openucx/ucx v1.22.0

on GitHub

1.22.0 (August 2, 2026)

Features:

UCP

• Added SGL datatype support for non-blocking PUT operations
• Added GET and PUT rendezvous protocols for RMA operations
• Added fault-tolerance recovery foundation
• Added AM, zcopy, multi, and PSN protocol failover support
• Added endpoint flush failover and first-fragment retransmission support
• Added multi-lane support for device operations
• Added memory registration flags to UCP/UCT/UCS APIs
• Enabled v2 operations on proxy and wireup endpoints
• Added transport and wireup debugging table dumps during context and endpoint creation
• Set four rails by default for Vera platforms

UCT

• Added UCT v2 capability flags for SGL zcopy operations
• Enhanced endpoint connectivity checks

RDMA CORE (IB, ROCE, etc.)

• Added plugin infrastructure for IB token query and plugin-provided RC extended operations
• Added per-endpoint RC transmit queue reservation for resource checks
• Added IB/MLX5 initiator small-fence WQE control flag
• Applied ECE returned by ibv_query_ece by default
• Enabled relaxed ordering by default for Vera CPU platforms

CUDA

• Added GDRCopy PCIe BAR1 export option
• Added POSIX file descriptor handle support for same-machine CUDA VMM IPC
• Added support for CUDA device aliasing with MPS MLOPart

UCS

• Added table builder debug API
• Added ucs_string_buffer_vappendf utility
• Added numeric range support for device configuration
• Added Vera CPU support

UCM

• Propagated memory attributes in allocation events

Tools

• Added software GET responder progress support in perftest

Build

• Added PR build modes
• Added Ubuntu 26.04, TencentOS 4.4, and GCC 15.2 build coverage
• Added GB300/aarch64 DL cluster test pipeline coverage
• Added parallel static checks
• Enabled on-demand DRP testing for release pipelines
• Added configure option to append a SONAME suffix

Bugfixes:

UCP

• Fixed endpoint fault-tolerance discard and failover flows
• Fixed fallback handling when no more lanes are available
• Fixed AM reply endpoint lookup during initialization
• Fixed memory release crash in active-message path
• Fixed rcache validity check before memory handle invalidation
• Fixed context creation when only accelerator devices are available
• Fixed error reporting when the transport/device resource limit is exceeded
• Fixed RMA rendezvous rkey size assertion
• Fixed bandwidth formatting macros
• Fixed single network-device filtering to use minimum distance
• Fixed protocol assertions and short selection diagnostics
• Fixed zero-length memory handle packing in trace logging
• Fixed rendezvous remote-memory flag estimation from rkey memory-domain maps
• Disabled RMA rendezvous protocols on non-Vera platforms
• Fixed active-message lane selection to preserve latency priority
• Fixed rendezvous multi-lane protocol selection to preserve combined minimum sizes
• Fixed unnecessary rendezvous PUT fences on ordered transports

ZE

• Fixed ZE driver enumeration, PCI fallback, IOV bounds, and DMA-BUF memory query semantics
• Fixed ZE memory query system-device detection

CUDA

• Fixed CUDA managed-memory Valgrind failures
• Fixed CUDA IPC remote-cache destruction during endpoint destroy
• Fixed CUDA cleanup for destroyed contexts
• Fixed CUDA IPC single-process unmap checks
• Fixed CUDA NVLink detection
• Fixed CUDA IPC remote-cache handling for GET and PUT paths
• Fixed CUDA IPC reachability for same-process interfaces
• Fixed CUDA copy memory-flag detection with active CUDA contexts

RDMA CORE (IB, ROCE, etc.)

• Fixed GGA GET zcopy purge completion
• Fixed GDA DMA-BUF offsets and on-demand DMA-BUF checks
• Fixed GDAKI Direct NIC matrix selection
• Fixed IB DEVX handling for UARs without WC support
• Fixed RoCE reachability for RTN_LOCAL routes
• Reverted ARM Neon BlueFlame writes in IB transport
• Fixed MLX5 DDP PUT fencing
• Fixed IB port-speed query logging for unsupported devices
• Fixed Direct NIC sibling matching when HCAs have no active ports
• Fixed DONT_FORK handling for kernels that support RDMA copy-on-fork
• Fixed GGA MMO fence ordering
• Restricted EMR relaxed ordering to device memory

Shared Memory

• Fixed POSIX reachability failure logging in containers

TCP

• Fixed TCP negative connect-test timeout

Packaging

• Fixed GDA RPM/DEB packaging and RPM development package builds

UCS

• Fixed stale async file descriptor event filtering
• Fixed IPv6 scope formatting
• Fixed time initialization to be multi-thread safe
• Fixed namespace identifier sizing for portability
• Fixed portability issues for non-glibc and clang environments

Tools

• Fixed perftest ZE allocator thread safety and device routing
• Fixed I/O demo data-size range parsing

Build

• Fixed GPUNetIO update failure reporting
• Fixed CMake policy version for examples tests
• Replaced ofed_info usage with native package-manager queries
• Fixed CI build result reporting on exceptions
• Excluded .ci/ changes from PR pipeline triggers

Known issues:

• JUCX package for this release was renamed to
  1.22.0-1, please avoid using 1.22.0.