Qwen3.8-27B · SGLang · NVFP4 + DSpark · conc-2

status: done

Configuration

Model	`Qwen/Qwen3.8-27B`
Company	Alibaba
Family	Qwen
Parameters	27B (dense, hybrid linear-attn)
Engine	SGLang + DSpark (speculative decoding)
Quant / precision	NVFP4
Why this quant	MiaAI-Lab SGLang DGX Spark recipe (https://github.com/MiaAI-Lab/Qwen3.8-27B-SGLang-DGX-Spark). RadixArk NVFP4 W4A4 of Qwen3.8-27B on the cookbook-pinned lmsysorg/sglang:qwen38-27b image, DSpark block-7 / 8 draft tokens (RadixArk/Qwen3.8-27B-DSpark, unquant), mem-fraction-static 0.90, FP8 KV, GDN bf16, extra_buffer_lazy, native 262K, YaRN off. conc-2 point of the ShareGPT sweep (client conc 1/2/4/8/16/32).
Download	RadixArk/Qwen3.8-27B-NVFP4
Context window	262144
Input modalities	text, image, video (served text-only here)

Measured results

Prefill tok/s	59.42
Decode tok/s	43.36
Peak memory (GB)	110.23 (system MemAvailable delta (10s sampling) over the DSpark boot (MAX_CONCURRENT_REQUESTS=32))
Completed	2026-08-20 03:46 +0800

Full run command

# MiaAI-Lab Spark recipe, native 262K, YaRN off, mem-fraction-static 0.90.
cd ~/Desktop/Qwen3.8-27B-SGLang-DGX-Spark
MAX_CONCURRENT_REQUESTS=32 ./start-dspark.sh
python3 ~/Desktop/autobench/scripts/bench-serving.py \
  --base-url http://127.0.0.1:8888 --model qwen3.8-27b-sglang \
  --dataset ~/Desktop/autobench/benchmark_data/ShareGPT_V3_unfiltered_cleaned_split.json \
  --num-prompts 1000 --max-seconds 900 --concurrency 2 --max-tokens 256
# RESULT prefill_toks=59.42 decode_toks=43.36 completed=167 errors=0 duration_s=912.5 hit_time_cap=True
# image lmsysorg/sglang:qwen38-27b@sha256:febfb971c7352570fc445c466ebd6ffc9d896024958e544a60f2137fd85856b1

qwen3.8-27b Alibaba Qwen NVFP4 16-40B Spark recipe conc-2

Notes

conc-2 point of the Qwen3.8-27B SGLang NVFP4 + DSpark Spark-recipe line. MiaAI-Lab start-dspark.sh (block-7 / 8 draft, unquant RadixArk/Qwen3.8-27B-DSpark), native 262K, YaRN off, --mem-fraction-static 0.90, FP8 KV, GDN bf16, extra_buffer_lazy. ShareGPT via scripts/bench-serving.py. This conc ran with MAX_CONCURRENT_REQUESTS=32 (mamba pool = 32×4 slots). Larger pools at 0.90 rest at ~7–9 GB free and the 6 GB watchdog SIGKILLs the container (unified-memory GB10, no graceful OOM).

Result (conc 2): prefill 59.42 / decode 43.36 tok/s aggregate; 167/1000 prompts, 0 errors, duration 912.5s (hit the 900 s time cap). TTFT median 339.9 ms, TPOT median 44.1 ms, req throughput 0.183/s.
vs vLLM on this box (unsloth NVFP4, ctx 65536 — not a matched pair, but the closest): vLLM base c1 decode 11.16 / c8 75.04; vLLM MTP c1 21.35 / c8 126.68. DSpark c1 23.4 ties MTP (+110% vs base); DSpark c8 118.3 is ~1.57× vLLM base and a hair under MTP 126.7. Context here is 262144 vs 65536.
Acceptance: ~2.5–3.0 tokens/step of 8 drafted (~22–30%). ShareGPT chat; MiaAI’s LRUCache code probe was ~51 tok/s single-stream. Cross-check: far from 70–85% MTP/EAGLE3 — expected for DSpark on general chat, not a misconfig.
Memory: peak 110.23 GB. 32-running boot: weights ~22.6 GB + DSpark draft ~2.4 GB + GDN mamba ~28 GB (128 slots) + FP8 KV. Dropping MAX_CONCURRENT_REQUESTS shrinks mamba, not the 0.90 reservation. Watchdog floor 6 GB.
Series: c1 · c2 · c4 · c8 · c16 · c32. Matched DFlash2: c1 · c2 · c4 · c8 · c16 · c32.