[object Object]

MoE models support expert parallelism (EP), which deploys experts on different devices to implement expert-level parallel computing.

Currently, two EP forms are implemented:

1. EP based on AllGather communication ([object Object])

2. EP based on AllToAll and communication-computing fusion ([object Object])

[object Object]

• The DeepSeek-V2, DeepSeek-V3, and DeepSeek-R1 models support this feature.
• If the number of parallel experts exceeds 32, DeepSeek-V3 and DeepSeek-R1 automatically enable the grouped matmul fused operator to improve computing performance.

[object Object]

describes the serving parameters required for enabling the Expert Parallel feature.

Table 1 Expert Parallel parameters in [object Object] of [object Object] [object Object][object Object]

[object Object]undefined[object Object]

Example when [object Object] is set to [object Object]:

[object Object]

Example when [object Object] is set to [object Object] in the long sequence scenario:

[object Object]

1. Set serving parameters. This feature must be used together with MindIE Motor. Add the corresponding parameters to the serving [object Object] file based on . For details about the serving parameters, see .
2. Start the service. For details, see "Quick Start" > "" in MindIE Motor Developer Guide.