Qwen3.6-27B · vLLM · NVFP4 + MTP · conc 2

status: done

Configuration

ModelQwen/Qwen3.6-27B
CompanyAlibaba
FamilyQwen
Parameters27B (dense)
EnginevLLM + MTP (num_speculative_tokens=3) (speculative decoding)
Quant / precisionNVFP4
Why this quantconc-2 MTP point of the Qwen3.6-27B NVFP4 sweep — the base+MTP speedup-decay curve (EXPERIMENTS.md
Downloadunsloth/Qwen3.6-27B-NVFP4
Context window65536
Input modalitiestext, image, video (served text-only here)

Measured results

Prefill tok/s56.13
Decode tok/s35.68
Peak memory (GB)109.42 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85) + MTP head)
Completed2026-07-01 23:14 +0800

Full run command

# conc-2 MTP. Same recipe as -mtp-c8 — only --max-num-seqs differs.
scripts/bench-vllm-serving.sh unsloth/Qwen3.6-27B-NVFP4 65536 2 1000 600 256 \
  --trust-remote-code --dtype bfloat16 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'

qwen3.6-27b Alibaba Qwen NVFP4 16-40B conc-2

Notes

conc-2 MTP point of the Qwen3.6-27B NVFP4 sweep — paired with the matched base -c2 for the base-vs-MTP speedup-decay curve (EXPERIMENTS.md #14).

  • Result (conc 2): prefill 56.13 / decode 35.68 tok/s aggregate; 0 errors; peak mem 109.4 GB.
  • MTP speedup at conc-2: +97.0% (35.68 vs base 18.11) — nearly a 2× single-stream-ish win on the dense 27B. The dense MTP advantage is largest at low batch: +80.6% (c1) → +97.0% (c2) → +62.6% (c8), far above the 35B-A3B MoE’s ~+25–40%, because a dense target has essentially no spare compute for the batch to reclaim.
  • Acceptance ~68%, accept-len ~3.0-of-3 — the MTP draft is highly efficient (per-position 0.85/0.69/0.55), matching the -c8 row’s ~71%. Acceptance is workload-driven and batch-stable here (unlike the gpt-oss EAGLE3 low-batch pathology).
  • Sweep: base -c2 · MTP -c1 · -c4 · -c8 · c32.