官方帖子:https://developer.sourcefind.cn/codes/modelzoo/deepseek-v4-int
Deepseek-V4-Flash-0731模型地址:https://modelscope.cn/models/hygon/DeepSeek-V4-Flash-0731-Channel-INT8-w8a8
Docker启动镜像容器:
```
docker run -it \
--shm-size 256g \
--network=host \
--name deepseek-v4-flash-int \
--privileged \
--device=/dev/kfd \
--device=/dev/dri \
--device=/dev/mkfd \
--group-add video \
--cap-add=SYS_PTRACE \
--security-opt seccomp=unconfined \
-u root \
-v /opt/hyhal/:/opt/hyhal/:ro \
-v /path/to/DeepSeek-V4-Flash-INT8/:/models/:ro \
-v /path/to/deepseek-v4-int/:/workspace/deepseek-v4-int/ \
-w /workspace/deepseek-v4-int \
harbor.sourcefind.cn:5443/dcu/admin/base/custom:dsv4-flash-k100ai-sglang0.5.12-20260728 bash
```
#注意: 如果需要支持工具调用和推理。官方仓库中的chat_template.jinja是有问题的。 建议替换start_sglang.sh脚本中sglang的启动参数如下:
sglang_args=(
--model-path "$MODEL_PATH"
--served-model-name "Deepseek-V4-Flash"
--host "$SGLANG_HOST"
--port "$SGLANG_PORT"
--trust-remote-code
--tp 8
--cuda-graph-max-bs 128
--chunked-prefill-size 8192
--quantization compressed-tensors
--kv-cache-dtype bfloat16
--max-running-requests 128
--reasoning-parser deepseek-v4
--tool-call-parser deepseekv4
--disable-radix-cache
--disable-flashinfer-autotune
)
启动命令:SERVED_MODEL_NAME=DeepSeek-V4-Flash ENABLE_MTP=true SPECULATIVE_NUM_STEPS=3 bash start_sglang.sh
运行后可调试:
curl http://localhost:30001/v1/chat/completions
-H "Content-Type: application/json"
-d '{
"model": "DeepSeek-V4-Flash",
"messages": [{"role": "user", "content": "你的知识更新截止时间?"}],
"chat_template_kwargs": {"thinking": true, "reasoning_effort": "max"}
}'
实测,平均推理速度:25 TPS (当前实测速度10TPS,待调参更新)
目前模型参数中包含DSpark参数,建议更新Docker镜像源