Qwen3.8-27B · vLLM · NVFP4 · conc-4

status: done

Configuration

ModelQwen/Qwen3.8-27B
CompanyAlibaba
FamilyQwen
Parameters27B (dense, hybrid linear-attn)
EnginevLLM
Quant / precisionNVFP4
Why this quantSame unsloth NVFP4 checkpoint as the conc-32 headline — only --max-num-seqs differs. Fills the c2→c8 gap in the concurrency curve, and the non-speculative baseline for the conc-4 MTP config.
Downloadunsloth/Qwen3.8-27B-NVFP4
Context window65536
Input modalitiestext, image, video (served text-only here)

Measured results

Prefill tok/s48.39
Decode tok/s38.62
Peak memory (GB)102.91 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85))
Completed2026-08-21 01:49 +0800

Full run command

# conc-4 point. Identical to the conc-32 base recipe except --max-num-seqs.
scripts/bench-vllm-serving.sh unsloth/Qwen3.8-27B-NVFP4 65536 4 1000 900 256 \
  --trust-remote-code --dtype bfloat16
# 151/1000 prompts in the 900 s window, 0 errors. ready after 343 s.
# TTFT median 314.5 ms, TPOT median 99.7 ms.

qwen3.8-27b Alibaba Qwen NVFP4 16-40B conc-4

Notes

38.6 tok/s at conc-4 — batch scaling is still essentially free here.

  • Result (conc 4): prefill 48.39 / decode 38.62 tok/s; TTFT median 314.5 ms, TPOT median 99.7 ms. 151/1000, 0 errors, hit the 900 s cap. Load 343 s.
  • Near-perfect doubling up to c8. Decode goes 11.16 → 20.11 → 38.62 → 75.04 tok/s for c1 → c2 → c4 → c8: per-step ratios of 1.80×, 1.92×, 1.94×. Below c8 this box is memory-bandwidth bound with compute to spare, so each added stream is nearly a free extra token stream. Per-stream TPOT over that whole range moves only 88.5 → 99.7 ms (+13%).
  • c4 costs a single user ~12% latency for 3.5× the aggregate throughput versus running alone. If the deployment has more than one user, there is no reason to run below c4.
  • Memory 102.9 GB, the low point of the sweep’s 103–107 GB band — noise on the gpu-memory-utilization reservation, not a concurrency effect. Resident breakdown on the conc-32 base.
  • Base-vs-MTP comparison on conc-4 + MTP.
  • Served text-only (mm_served: false), ctx 65536, consistent with the rest of the sweep.