Qwen3.6-27B · vLLM · NVFP4 + MTP · conc 8

status: done

Configuration

ModelQwen/Qwen3.6-27B
CompanyAlibaba
FamilyQwen
Parameters27B (dense)
EnginevLLM + MTP (speculative decoding)
Quant / precisionNVFP4
Why this quantConcurrency-8 point of the Qwen3.6-27B NVFP4 + native-MTP sweep — same stack as the conc-32 run (unsloth NVFP4 base + in-repo MTP), lower batch for the latency-characterization point. Acceptance should hold ~constant vs conc-32 (workload-driven); what changes is whether the MTP speedup materializes.
Downloadunsloth/Qwen3.6-27B-NVFP4
Context window65536
Input modalitiestext, image, video (served text-only here)

Measured results

Prefill tok/s120.27
Decode tok/s109.05
Peak memory (GB)106.65 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85) + MTP head)
Completed2026-06-23 10:33 +08

Full run command

# conc-8 latency point (500 prompts / 300 s cap, matching the FP8-MTP -c8 convention).
scripts/bench-vllm-serving.sh unsloth/Qwen3.6-27B-NVFP4 65536 8 500 300 256 \
  --trust-remote-code --dtype bfloat16 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'
# 134/500 prompts (hit 300 s cap), 0 errors. TTFT median 579.5 ms, TPOT median 60.8 ms.

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

Notes

conc-8 point of the Qwen3.6-27B NVFP4 + MTP sweep. Lower-batch latency characterization (cap 500 prompts / 300 s — reached 134 at the time cap, 0 errors).

  • Result (conc 8): prefill 120.3 / decode 109.05 tok/s aggregate; per-stream TTFT median 579.5 ms, TPOT median 60.8 ms (≈16 tok/s/stream). Peak mem 106.7 GB.
  • Acceptance ~71% / mean accept-len 3.1 (per-position 0.87 / 0.69 / 0.55) — holds vs the conc-32 run’s 67% (even a touch higher), confirming acceptance is workload-driven, not concurrency-driven. Aggregate tok/s is lower than conc-32 (274) purely because 8 streams push fewer total tokens — the meaningful low-conc metric is the per-stream TPOT above.