If dp_size = tp_size is still required for deepseek model?

[luzengxiangcn]

According to deepseek optimization, the deepseek model with dp+tp attention is now supported. If dp_size = tp_size is still required？

In server argument doc:

enable_dp_attention: Enable [Data Parallelism Attention](https://lmsys.org/blog/2024-12-04-sglang-v0-4/#data-parallelism-attention-for-deepseek-models) for Deepseek models. 
Note that you need to choose dp_size = tp_size for this.

[zhaochenyang20]

Well. I think still need this. But we've automiatically done this.

[luzengxiangcn]

cc @zhaochenyang20. Does it mean dp or tp is supported, and i can't use it in same time, eg tp=4, dp=2. We need this feature to deploy deepseek-v3 across multiple nodes. TP in same node, dp cross all nodes.

[zhaochenyang20]

The dp in data parallelism attention has different meaning. Check this:

https://docs.sglang.ai/references/deepseek.html#multi-head-latent-attention-mla-throughput-optimizations

[luzengxiangcn]

@zhaochenyang20

The dp in data parallelism attention has different meaning. Check this:

https://docs.sglang.ai/references/deepseek.html#multi-head-latent-attention-mla-throughput-optimizations

Gocha！
What I am looking for is：

We are planning to apply this deployment architecture to reduce network pressure between nodes, while not increasing inference time too much when throughput is low.

[zhaochenyang20]

@luzengxiangcn Cool. Nice work. Hope to see it soon.