Qwen3.6-27B · vLLM · FP8 + MTP

status: done

Configuration

ModelQwen/Qwen3.6-27B
CompanyAlibaba
FamilyQwen
Parameters27B (dense)
EnginevLLM + MTP (speculative decoding)
Quant / precisionFP8
Why this quantQwen3.6-27B FP8 + the model's own native MTP module (mtp.safetensors ships in the base repo) — built-in multi-token-prediction speculative decoding, no separate draft.
DownloadQwen/Qwen3.6-27B-FP8
Context window65536
Input modalitiestext, image (served text-only here)

Measured results

Prefill tok/s241.68
Decode tok/s240.92
Peak memory (GB)107.09 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85) + MTP head)
Completed2026-06-22 18:46 +08

Full run command

# base Qwen3.6-27B-FP8 + native MTP (mtp.safetensors ships in-repo) via vLLM --speculative-config.
docker run -d --gpus all --ipc=host -p 8000:8000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface --env HF_TOKEN=*** \
  vllm/vllm-openai:cu130-nightly Qwen/Qwen3.6-27B-FP8 \
  --host 0.0.0.0 --port 8000 --max-model-len 65536 --gpu-memory-utilization 0.85 --max-num-seqs 32 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'
python3 scripts/bench-serving.py --base-url http://localhost:8000 --model Qwen/Qwen3.6-27B-FP8 \
  --dataset benchmark_data/ShareGPT_V3_unfiltered_cleaned_split.json \
  --num-prompts 1000 --max-seconds 900 --concurrency 32 --max-tokens 256

qwen3.6-27b Alibaba Qwen FP8 16-40B conc-32

Notes

Native MTP is a clean +56% win on Qwen3.6-27B — and the first config with a captured acceptance rate, which explains exactly why. Qwen3.6-27B FP8 + the model’s built-in MTP head, on vLLM.

  • Workload: ShareGPT V3, concurrency 32. 867/1000, 0 errors — clean, hit the cap. Loaded + CUDA-graph captured in 504 s (MTP adds compile time).
  • Throughput (aggregate, conc 32): prefill 241.7 tok/s, decode 240.9 tok/s vs the base Qwen3.6-27B’s 154.7+56% decode from MTP at conc 32.
  • Acceptance rate (captured from vLLM’s SpecDecoding metrics): mean acceptance length ≈ 3.0, avg draft acceptance ≈ 67%, per-position 0.84 / 0.67 / 0.51 (for the 3 draft positions). With num_speculative_tokens=3, ~3 tokens are emitted per target step on average (the always-accepted token + ~2 accepted drafts). High acceptance on ShareGPT’s fairly predictable chat continuations is what converts into the +56%.
  • Why this won where gpt-oss-120b’s EAGLE3 lost — same conc 32, opposite size. A 27B model still has GB10 compute headroom at batch 32, so the MTP draft+verify rides on otherwise-idle FLOPs and the ~3× tokens/step turns into real throughput. The 120B gpt-oss was already compute-saturated, so the same idea went −45%. Acceptance ~67% here is healthy; even with good acceptance, a saturated large model wouldn’t benefit. Headroom × acceptance is the predictor — and this 27B has both.
  • Memory: 107.1 GB is the vLLM --gpu-memory-utilization 0.85 reservation (+ the small MTP head), not the footprint (FP8 weights ≈ 27 GB).