Skip to content

feat(server): support native CUDA IPC clients#407

Open
xiaguan wants to merge 1 commit into
masterfrom
feat/rust-cuda-ipc-client-support
Open

feat(server): support native CUDA IPC clients#407
xiaguan wants to merge 1 commit into
masterfrom
feat/rust-cuda-ipc-client-support

Conversation

@xiaguan

@xiaguan xiaguan commented Jul 19, 2026

Copy link
Copy Markdown
Collaborator

Summary

  • Accept native CUDA IPC registrations as one CudaIpcTensor entry per layer, including allocation handle, view range, and block stride.
  • Reuse the existing strided registration, in-process load completion, and flush APIs; the Python wrapper path remains unchanged.
  • Reject old/new wire mismatches through an exact native-CUDA-IPC capability version.

Validation

  • prek run
  • cargo test --release -p pegaflow-server --no-default-features --features cuda-13 --lib (23 passed)
  • cargo test --release -p pegaflow-core --lib (124 passed, 1 ignored)
  • cargo check --release -p pegaflow-server --no-default-features --features cuda-13 --all-targets
  • cargo check --release -p pegaflow-py --no-default-features --features cuda-13

@xiaguan
xiaguan force-pushed the feat/rust-cuda-ipc-client-support branch 2 times, most recently from 5b73e5b to 6173278 Compare July 20, 2026 07:36
@xiaguan
xiaguan changed the base branch from master to fix/drain-before-unregister July 20, 2026 07:37
@xiaguan
xiaguan force-pushed the feat/rust-cuda-ipc-client-support branch 2 times, most recently from 2892cba to 933f548 Compare July 20, 2026 07:52
@xiaguan
xiaguan changed the base branch from fix/drain-before-unregister to master July 20, 2026 08:24
@xiaguan xiaguan closed this Jul 20, 2026
@xiaguan xiaguan reopened this Jul 20, 2026
Signed-off-by: xiaguan <751080330@qq.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant