Open Models Almanac
Methodssoftware implementationkernel & quantization library

specific method · filed under software implementation

MegaMoE

MegaMoE, in DeepSeek's DeepGEMM library, runs an MoE layer's dispatch, both expert matrix products, the activation and the combine as one pipelined kernel that overlaps the exchange of tokens between GPUs with the expert arithmetic. It needs expert parallelism over NVLink with symmetric memory on every GPU, and each variant supports specific number formats.

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

Also called MegaMoE expert-parallel serving backend, MegaMoE MoE backend, MegaMoE MoE execution backend.

sources
2
model
1
lab adopt it
1
strongest
used

Documented in

Evidence

3 quotes from 2 sources, one card per model, each opening with its strongest.

How sources treat it
optional 2used 1

Step-5-Preview

StepFunadopted2 sources

This configuration uses MegaMoE for the FP8 routed experts.

usedinference servingin Step 5 PreviewStepFun
2 more from this model

On B300 (Blackwell), the TEP4 and DEP4 commands select MegaMoE with `--moe-backend deep_gemm_mega_moe`.

optionalinference servingin Step 5 PreviewStepFun

the TEP4 and DEP4 commands select MegaMoE with --moe-backend deep_gemm_mega_moe

optionalinference servingin Step 5 PreviewStepFun
↑