8000 [core][gpu-objects] RDMA support for data transfer · Issue #53976 · ray-project/ray · GitHub
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content
[core][gpu-objects] RDMA support for data transfer #53976
New issue
New issue
Open
Open
[core][gpu-objects] RDMA support for data transfer#53976
Labels
P1Issue that should be fixed within a few weekscoreIssues that should be addressed in Ray CoreenhancementRequest for new feature and/or capabilitygpuGPU related issuesgpu-objectsperformance
@stephanie-wang

Description

@stephanie-wang
stephanie-wang
opened on Jun 20, 2025

Description

The initial tensor transport is based on NCCL, but RDMA is better for cases where we only need point-to-point and if the group/world size is dynamic.

There are a few options for RDMA that we plan to look into:

  • NVIDIA NIXL
  • UCX
  • a native solution

Use case

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    P1Issue that should be fixed within a few weekscoreIssues that should be addressed in Ray CoreenhancementRequest for new feature and/or capabilitygpuGPU related issuesgpu-objectsperformance

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions
      0