[Roadmap]: NVSHMEM Roadmap Q3 2026

[gab9talavera]

NVSHMEM Roadmap (Q3 2026)

This issue tracks planned NVSHMEM development and releases for Q3 2026.
Plans are subject to change as the team iterates, completes validation, and receives feedback.
If you have suggestions for features, please open a feature request or comment below.

---

Recently Released: NVSHMEM 3.7.0-0

Release Highlights

• TMA-backed NVLink put / get: Added TMA-backed implementations for NVLink put and get operations, including support for global-memory and shared-memory local buffers plus shared-memory registration APIs.
• GPUNetIO remote transport: Added GPUNetIO support with GPU-initiated communication and DOCA SDK integration.
• Flush APIs: Added nvshmemx_flush APIs to support source-buffer reusability without requiring remote visibility.
• Experimental logical endpoint / CFT support: Added experimental logical endpoint and CFT handle support for fabric-PTX unicast communication.
• Floating-point atomics: Added floating-point atomic add and fetch-add APIs with P2P and proxy-backed IBRC support.
• OpenSHMEM bootstrap improvements: Added teams-based OpenSHMEM bootstrap support using SHMEM_TEAM_WORLD, with fallback to legacy active-set collectives.
• InfiniBand routing and reliability controls: Added GID-based routing for IBRC and IBGDA, plus PKey index, QP ack timeout, and retry-count controls.
• Build, packaging, and diagnostics: Added Python wheel CMake target support, multi-architecture device libraries, C++17 as the minimum C++ version, Apache-2.0 licensing, and improved status strings in error logs.
• NVSHMEM4Py 0.3.1: Updated Numbast integration, improved Python / Numba-CUDA dependency handling, removed a hardcoded CUDA 13 build requirement, and fixed CuTe / Numba device collective generation issues.
• External contributions: Added NUMA-aware CPU affinity, NIC assignment policy controls, libfabric RMA batching support, and EFA-focused libfabric progress and GDRCopy improvements.

For the full release notes, see: NVSHMEM 3.7.0 Release

---

Q3 Roadmap - Aug '26-Oct '26

• CFT support: Complete Compute Fabric Transport enablement and validation. CFT is a handle-based GPU fabric communication model that uses logical endpoints plus offsets instead of raw pointers, mainly to scale better and improve error handling in large multi-GPU systems.
• AI skills library: Introduce AI skills to make NVSHMEM easier to use in AI and accelerated-computing workflows.
• NVSHMEM4Py for numba-cuda-mlir: Add NVSHMEM4Py support for numba-cuda-mlir, NVIDIA’s MLIR-based CUDA Python kernel compiler, so Python and compiler-driven workflows can use NVSHMEM more naturally.
• Multi-NIC support for IBRC and IBDevX: Extend Multi-NIC support across IB/RoCE transports, including support for multi-NIC mapping through NVSHMEM_HCA_LIST and NVSHMEM_HCA_PE_MAPPING.
• Batched RMA: Batch RMA operations to enable message coalescing and reduce overhead for workloads with many small operations.

---

Features Under Consideration

• SM‑initiated CE collectives: Explore enabling GPU‑initiated transfers on copy engines to increase overlap between communication and computation.

Contributions, feedback, issue submissions, and discussion are welcome.