Repository navigation

Issues

New issue

What is UCX's policy of choosing transport?
#5320 · rzambre opened on Jun 23, 2020
2
Error in mlx5dv_create_qp in the DC transport
#5749 · lyu opened on Sep 29, 2020
34
Error: Transport retry count exceeded on mlx5_0:1/RoCE
#6000 · afernandezody opened on Dec 8, 2020
9

ucp_ep_rkey_unpack() fails the whole rkey when its sysv segment cannot be attached
openucx/ucx#12113 · opened on Oct 18, 2026
ucp_put_bw with CUDA device API hangs with 8 KiB and many channels on CX8
openucx/ucx#12103 · opened on Oct 16, 2026
UCT/CUDA_IPC: uct_cuda_ipc_mkey_pack leaves struct padding in the packed rkey uninitialized
openucx/ucx#12099 · opened on Oct 15, 2026
UCP/UCT: stack buffer overflow during initialization when SR-IOV VF changes increase IB MD resource count
openucx/ucx#12079 · opened on Oct 13, 2026
2
ucp_init() can deadlock on the GLIBC loader lock in cuInit()
openucx/ucx#12078 · opened on Oct 13, 2026
UCS/NETLINK: route-dump reader blocks forever when NLMSG_DONE shares a datagram with the last route (hangs ucp_ep_create)
openucx/ucx#12070 · opened on Oct 10, 2026
UCP/EP: failed ucp_ep_create leaks the lanes connected before the failing lane
openucx/ucx#12021 · opened on Oct 3, 2026
UCP/EP: flush-mode ucp_ep_close_nbx keeps worker-address endpoints to live peers until ucp_worker_destroy
openucx/ucx#12020 · opened on Oct 3, 2026
make dist silently produces an empty tarball when uid/gid exceeds the ustar limit
openucx/ucx#12018 · opened on Oct 2, 2026
CUDA IPC-imported pointers (cuIpcOpenMemHandle) are not tracked by the memtype cache: host memcpy into device memory (SIGSEGV) on eager, and no fallback when registration fails on rendezvous
openucx/ucx#12013 · opened on Oct 1, 2026
Intra-Node Performance Degradation (due to multirail?)
openucx/ucx#11986 · opened on Sep 24, 2026
CUDA IPC: support VMM POSIX FD sharing across PID namespaces
openucx/ucx#11978 · opened on Sep 22, 2026
rc_verbs on RoCE LAG bond: 23 GB/s → 2.9 MB/s after #11262 (ZTR-RTT CC enabled by default)
openucx/ucx#11955 · opened on Sep 16, 2026
TCP: cannot be selected as an RMA BW lane with peer error handling, so ucp_put falls back to AM emulation
openucx/ucx#11929 · opened on Sep 8, 2026
1
Segfault when operating on HIP asynchronously allocated memory on ROCm 10.0.0
openucx/ucx#11898 · opened on Sep 4, 2026
uct_iface_open failure on unusable RDMA device aborts entire worker instead of skipping
openucx/ucx#11820 · opened on Aug 24, 2026
UCP/WIREUP: are misaligned typed accesses in packed addresses intentional?
openucx/ucx#11806 · opened on Aug 20, 2026
cuda_ipc: imported remote mappings are never released until importer process exit (no per-peer eviction) — pins exporter VMM/fabric memory domain-wide
openucx/ucx#11753 · opened on Aug 11, 2026
11
UCT/IB/MLX5: Should rc_mlx5 support bounded batch CQ draining?
openucx/ucx#11729 · opened on Aug 5, 2026
UCT/IB/MLX5: Non-DevX rc_mlx5 CQ notification misses after Direct Verbs polling
openucx/ucx#11728 · opened on Aug 5, 2026
2
UMR QP INIT2RTR fails on LAG-bonded devices, causing segfault on memory registration
openucx/ucx#11709 · opened on Jul 31, 2026
#11547 concentrates multi-rank GPU RMA traffic onto a subset of NICs (4 of 8), causing severe oversubscription and multi-× bandwidth loss
openucx/ucx#11680 · opened on Jul 20, 2026
2
cuda_copy transport in UCX_TLS degrades LLM inference throughput
openucx/ucx#11652 · opened on Jul 14, 2026
3
Version 1.21.0 does not build without gpunetio
openucx/ucx#11591 · opened on Jun 25, 2026
Feature request: allow UCP AMO offload for partially capable UCT atomic masks
openucx/ucx#11572 · opened on Jun 23, 2026

Navigation Menu

Repository navigation

Issues

Issues

ucp_ep_rkey_unpack() fails the whole rkey when its sysv segment cannot be attached

ucp_put_bw with CUDA device API hangs with 8 KiB and many channels on CX8

UCT/CUDA_IPC: `uct_cuda_ipc_mkey_pack` leaves struct padding in the packed rkey uninitialized

UCP/UCT: stack buffer overflow during initialization when SR-IOV VF changes increase IB MD resource count

ucp_init() can deadlock on the GLIBC loader lock in cuInit()

UCS/NETLINK: route-dump reader blocks forever when NLMSG_DONE shares a datagram with the last route (hangs ucp_ep_create)

UCP/EP: failed ucp_ep_create leaks the lanes connected before the failing lane

UCP/EP: flush-mode ucp_ep_close_nbx keeps worker-address endpoints to live peers until ucp_worker_destroy

`make dist` silently produces an empty tarball when uid/gid exceeds the ustar limit

CUDA IPC-imported pointers (`cuIpcOpenMemHandle`) are not tracked by the memtype cache: host `memcpy` into device memory (SIGSEGV) on eager, and no fallback when registration fails on rendezvous

Intra-Node Performance Degradation (due to multirail?)

CUDA IPC: support VMM POSIX FD sharing across PID namespaces

rc_verbs on RoCE LAG bond: 23 GB/s → 2.9 MB/s after #11262 (ZTR-RTT CC enabled by default)

TCP: cannot be selected as an RMA BW lane with peer error handling, so ucp_put falls back to AM emulation

Segfault when operating on HIP asynchronously allocated memory on ROCm 10.0.0

`uct_iface_open` failure on unusable RDMA device aborts entire worker instead of skipping

UCP/WIREUP: are misaligned typed accesses in packed addresses intentional?

cuda_ipc: imported remote mappings are never released until importer process exit (no per-peer eviction) — pins exporter VMM/fabric memory domain-wide

UCT/IB/MLX5: Should rc_mlx5 support bounded batch CQ draining?

UCT/IB/MLX5: Non-DevX rc_mlx5 CQ notification misses after Direct Verbs polling

UMR QP INIT2RTR fails on LAG-bonded devices, causing segfault on memory registration

#11547 concentrates multi-rank GPU RMA traffic onto a subset of NICs (4 of 8), causing severe oversubscription and multi-× bandwidth loss

`cuda_copy` transport in `UCX_TLS` degrades LLM inference throughput

Version 1.21.0 does not build without gpunetio

Feature request: allow UCP AMO offload for partially capable UCT atomic masks