H20部署DeepSeek-V4-Flash-0731实战：FP8 KV Cache、DSpark投机解码与专家并行调优

硬件层面

模型权重

模型权重：deepseek-ai/DeepSeek-V4-Flash-0731 · Hugging Face

vllm镜像：vllm-openai:v0.26.0-cu129-ubuntu2404

sglang镜像：sglang:v0.5.16-cu130

vllm启动脚本

sudo docker run -itd \
  --restart=always \
  --name vllm-deepseek-v4-flash-0731 \
  --gpus all \
  --shm-size=32g \
  -p 8000:8000 \
  -v /data1/models/DeepSeek-V4-Flash-0731:/models/DeepSeek-V4-Flash-0731 \
  -e VLLM_ENGINE_READY_TIMEOUT_S=3600 \
  harbor-cmp.zoomlion.com/library/vllm/vllm-openai:v0.26.0-cu129-ubuntu2404 /models/DeepSeek-V4-Flash-0731 \
  --served-model-name deepseek-v4-flash \
  --trust-remote-code \
  --kv-cache-dtype fp8 \
  --block-size 256 \
  --enable-expert-parallel \
  --tensor-parallel-size 8 \
  --tokenizer-mode deepseek_v4 \
  --tool-call-parser deepseek_v4 \
  --enable-auto-tool-choice \
  --reasoning-parser deepseek_v4 \
  --host 0.0.0.0 \
  --max-model-len 131072 \
  --api-key 4312f57efa11b812a0a5a055058bc469

vllm调优配置

--kv-cache-dtype fp8_ds_mla \
--max-model-len 262144 \
--no-enable-flashinfer-autotune \
--compilation-config '{"mode":0,"cudagraph_mode":"FULL_DECODE_ONLY"}' \
--speculative-config '{"method":"dspark","num_speculative_tokens":5,"draft_sample_method":"greedy"}'

sglang启动脚本

sudo docker run -itd \
  --restart=always \
  --name vllm-deepseek-v4-flash-0731 \
  --gpus all \
  --shm-size=32g \
  -p 8000:8000 \
  -v /data1/models/DeepSeek-V4-Flash-0731:/models/DeepSeek-V4-Flash-0731 \
  -e VLLM_ENGINE_READY_TIMEOUT_S=3600 \
  harbor-cmp.zoomlion.com/library/vllm/vllm-openai:v0.26.0-cu129-ubuntu2404 /models/DeepSeek-V4-Flash-0731 \
  --served-model-name deepseek-v4-flash \
  --trust-remote-code \
  --kv-cache-dtype fp8 \
  --block-size 256 \
  --enable-expert-parallel \
  --tensor-parallel-size 8 \
  --tokenizer-mode deepseek_v4 \
  --tool-call-parser deepseek_v4 \
  --enable-auto-tool-choice \
  --reasoning-parser deepseek_v4 \
  --host 0.0.0.0 \
  --max-model-len 131072 \
  --api-key 4312f57efa11b812a0a5a055058bc469