP4’s 74.7 (which activates only ~3B/token) — the expected dense-vs-MoE gap when every token pays the full 27B. This is why MTP matters more here.
• MTP speedup at conc-1: MTP 16.85 vs this base 9.33 = +80.6% — the tree-free MTP draft nearly doubles single-stream decode on the dense model (vs the MoE’s more modest +25.6%), because the dense target has the spare-nothing single-stream that spec-decode helps most.
• TPOT/TTFT are the usual conc-1 single-stream figures; decode tok/s is the reported metric.
• Sweep siblings (base): -c2 · -c4 · -c8 · -c16 · c32 (main). MTP counterpart: -mtp-c1.