8000 End-to-end serving evaluation · Issue #11 · afdEP/afdEP-comm · GitHub
[go: up one dir, main page]

Skip to content

End-to-end serving evaluation #11

Open

Description

@Etelis

Every afdEP result so far is communication-only: one dispatch plus combine round trip under CUDA graphs. Nothing yet shows that the lower round trip becomes lower inter-token latency or higher throughput in serving.

Prerequisites

Tasks

  • A serving integration on v1's API, such as an afd-plugin connector for vLLM. SGLang's AFD path on sunfanyi/DeepEP serves as a comparison.
  • TPOT/ITL and throughput against Masked DeepEP and against the transport the stack uses today, at decode-heavy batch sizes and several Attention:FFN ratios (3:1, 7:1, 11:1), in one node and across nodes.
  • Profile where communication sits on the critical path: its overlap with attention and expert GEMMs, and SM contention under load.
  • Publish a record and a README section labeled as serving results.

Activity

  1. added
    enhancementNew feature or request
    servingIntegration with serving stacks and end-to-end results
    on Oct 23, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or requestservingIntegration with serving stacks and end-to-end results

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions

      0