afdEP /
afdEP-comm
Expert-parallel communication for Attention–FFN Disaggregation (AFD), built on DeepEP V2
Updated 43 min ago
+27 stars gained this week
Stars gained each day, last 30 days
Oct 4Oct 19Today
This week Before
In short
Needs NVIDIA Hopper GPUs with NVLink or RDMA and a CUDA build of DeepEP V2 kernels, run across N Attention and M FFN GPUs. Use it as a Python library inside an AFD inference engine; routing, quantization and expert compute stay with the caller.
Summary written by AI from the repository’s description and README.
afdEP / afdEP-comm
View on GitHub(opens in new tab)