opt ll dispatch layered algo

[alpha-baby]

introduce

algo opt for dispatch in low-latency mode:

In the dispatch kernel of DeepEP's low-latency mode, the original algorithm directly sends data to the destination rank via the RDMA cross-orbit network. A drawback of this algorithm is that it results in excessive duplicate data being transmitted over the RDMA network. Now, drawing inspiration from the approach used in normal mode, we can improve the dispatch kernel in low-latency mode by first sending data to the same-orbit rank on the cross-node, and then forwarding it to the actual destination rank via the NVLink interconnect.

Note: This feature conflicts with the existing Elasticity Support to DeepEP for Fault-Tolerant EP Inference functionality, and the two features cannot be enabled simultaneously.

before:

after:

performance

benchmark:

use

This feature is enabled by default and requires no additional activation from the user. To disable it, please set the following environment variable: DEEPEP_DISABLE_LL_DISPATCH_OPT=1.

[alpha-baby]

benchmark test pass on env: one/two/four 8*H200,

[wangfakang]

LGTM. Thanks.

[ywj55555]

Hello, I'd like to ask why similar optimizations weren't made to combine?