Skip to content

PR Snapshot 2026 09 23

Tomislav Janjusic edited this page Sep 23, 2026 · 4 revisions

PRs Snapshot 2026-09-23 — open-mpi/ompi Open Pull Requests

Snapshot date: 2026-09-23

Total open PRs: 92

Legend

All PRs below are open/unmerged — "merged" in any column always refers to a related PR, never the row itself.

Field Meaning
Target branch (section header) PR's actual base branch — authoritative
Target:/target: label Auto-label mirroring base branch (redundant, just two casings)
backport: vX.Y.x label Intent to backport after merge — not yet done
Created (age) Open date; age = days/months/years since snapshot
LOC (+/-) Lines added / deleted
Merge/Backport Status Source PR + sibling-backport state, or why it's standalone
Review Status ⏳/✅/🔴 decision, who approved/requested changes/is awaiting review
Draft/WIP Split into its own section at bottom

Open PRs (74)

Target branch: main (65 open PRs)

# Title Author Created (age) Labels Merge/Backport Status LOC (+/-) Review Status Description
#14499 Tell a bridged intercomm's peers which of them share our node rhc54 2026-09-23 (0d) Target: main - +154 / -26 ⏳ Review required — no reviewer assigned, no reviews yet MPI_Intercomm_create between two groups that have never been connected (for example two jobs spawned by one parent and bridged through it) has the two leaders swap their sides' procs with ompi_proc_pack()/`ompi_pr...
#14498 3rd-party: update embedded hwloc from 2.7.1 to 2.15.0 edgargabriel 2026-09-23 (0d) Target: main, backport: v6.0.x Not yet backported - pending merge to main first. +22 / -2 ⏳ Review required — awaiting: @hppritcha Bump the embedded hwloc tarball shipped in 3rd-party/ to the 2.15.0 release. The build system derives the tarball and directory names from the $hwloc_version constant in autogen.pl, and the documentation derives the e...
#14496 CI: Test ntsemah 2026-09-23 (0d) Target: main - +1 / -0 ✅ Approved — approved by janjust no description
#14495 Add LL/SC support for RISC-V platforms devreal 2026-09-23 (0d) Target: main - +332 / -19 ⏳ Review required — awaiting: @jsquyres - Introduces explicit acquire-ll, linked load with acquire semantics - Adds implementation of ll/sc for 64bit RISC-V platforms. Note: The RISC-V "A" extension spec (Volume I, §A.6 "Eventual Success of Store-Conditiona...
#14494 Add memory clobber to opal_atomic_ll for arm and PPC devreal 2026-09-22 (1d) Target: main, backport: v6.0.x Not yet backported - pending merge to main first. +15 / -14 ⏳ Review required — awaiting: @jsquyres This was flagged while working on https://github.com/open-mpi/ompi/pull/13969. That PR has become unmanageable for me so I'm splitting out the noncontroversial parts here. More to come as time permits.
#14493 osc/ucx: fence between endpoint close and worker destroy at finalize bwestheimer 2026-09-22 (1d) Target: main - +85 / -12 ⏳ Review required — no reviewer assigned, no reviews yet ompi/mca/osc/ucx/osc_ucx_component.c:component_finalize() closes the UCP endpoints and then destroys the UCP workers back to back, with nothing synchronizing the ranks in between.
#14492 comm_revoke_local: Check for null before revoking c_coll Matthew-Whitlock 2026-09-22 (1d) Target: main - +2 / -2 ⏳ Review required — no reviewer assigned, no reviews yet There's a window during communicator creation that a failure can lead to the communicator being constructed being revoked before it has been assigned a coll component in c_coll. Open MPI then segfaults. This fix just ...
#14474 Rename --enable-standard-abi to --enable-forum-abi, and use "MPI Forum ABI" consistently jsquyres 2026-09-21 (2d) Target: main, backport: v6.0.x Not yet backported - pending merge to main first. +160 / -146 ⏳ Review required — commented only: hppritcha, jsquyres Follow-on to #14467 (the ompi-forum-abi*.pc rename). The docs settled on "the Open MPI ABI" and "the MPI Forum ABI" as the names of Open MPI's two ABIs, but the configure option, the installed header directory, and ...
#14472 DOCS: Document OFI domain selection and LINKx usage rakumarp 2026-09-21 (2d) Target: main - +152 / -3 ⏳ Review required — commented only: hppritcha, rakumarp These changes improve Open MPI documentation for OpenFabrics Interfaces and Libfabric usage. The existing documentation did not explain how Open MPI selects Libfabric domains for MPI processes, or how users can config...
