Qwen3.6-27B · vLLM · NVFP4 + MTP · conc 1

status: done

Configuration

ModelQwen/Qwen3.6-27B
CompanyAlibaba
FamilyQwen
Parameters27B (dense)
EnginevLLM + MTP (speculative decoding)
Quant / precisionNVFP4
Why this quantSingle-stream (conc-1) point of the Qwen3.6-27B NVFP4 + native-MTP sweep — the best-case MTP latency win (no batch contention), same stack as the conc-32/conc-8 runs. Single-stream decode tok/s + TPOT is where spec-decode helps most.
Downloadunsloth/Qwen3.6-27B-NVFP4
Context window65536
Input modalitiestext, image, video (served text-only here)

Measured results

Prefill tok/s3.35
Decode tok/s16.85
Peak memory (GB)107.59 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85) + MTP head)
Completed2026-06-23 10:48 +08

Full run command

# conc-1 single-stream latency point (200 prompts / 300 s cap).
scripts/bench-vllm-serving.sh unsloth/Qwen3.6-27B-NVFP4 65536 1 200 300 256 \
  --trust-remote-code --dtype bfloat16 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'
# only 20/200 prompts in the 300 s window (single stream + 515 s cold load). TTFT 336.6 ms, TPOT 51.5 ms.

qwen3.6-27b Alibaba Qwen NVFP4 16-40B conc-1

Notes

conc-1 (single-stream) point of the Qwen3.6-27B NVFP4 + MTP sweep. Best-case MTP latency — no batch contention.

  • Result (conc 1): decode 16.85 tok/s single-stream; TTFT median 336.6 ms, TPOT median 51.5 ms (≈19 tok/s/stream — bandwidth-bound for a 27B at ~4-bit on GB10). Peak mem 107.6 GB.
  • Small-sample caveat: only 20 prompts completed in the 300 s window (single stream + a 515 s cold load left little time), so acceptance is noisy — windows ranged 53–69%, mean accept-len ~2.6–3.1. Treat as directional; the conc-8/conc-32 runs (134/986 prompts) are the reliable acceptance numbers (~67–71%). The single-stream TPOT above is the solid metric here.