Qwen3.6-35B-A3B · vLLM · NVFP4 · conc 8

# conc-8 base (no spec). Same recipe as the published conc-32 base — only --max-num-seqs differs.
scripts/bench-vllm-serving.sh nvidia/Qwen3.6-35B-A3B-NVFP4 65536 8 1000 600 256 \
  --quantization modelopt --trust-remote-code --reasoning-parser qwen3
# 574/1000 prompts (hit 600 s cap), 0 errors. ready after 385 s.

conc-8 base (no-spec) point of the Qwen3.6-35B-A3B NVFP4 sweep — matched baseline for the MTP/DFlash conc-8 rows (EXPERIMENTS.md #4/#14), so the post can quote the MTP/DFlash speedups at conc-8 against a real base instead of only conc-32. Same NVIDIA ModelOpt NVFP4 recipe as the published conc-32 base.

• Result (conc 8): prefill 264.33 / decode 241.82 tok/s aggregate; 574/1000 prompts (hit the 600 s cap), 0 errors; peak mem 107.9 GB.
• MTP speedup at conc-8: the original June MTP c8 (289.1, a 300 s-cap run) gave +19.5%, but a matched 600 s-cap recheck (2026-07-01) measured MTP c8 = 304.0, i.e. +25.7% over this base — the +19.5% “dip” was a short-cap sampling artifact. With that correction the MTP-vs-base ratio is a robust ~+25–30% across the sweep — +32.5% (c1) / +42.4% (c2) / +33.8% (c4) / +25.7% (c8) / +30.3% (c16) / +25.7% (c32) — essentially flat with a modest low-batch (c2–c4) bump, NOT a monotone decay (full discussion on the -c16 page). DFlash at conc-8 (269.9) = +11.6% over base, positive but below MTP.
• TPOT 0.0 = qwen3 reasoning-parser client artifact — decode tok/s is the reported metric.
• Sweep siblings: -c1 · -c2 · -c4 · -c16 · c32 (main). MTP counterpart: -mtp-c8.