#14470 opal/util: strip trailing separators in opal_os_dirpath_create() tonyskjellum 2026-09-20 (3d) Target: main - +93 / -7 ⏳ Review required — commented only: hppritcha Fixes #14469. opal_os_dirpath_destroy() was reworked in #14257 to ignore trailing separators before it opens anything, because a trailing separator makes the kernel resolve the final component as a directory and fol...
#14460 DOCS: add blurb about creative commons hppritcha 2026-09-17 (6d) Target: main - +38 / -1 ⏳ Review required — awaiting: @jsquyres this is need for HPSF activity
#14458 common/ofi: select a device near the accelerator by default yinliaws 2026-09-17 (6d) Target: main - +48 / -4 ⏳ Review required — no reviewer assigned, no reviews yet Device selection already knows how to find the NIC closest to the accelerator a process is using: find_nearest_provider_from_accelerator() compares the PCI locality of each candidate against the accelerator's own and ...
#14449 coll/ucc: pass void pointers to %p yqtian-se 2026-09-16 (7d) Target: main - +17 / -17 ⏳ Review required — no reviewer assigned, no reviews yet UCC_VERBOSE ultimately passes its variadic arguments to the printf-checked opal_output_verbose() function. Seventeen persistent-collective logging calls use %p for an mca_coll_ucc_req_t * argument. The C forma...
#14429 Add Bine tree collective algorithms Dragoned 2026-09-14 (9d) Target: main - +5504 / -234 ⏳ Review required — commented only: Dragoned, hppritcha Add implementations of the BINE (Binomial Negabinary tree) family of collective algorithms, based on "Bine Trees: Enhancing Collective Operations by Optimizing Communication Locality" (De Sensi et al., SC'25, https://...
#14426 Add --disable-ompi-abi configure option for ABI-only installations hppritcha 2026-09-13 (10d) Target: main - +195 / -26 ⏳ Review required — commented only: bwbarrett This commit adds support for a new configuration mode that produces MPI ABI-only installations containing only the standard MPI ABI library (libmpi_abi.so), the standard mpi.h header, and standard wrapper compilers. W...
#14425 btl/tcp: arbitrate an inbound connection from the component bosilca 2026-09-13 (10d) Target: main - +7478 / -1256 ⏳ Review required — no reviewer assigned, no reviews yet Depends on #14377, and through it on #14423 and #14424. The branch is cut from topic/delayed_connections, so this PR necessarily shows all 24 commits of the stack below it — a pull request's base branch has to l...
#14420 topo/base: drop the parent's topology when creating a dist graph metereconsulting 2026-09-11 (12d) Target: main - +21 / -0 ⏳ Review required — no reviewer assigned, no reviews yet A distributed graph communicator created from a Cartesian or graph communicator is treated by Open MPI as the parent's topology type. Open MPI v5.0.10, MPI_Dist_graph_create_adjacent called on a 3-D periodic Cartesi...
#14377 Establish connections lazily: defer the modex wait and add_procs out of MPI_Init bosilca 2026-09-02 (21d) bug, enhancement, Target: main, backport: v5.0.x, backport: v6.0.x Not yet backported - pending merge to main first. +6451 / -1139 ⏳ Review required — commented only: hppritcha Depends on #14423, and through it on #14424. Those two are the base of this stack: the first thirteen commits here are theirs, three from #14424 and ten from #14423. All thirteen are fixes that stand on their own and ...
#14376 oshmem/memheap: validate symmetric memory layouts nbellalou 2026-09-01 (22d) Target: main - +1349 / -219 ⏳ Review required — awaiting: @janjust OpenSHMEM currently discovers static symmetric storage by scanning writable mappings in /proc/self/maps. Anonymous runtime mappings can therefore enter the memheap table, even though those mappings may differ betwee...
#14332 Work around buggy libofi CXI provider devreal 2026-08-18 (1m) Target: main, backport: v6.0.x Not yet backported - pending merge to main first. +22 / -9 ⏳ Review required — awaiting: @hppritcha, @jiaxiyan libofi's CXI provider happily returns an already free'd fi_fabric object from its cache, leading to a Segfault when we try to create a new fabric. This happens in OMPI because we query both mtl/ofi and btl/ofi, which ...
