Qwen3.6-35B-A3B · vLLM · NVFP4 · conc 2

status: done

Configuration

ModelQwen/Qwen3.6-35B-A3B
CompanyAlibaba
FamilyQwen
Parameters35B / 3B (MoE)
EnginevLLM
Quant / precisionNVFP4
Why this quantconc-2 base (non-spec) point of the Qwen3.6-35B-A3B NVFP4 sweep — matched no-spec baseline for the MTP/DFlash conc-2 rows (EXPERIMENTS.md
Downloadnvidia/Qwen3.6-35B-A3B-NVFP4
Context window65536
Input modalitiestext, image, video (served text-only here)

Measured results

Prefill tok/s147.36
Decode tok/s113.19
Peak memory (GB)107.8 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85))
Completed2026-07-01 14:07 +0800

Full run command

# conc-2 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 2 500 600 256 \
  --quantization modelopt --trust-remote-code --reasoning-parser qwen3
# 267/500 prompts (hit 600 s cap), 0 errors. ready after 404 s.

qwen3.6-35b-a3b Alibaba Qwen NVFP4 16-40B Spark recipe conc-2

Notes

conc-2 base (no-spec) point of the Qwen3.6-35B-A3B NVFP4 sweep — matched baseline for the MTP/DFlash conc-2 rows (EXPERIMENTS.md #4/#14). Same NVIDIA ModelOpt NVFP4 recipe as the published conc-32 base; only --max-num-seqs changes.

  • Result (conc 2): prefill 147.36 / decode 113.19 tok/s aggregate; 267/500 prompts (hit the 600 s cap), 0 errors; peak mem 107.8 GB.
  • MTP speedup at conc-2: MTP 161.2 vs this base 113.19 = +42.4%larger than the conc-1 speedup (+25.6%). The base-vs-MTP win is not monotone-decaying from conc-1; it rises into low-batch before decaying at high concurrency (see the c4/c8/c16 siblings for the full curve). Worth flagging against the post’s “MTP win shrinks as the batch fills” framing — the shrink is real at the high end, but there’s a low-batch peak first.
  • TPOT 0.0 = qwen3 reasoning-parser client artifact — decode tok/s is the reported metric.
  • Sweep siblings: -c1 · -c4 · -c8 · -c16 · c32 (main). MTP counterpart: -mtp-c2.