DGX Spark 单机部署 Qwen3.8-27B：NVFP4 量化 + DSpark 投机解码实践

#!/usr/bin/env bash

# ============================================================

# Qwen3.8-27B-NVFP4 + DSpark 投机 单节点 DGX Spark (GB10/SM120) SGLang 启动脚本

# 依据: PR #34966 官方验证命令 (NVFP4 + DSpark) + SGLang cookbook GB10 适配

#   - SM120 native FP4 runner (不需要 --fp4-gemm-backend marlin, 那是 SM89 用的)

#   - KV cache 用 fp8_e4m3 减半 (官方验证命令), 否则 bf16 KV 50GB 会爆内存

#   - --mamba-ssm-dtype bfloat16 减半 Mamba state

#   - --mem-fraction-static 0.88 (官方验证命令, 之前 0.95 是爆内存主因之一)

# ============================================================

set -euo pipefail



MODEL_DIR=${MODEL_DIR:-/data/models/Qwen3.8-27B-NVFP4}

DRAFT_DIR=${DRAFT_DIR:-/data/models/Qwen3.8-27B-DSpark}

IMAGE=lmsysorg/sglang:qwen38-27b

NAME=sglang-qwen38

PORT=8000                 # 统一 8000 端口 (与旧 vLLM 接入一致)



# DSpark 投机 (PR #34966 官方验证命令参数, 减去镜像不支持的 --speculative-draft-kv-cache-dtype)

SPEC_ARGS="--speculative-algorithm DSPARK \

--speculative-draft-model-path /models/Qwen3.8-27B-DSpark \

--speculative-dspark-block-size 7 \

--speculative-draft-model-quantization unquant \

--speculative-draft-attention-backend flashinfer"



echo ">>> 启动前检查"

docker ps -a --format '{{.Names}}' | grep -q "^${NAME}$" && { echo "残留容器, 删除"; docker rm -f $NAME; }

ss -tlnp | grep -q ":${PORT} " && { echo "端口 ${PORT} 被占用!"; exit 1; }

free -h | head -2



echo ">>> 启动 Qwen3.8-27B-NVFP4 + DSpark via SGLang on :${PORT}"

docker run -d --name $NAME \

  --gpus all --ipc=host --shm-size 32g \

  -p ${PORT}:30000 \

  -v ${MODEL_DIR}:/models/Qwen3.8-27B-NVFP4:ro \

  -v ${DRAFT_DIR}:/models/Qwen3.8-27B-DSpark:ro \

  ${IMAGE} \

  python3 -m sglang.launch_server \

  --model-path /models/Qwen3.8-27B-NVFP4 \

  --served-model-name Qwen3.8-27B-NVFP4 \

  --host 0.0.0.0 --port 30000 \

  --dtype bfloat16 \

  --context-length 262144 \

  --mem-fraction-static 0.88 \

  --max-running-requests 2 \

  --kv-cache-dtype fp8_e4m3 \

  --attention-backend flashinfer \

  --mamba-ssm-dtype bfloat16 \

  --mamba-radix-cache-strategy extra_buffer \

  --disable-prefill-cuda-graph \

  --chunked-prefill-size 8192 \

  --reasoning-parser qwen3 --tool-call-parser qwen3_coder \

  --enable-cache-report \

  --trust-remote-code \

  ${SPEC_ARGS}



echo ">>> 监控日志 (等 server ready)"

for i in $(seq 1 90); do

  sleep 10

  docker logs $NAME 2>&1 | grep -qE "server is ready|fired up" && break

done

docker logs $NAME 2>&1 | tail -20



echo ">>> 验证"

curl -s http://localhost:${PORT}/v1/models | head -c 300

echo

echo ">>> 推理验证"

curl -s http://localhost:${PORT}/v1/chat/completions \

  -H 'Content-Type: application/json' \

  -d '{"model":"Qwen3.8-27B-NVFP4","messages":[{"role":"user","content":"Count 1 to 30, one per line"}],"max_tokens":60,"chat_template_kwargs":{"enable_thinking":false}}' | python3 -m json.tool | head -25