为什么4090单卡可以进行推理，而多卡会爆显存呢（More than 4090 cards burst video memory）

[3214894586]

我使用4090进行单GPU推理并用qwen-3b拓展提示词，正常生成了视频。但是4张4090使用提示词拓展就爆显存了。这是什么问题？明明显存多了那么多，求解答！
I used the 4090 for single GPU inference and extended the prompt words with qwen-3b, and the video was generated normally. But 4 4090 cards use prompt words to expand and explode the video storage. What is the problem? Obviously, there are so many savings, ask for answers!
40901：

40904：

[yizhixu]

现在还没有开源模型并行的代码吧？
数据并行的话多卡通信有显存开销的