Qwen3.6-35B-A3B · vLLM · NVFP4 + MTP · conc 1

status: done

Configuration

ModelQwen/Qwen3.6-35B-A3B
CompanyAlibaba
FamilyQwen
Parameters35B / 3B (MoE)
EnginevLLM + MTP (speculative decoding)
Quant / precisionNVFP4
Why this quantSingle-stream point of the Qwen3.6-35B-A3B NVFP4 + native-MTP sweep — same stack as the conc-32 run (nvidia ModelOpt NVFP4 base on marlin + in-repo MTP head on triton). Best-case MTP latency with no batch contention; the per-stream TPOT here is the headline number.
Downloadnvidia/Qwen3.6-35B-A3B-NVFP4
Context window65536
Input modalitiestext, image, video (served text-only here)

Measured results

Prefill tok/s116.8
Decode tok/s93.91
Peak memory (GB)108.57 (system MemAvailable delta (10s sampling) — vLLM static KV (util 0.85) + MTP head)
Completed2026-06-23 13:02 +08

Full run command

# conc-1 single-stream latency point (200 prompts / 300 s cap). Same NVIDIA Spark MTP recipe.
scripts/bench-vllm-serving.sh nvidia/Qwen3.6-35B-A3B-NVFP4 65536 1 200 300 256 \
  --quantization modelopt --trust-remote-code --reasoning-parser qwen3 --moe-backend marlin \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3,"moe_backend":"triton"}'
# 111/200 prompts (hit 300 s cap), 0 errors. ready after 420 s. TTFT median 2394 ms.

qwen3.6-35b-a3b Alibaba Qwen NVFP4 16-40B Spark recipe conc-1

Notes

conc-1 (single-stream) point of the Qwen3.6-35B-A3B NVFP4 + MTP sweep. Best-case MTP latency — no batch contention. Pairs with the conc-32 done run (decode 541 tok/s agg) and the conc-8 sibling.

  • Result (conc 1): prefill 116.8 / decode 93.91 tok/s single-stream aggregate; 111/200 prompts (hit 300 s cap), 0 errors; peak mem 108.57 GB; TTFT median 2394 ms. The single-stream decode is the headline low-conc number — no batch contention, so this is best-case MTP latency for this checkpoint.
  • Acceptance holds: mean accept-len ~3.0 (range 2.79–3.24, median 2.99), avg draft acceptance ~66%, per-position ~0.84/0.66/0.51 — identical to conc-8 (~3.0) and conc-32 (~3.0). Acceptance is workload-driven, not concurrency-driven — same conclusion across the whole 35B-A3B and 27B sweeps.
  • Single-stream accepted throughput ~70 tok/s (SpecDecoding log) vs ~93.9 tok/s client decode — the spec head sustains ~70 accepted tokens/s of useful generation per stream at acc-len 3.
  • 600 s-cap recheck (2026-07-01): this page’s 93.91 was a 300 s / 200-prompt run; a matched 600 s-cap re-measurement gave decode 99.04 tok/s (233/400, 0 err) — ~5% higher (the short cap over-weights warmup; same effect lifted MTP c8 289→304). This is the money-chart-relevant MTP c1 baseline: against it, DFlash’s conc-1 lead is only ~+2.9% (101.9 vs 99.04), not the +8.5% quoted against the 93.91 short-cap number — see …ultimate-dflash and the base -vllm-c1.
  • TPOT caveat: the qwen3 reasoning-parser zeros the client TPOT median (reads 0.0) — corroborate with the in-engine SpecDecoding metrics above.