kvcache-ai/Mooncake

[TE] Refactor: unify libfabric-based transports (EFA / CXI) shared infrastructure

开放

#2,657 创建于 2026年6月29日

 (3 条评论) (0 个反应) (0 位负责人)C++ (803 个派生)auto 404
good first issue

仓库指标

星标
 (5,470 个星标)
PR 合并指标
 (平均合并 4天 6小时) (30 天内合并 224 个 PR)

描述

Background

With #2535 (CXI backend) merged, we now have two transports built on libfabric that share significant structural overlap but are implemented as independent forks. This was the pragmatic choice for initial integration (see discussion in #2535), but leaves maintenance debt.

Scope

  1. Topology — CXI currently reuses InfiniBand device discovery paths in topology.cpp. CXI is not IB; it should have its own topology functions or a shared libfabric_topology abstraction (ref: @alogfans' comment).

  2. Shared base class or utilities — Extract common libfabric patterns (endpoint management, CQ polling, MR registration, error handling) into a shared layer that both EFA and CXI can consume, reducing duplicated code.

  3. mr_key_t consolidation — The current #if defined(USE_EFA) || defined(USE_CXI) guards work but will grow unwieldy if more libfabric providers are added. Consider a compile-time or runtime abstraction.

Non-goals

  • Changing the public API or metadata wire format
  • Merging EFA and CXI into a single transport (they have meaningful behavioral differences in MR handling and transfer semantics)

References

  • #2535 — CXI backend PR
  • #2564 — 64-bit MR key fix

贡献者指南