#14331 accelerator/cuda: fix memory pool and memory context handling devreal 2026-08-18 (1m) Target: main, backport: v5.0.x, backport: v6.0.x Not yet backported - pending merge to main first. +26 / -11 🔴 Changes requested — changes requested by bosilca — awaiting: @Akshay-Venkatesh, @janjust This PR addresses two issues I encountered recently: 1) When we detect a context for the memory we're probing in check_addr we would just return without setting the output device ID and after setting the memory cont...
#14306 Emit MPI Standard ABI handles in MPI_T event payloads hppritcha 2026-08-10 (1m) Target: main, backport: v6.0.x Not yet backported - pending merge to main first. +2471 / -78 🔴 Changes requested — changes requested by jsquyres — awaiting: @devreal, @jsquyres Open MPI's MPI_T event producers publish MPI object handles (communicator, window, session, error handler, file) in their event payloads. Those values were always the internal object pointer, which is correct only for...
#14262 OSC/UCX: Fix non-contiguous origin iovec traversal in MINLOC/MAXLOC accumulate bwestheimer 2026-08-02 (1m) Target: main - +432 / -87 🔴 Changes requested — changes requested by bosilca — awaiting: @janjust The three MINLOC/MAXLOC reduce loops in osc_ucx_comm.c walked the origin buffer through the iovec list built by create_iov_list(), stepping an absolute pointer by the staging buffer's element stride and handing ra...
#14260 OSC/UCX: Size accumulate staging buffers with the datatype span bwestheimer 2026-08-02 (1m) Target: main - +43 / -29 🔴 Changes requested — changes requested by bosilca — awaiting: @janjust Three OSC/UCX accumulate paths allocated primitive staging arrays as true_extent * count, even though the data is produced and consumed at the datatype extent. For padded pair types and counts greater than one, that
#14256 Fix CodeQL integer-multiplication overflow alerts jsquyres 2026-08-02 (1m) Target: main, backport: v6.0.x Not yet backported - pending merge to main first. +91 / -65 🔴 Changes requested — changes requested by bosilca — approved by devreal, edgargabriel, hppritcha This triages and fixes 34 of the 38 open cpp/integer-multiplication-cast-to-long CodeQL code scanning alerts (the remaining 4 are dismissed as bounded by construction — see below). Every site was individually triage...
#14254 Fix CodeQL snprintf/scanf alerts in monitoring, tuned, and comm_method jsquyres 2026-08-02 (1m) Target: main, backport: v6.0.x Not yet backported - pending merge to main first. +44 / -18 ✅ Approved — approved by devreal This fixes the three remaining small CodeQL code scanning alerts, plus two adjacent bugs found while fixing them. Four commits: common/monitoring ([alert 41](https://github.com/open-mpi/ompi/security/code-scanning...
#14253 Fix CodeQL snprintf-overflow and format-specifier alerts in coll jsquyres 2026-08-02 (1m) Target: main, backport: v5.0.x, backport: v6.0.x Not yet backported - pending merge to main first. +50 / -30 ✅ Approved — approved by bosilca — awaiting: @devreal This fixes four CodeQL code scanning alerts, in two commits. Commit 1 resolves alert 40 (cpp/overflowing-snprintf) in coll/han: the loop building t...
#14207 osc/ucx: Honor origin datatype lower bounds in accumulates jeffhammond 2026-07-24 (1m) Target: main, backport: v6.0.x Not yet backported - pending merge to main first. +6 / -3 ⏳ Review required — awaiting: @devreal Contiguous derived datatypes may use a nonzero true lower bound, including absolute hindexed displacements paired with MPI_BOTTOM. The UCX accumulate reduction paths classified these datatypes as contiguous but passed...
#14204 osc/rdma: Preserve remote base for local window peers jeffhammond 2026-07-24 (1m) Target: main, backport: v6.0.x Not yet backported - pending merge to main first. +12 / -7 ⏳ Review required — awaiting: @devreal MPI_Win_allocate maps a node-local shared segment into every process. When CPU and NIC atomics cannot be mixed, local accumulates are sent through the node leader's BTL endpoint and must use the leader's registered vi...
