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

status: done

Configuration

ModelQwen/Qwen3.6-35B-A3B
CompanyAlibaba
FamilyQwen
Parameters35B / 3B (MoE)
EnginevLLM + MTP (speculative decoding)
Quant / precisionNVFP4
Why this quantconc-2 fine-grained point of the Qwen3.6-35B-A3B NVFP4 + native-MTP sweep — same stack as conc-1/8/32 (nvidia ModelOpt NVFP4 base on marlin + in-repo MTP head on triton). Fills the 1→8 crossover region for the money chart (EXPERIMENTS.md
Downloadnvidia/Qwen3.6-35B-A3B-NVFP4
Context window65536
Input modalitiestext, image, video (served text-only here)

Measured results

Prefill tok/s176.53
Decode tok/s161.21
Peak memory (GB)106.0 (system MemAvailable delta (10s sampling) — vLLM static KV (util 0.85) + MTP head)
Completed2026-07-01 12:20 +0800

Full run command

# conc-2 fine-grained point. Same NVIDIA DGX Spark MTP recipe as conc-1/8/32 — only --max-num-seqs differs.
scripts/bench-vllm-serving.sh nvidia/Qwen3.6-35B-A3B-NVFP4 65536 2 1000 900 256 \
  --quantization modelopt --trust-remote-code --reasoning-parser qwen3 --moe-backend marlin \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3,"moe_backend":"triton"}'
# 569/1000 prompts (hit 900 s time cap), 0 errors. ready after 433 s. TTFT median 2996 ms.

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

Notes

conc-2 point of the Qwen3.6-35B-A3B NVFP4 + MTP sweep — fine-grained resolution for the money chart’s 1→8 crossover region (EXPERIMENTS.md #13/#14). Boots the same NVIDIA ModelOpt NVFP4 base (marlin) + in-repo MTP head (triton) as the published conc-1/8/32 rows; only --max-num-seqs changes.

  • Result (conc 2): prefill 176.53 / decode 161.21 tok/s aggregate; 569/1000 prompts (hit the 900 s time cap), 0 errors; peak mem 106.0 GB. Sits cleanly between conc-1 (93.9) and conc-8 (289.1), so the MTP decode curve is monotonic through the low-batch region — no anomaly.
  • Acceptance holds ~constant: mean accept-len ~3.0 (2.86–3.18), avg draft acceptance ~66–69%, per-position 0.84 / 0.67 / 0.53 — indistinguishable from conc-8 (~3.0 / ~67%) and conc-32 (~3.0 / 66–69%). Confirms the workload-driven, not concurrency-driven acceptance rule for the in-repo MTP head.
  • TPOT median reads 0.0 — the usual qwen3 reasoning-parser client artifact; trust the aggregate decode tok/s and the in-engine SpecDecoding throughput (~116 tok/s accepted).
  • Sweep siblings: -c1 · -c4 · -c8 · -c16 · maxctx/c32.