Open Models Almanac
Methodsinference & servingserving parallelism

implementation detail · filed under inference & serving

DeepEP

DeepEP is DeepSeek's library of all-to-all kernels that send each token to the GPUs holding its experts and gather the results back. Its first version had a high-throughput mode for training and prefill and a low-latency, CUDA-graph-capturable mode for decode; version 2 merged them into one interface, and version 2.5 removed the first.

Written for the explainer Kernels and graph compilation: where the cost model meets the GPU.

Also called --moe-a2a-backend deepep, DeepEP expert parallelism across nodes, DeepEP for expert parallelism.

sources
4
models
4
lab adopt it
1
strongest
used

Documented in

Evidence

4 quotes from 4 sources, one card per model, each opening with its strongest.

How sources treat it
optional 2used 2

Further reading

Picked by hand, not extracted: where to read more, not evidence for anything on this page.

↑