#14026 DOCS: add more explicit info about fortran interfaces hppritcha 2026-06-16 (3m) Target: main - +125 / -0 ✅ Approved — approved by edgargabriel text generated largely by claude based off of recent commits done by human to add new functions to mpif-h, mpi f90, mpi f08 interfaces.
#13969 opal/lifo: fix missing acquire barriers in pop on weak-memory architectures devreal 2026-06-13 (3m) Severity: critical, Target: main, Target: v6.0.x - +34 / -17 🔴 Changes requested — changes requested by jsquyres — awaiting: @bosilca, @hppritcha opal/lifo: fix missing acquire barriers in pop on weak-memory architectures Three release/acquire pairs were broken in opal_lifo_pop_atomic across the three compiled variants of the function. All atomics in OPAL are
#13958 mca: fix deprecated variable source warning francdoc 2026-06-11 (3m) Target: main - +35 / -19 ✅ Approved — approved by rhc54 Linked to #12437. This PR fixes deprecated MCA synonym warning source reporting. Previously, if mca_base_component_show_load_errors was set in an MCA params file and the deprecated mca_component_show_load_errors n...
#13949 coll/ucc: give pset-derived communicators their own OOB and context bosilca 2026-06-09 (3m) Target: main, backport: v6.0.x Not yet backported - pending merge to main first. +527 / -152 ⏳ Review required — awaiting: @Sergei-Lebedev, @janjust UCC previously bootstrapped a single, process-global UCC context whose out-of-band (OOB) channel was bound to MPI_COMM_WORLD, and built every team on that context using an endpoint map into the world rank space. That ...
#13942 Add pdeb modular process debugger helper bosilca 2026-06-05 (3m) Target: main, backport: v6.0.x Not yet backported - pending merge to main first. +2948 / -0 ⏳ Review required — awaiting: @jsquyres I developed these tools over the years to compensate for the lack of parallel debuggers and to help debug OMPI internals. It now seems like the right time to contribute them back to the community, in case others find ...
#13903 rocm/accelerator: add HIP Virtual Memory Management (VMM) IPC support mshanthagit 2026-05-21 (4m) Target: main - +646 / -66 🔴 Changes requested — changes requested by bosilca Implements HIP VMM-based IPC for the ROCm accelerator component and fixes misc bugs.
#13896 osc/rdma: extend MPI_Win_shared_query to MPI_Win_create windows via smsc devreal 2026-05-14 (4m) Target: main - +159 / -42 ⏳ Review required — commented only: copilot-pull-request-reviewer Follow-up to https://github.com/open-mpi/ompi/pull/13330 to support MPI_Win_shared_query on windows created using MPI_WIN_CREATE in osc/rdma. This is supported if xpmem is available via smsc. The mapping is created ...
#13876 Add hook/fault_injector debugging tool Matthew-Whitlock 2026-05-06 (4m) Target: main - +338 / -0 🔴 Changes requested — awaiting: @bosilca Any interest in making this a core ompi feature? edit: I'm really asking if there's any willingness to let this be an ompi feature - I can supply the interest :) It's a helpful tool for debugging fault tolerance funct...
#13830 HPE Cassini (CXI) Hardware Performance Counters (HWPC) Feature sadlochr 2026-04-20 (5m) Target: main - +6663 / -19 🔴 Changes requested — approved by ryanhankins This feature allows users to track and output various Cassini hardware performance counters over execution time. During MPI_Init(), it will determine which hardware counters to track depending on user input. During th...
#13789 opal/sys: add RISC-V 64-bit atomic LL/SC and full atomic support devreal 2026-03-24 (5m) Target: main - +605 / -39 ⏳ Review required — commented only: copilot-pull-request-reviewer Add architecture-specific atomic operations for RISC-V 64-bit: - atomic_llsc.h: load-reserved/store-conditional macros using lr.w.aq/ sc.w.rl (32-bit) and lr.d.aq/sc.d.rl (64-bit) with acquire/release ordering, matchi...
#13667 Properly del_proc spawned procs upon instance finalize Dupratj 2026-01-20 (8m) Target: main - +153 / -17 ⏳ Review required — awaiting: @hppritcha Upon finalize the ressources allocated for a spawned proc are not all released. This Pull request is a take over of 13445
