Skip to content

Clarification on Batch Size Definition #20

Description

@zaxliu

Hi DeepSeek Team, thanks for the great work on DeepSeek-V3 and the detailed deployment profile.

I have a clarification question regarding PD (Prefill–Decode) separation during deployment.
When you mention the decoding batch size, does this refer to:

  1. The number of active sequences being decoded concurrently in the system (i.e., batch size at the sequence/request level), or
  2. The number of tokens processed by experts in the decoding stage (i.e., token-level batching inside the MoE/expert execution)?

Understanding this distinction would be very helpful for correctly interpreting the performance analysis and for reproducing similar deployment settings.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions