Qwen3.8-27B · SGLang · NVFP4 + DFlash2 · conc-8

status: blocked

Configuration

ModelQwen/Qwen3.8-27B
CompanyAlibaba
FamilyQwen
Parameters27B (dense, hybrid linear-attn)
EngineSGLang + DFlash2 (speculative decoding)
Quant / precisionNVFP4
Why this quantSame Spark-recipe NVFP4 target (RadixArk/Qwen3.8-27B-NVFP4) as the DSpark line, but DFlash2 block-diffusion draft (z-lab/Qwen3.8-27B-DFlash2 @50307d4) on the derived lmsysorg/sglang:qwen38-27b-dflash2 image. mem-fraction-static 0.90. conc-8 point of the ShareGPT sweep (client conc 1/2/4/8/16/32).
DownloadRadixArk/Qwen3.8-27B-NVFP4
Context window262144
Input modalitiestext, image, video (served text-only here)

Measured results

Prefill tok/s
Decode tok/s
Peak memory (GB)
Completed2026-08-20 05:25 +0800

Full run command

# Not run — memory ceiling. Last attempted boot used MAX_CONCURRENT_REQUESTS=5.
cd ~/Desktop/Qwen3.8-27B-SGLang-DGX-Spark
MAX_CONCURRENT_REQUESTS=5 ./start-dflash.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 8 --max-tokens 256

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

Notes

BLOCKED — memory ceiling above DFlash2 conc-4. The 4-running boot (MAX_CONCURRENT_REQUESTS=5, 5 mamba slots/req → 4 running, --cuda-graph-max-bs-decode 4) SIGKILL’d 69 s into conc-4 (27 ok / 973 err, exit 137) after a scheduler GC freeze on top of the 0.90 reservation (~7 GB free at ready). The 8-running DFlash2 boot already died mid conc-2. Do not retry conc-8 at 0.90.

  • Last partial point: c4 decode 85.1 tok/s (n=27, noisy).
  • Clean last point: c1 decode 30.0 (116/1000, 0 errors).
  • Series: c1 · c2 · c4 · c8 · c16 · c32.