Qwen3.6-27B · vLLM · NVFP4 · conc 8

status: done

Configuration

ModelQwen/Qwen3.6-27B
CompanyAlibaba
FamilyQwen
Parameters27B (dense)
EnginevLLM
Quant / precisionNVFP4
Why this quantconc-8 base (non-spec) point of the Qwen3.6-27B NVFP4 sweep — matched no-spec baseline for the MTP conc-8 row (EXPERIMENTS.md
Downloadunsloth/Qwen3.6-27B-NVFP4
Context window65536
Input modalitiestext, image, video (served text-only here)

Measured results

Prefill tok/s68.25
Decode tok/s67.06
Peak memory (GB)107.09 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85))
Completed2026-07-01 22:32 +0800

Full run command

# conc-8 base (no spec). Same recipe as the published conc-32 base — only --max-num-seqs differs.
scripts/bench-vllm-serving.sh unsloth/Qwen3.6-27B-NVFP4 65536 8 1000 600 256 \
  --trust-remote-code --dtype bfloat16

qwen3.6-27b Alibaba Qwen NVFP4 16-40B conc-8

Notes

conc-8 base (no-spec) point of the Qwen3.6-27B NVFP4 sweep — matched baseline for the MTP conc-8 row (EXPERIMENTS.md #4/#14).

  • Result (conc 8): prefill 68.25 / decode 67.06 tok/s aggregate; 0 errors; peak mem 107.1 GB.
  • Scaling starts to bend by conc-8 (9.33 → 18.11 → 35.54 → 67.06 at c1/2/4/8 ≈ ×1.94/×1.96/×1.89) — still strong but no longer perfectly linear as the dense 27B begins to fill the compute.
  • MTP speedup at conc-8: MTP -mtp-c8 109.05 vs this base 67.06 = +62.6% — the dense model keeps a large MTP win at moderate batch (vs the 35B-A3B MoE’s ~+20% at c8), because a dense target has less spare compute to lose to draft overhead. MTP accepts ~71% / accept-len ~3.1.
  • Sweep siblings (base): -c1 · -c2 · -c4 · -c16 · c32 (main).