Qwen3.8-27B · SGLang · NVFP4 + DSpark · conc-16

status: blocked

Configuration

ModelQwen/Qwen3.8-27B
CompanyAlibaba
FamilyQwen
Parameters27B (dense, hybrid linear-attn)
EngineSGLang + DSpark (speculative decoding)
Quant / precisionNVFP4
Why this quantMiaAI-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-16 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=8.
cd ~/Desktop/Qwen3.8-27B-SGLang-DGX-Spark
MAX_CONCURRENT_REQUESTS=8 ./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 16 --max-tokens 256

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

Notes

BLOCKED — memory ceiling above DSpark conc-8. The 8-running boot (MAX_CONCURRENT_REQUESTS=8, mamba 32 slots) SIGKILL’d mid conc-8 after 406 successful requests (watchdog floor 6 GB, container exit 137, unified-memory GB10). 16- and 32-running boots already died earlier (32-boot during autotune at an 8 GB floor; 16-boot at the conc-8 start). Raising max-mamba-cache-size to actually admit 16 running requests adds ~7–14 GB of GDN state on top of the 0.90 reservation and is what takes the box down. Do not retry the identical 0.90 / conc-16 config. Dropping mem-fraction below 0.90 would be a separate util-adjusted regime.

  • Last good point: c8 decode 118.3 tok/s (partial).
  • Series: c1 · c2 · c4 · c8 · c16 · c32.