Qwen3.6-27B · vLLM · NVFP4 · conc 16

status: done

Configuration

ModelQwen/Qwen3.6-27B
CompanyAlibaba
FamilyQwen
Parameters27B (dense)
EnginevLLM
Quant / precisionNVFP4
Why this quantconc-16 base (non-spec) point of the Qwen3.6-27B NVFP4 sweep — matched no-spec baseline for the MTP conc-16 row (EXPERIMENTS.md
Downloadunsloth/Qwen3.6-27B-NVFP4
Context window65536
Input modalitiestext, image, video (served text-only here)

Measured results

Prefill tok/s147.97
Decode tok/s116.57
Peak memory (GB)107.72 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85))
Completed2026-07-01 22:58 +0800

Full run command

# conc-16 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 16 1000 600 256 \
  --trust-remote-code --dtype bfloat16

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

Notes

conc-16 base (no-spec) point of the Qwen3.6-27B NVFP4 sweep — completes the base curve at c1/2/4/8/16/32 (matched baseline for the MTP conc-16 row, EXPERIMENTS.md #4/#14).

  • Result (conc 16): prefill 147.97 / decode 116.57 tok/s aggregate; 0 errors; peak mem 107.7 GB.
  • The full base decode curve: 9.33 → 18.11 → 35.54 → 67.06 → 116.57 → 187.74 tok/s at c1/2/4/8/16/32. Scaling factor per doubling falls from ~1.94 (c1→2) to 1.74 (c8→16) to 1.61 (c16→32) — the dense 27B transitions from bandwidth-bound (near-linear) toward compute-bound (diminishing returns) as the batch fills.
  • MTP-vs-base delta on the -mtp-c16 page — this is where spec-decode’s edge is expected to compress, since a filling dense batch leaves less spare compute for the draft.
  • Sweep siblings (base): -c1 · -c2 · -c4 · -c8 · c32 (main).