#13656 coll/ucc: enable asymmetric datatype check in UCC without OMPI fallback QiaoK 2026-01-15 (8m) bug, Target: main - +52 / -0 ⏳ Review required — awaiting: @Sergei-Lebedev This change enables UCC to perform asymmetric datatype checks internally for gather, gatherv, scatter, and scatterv operations. This feature is only available in UCC 1.8 or greater. The motivation for this change is t...
#13654 wipe the generalized request upon demand bosilca 2026-01-14 (8m) Target: main - +19 / -12 ⏳ Review required — no reviewer assigned, no reviews yet If the user calls MPI_Request_free on a generalized request, make sure the request, both user side and MPI side, is properly released. However, make sure the generalized request free_fn callback function will be calle...
#13527 pml/ob1: introduce a new protocol for intermediate-sized messages hjelmn 2025-11-21 (10m) Target: main - +670 / -142 ⏳ Review required — commented only: bosilca, hjelmn This commit introduces the multi-eager protocol to ob1. This protocol works by fragmenting into multiple eager-sized messages and sending them in parallel to the destination. On the receiver the first fragment is matc...
#13526 btl/uct: update btl default limits hjelmn 2025-11-21 (10m) Target: main - +10 / -8 🔴 Changes requested — changes requested by bosilca When a module is created for a UCT transport we take the min of the transport limits and whatever the default is in the module. This is done to support user provided limits. The issue is the limits are too low and the...
#13519 ucx: meet at the pmix fence before disconnecting to avoid an infinite loop jeking3 2025-11-14 (10m) Target: main - +8 / -2 🔴 Changes requested — changes requested by bosilca — awaiting: @janjust While moving a job from a small number of GPU nodes to a larger number of CPU nodes, I was able to reliably reproduce #11087 in my environment. While in the debugger, I found that opal_common_ucx_mca_pmix_fence was ...
#13433 interlib: fix thread safety issue ggouaillardet 2025-10-10 (11m) bug, Target: main - +6 / -3 ⏳ Review required — no reviewer assigned, no reviews yet getenv() is not thread-safe so it should not be invoked within a callback subroutine. So read the environment at initialization and cache the result. Thanks Zdeněk Salvet for reporting this issue. Refs #13428
#13195 btl/uct: allow connections to be formed using a separate memory domain hjelmn 2025-04-14 (1y5m) Target: main - +459 / -217 ⏳ Review required — commented only: bosilca, hjelmn It is possible that the current memory domain does not have an adequate transport for forming endpoint to endpoint connections. When this is the case the btl will fail to function. To support these situations this CL ...
#13147 Extract the install prefix from the shared library. bosilca 2025-03-17 (1y6m) Target: main - +225 / -22 ⏳ Review required — commented only: bosilca, jsquyres This is part of a multi-project effort, a similar PR will be created in OpenPMIX and OMPI. The goal of each of these changes is the same: instead of using build-time generated prefix that ignore a project rebase, take...
#13088 RFC: Provide equivalence of MPICH_ASYNC_PROGRESS hominhquan 2025-02-11 (1y7m) Target: main - +303 / -45 ⏳ Review required — commented only: bosilca, devreal, hominhquan, jsquyres This PR is follow-up of #13074.
#13039 part-persist: implement message aggregation AxelSchneewind 2025-01-15 (1y8m) Target: main - +352 / -44 🔴 Changes requested — commented only: AxelSchneewind, mdosanjh Created a component for partitioned communication that supports message aggregation, based on the existing part-persist component. The goal is to prevent drops in effective bandwidth when using too fine-grained partit...
#12935 part: multithreading deadlocks fixes and safety checks Keluaa 2024-11-26 (1y9m) Target: main - +172 / -148 ⏳ Review required — no reviewer assigned, no reviews yet This PR fixes multiple deadlocks and issues encountered with partitioned communications. The first deadlocks occur when one thread is in opal_progress and others are working on the partitioned request: - `MPI_Pready...
#12826 mca/base: add a new MCA variable type for include lists hjelmn 2024-09-24 (1y11m) Target: main - +910 / -26 ⏳ Review required — awaiting: @bwbarrett, @jsquyres It is not uncommon in Open MPI to register a string variable to allow specifying an exclude or exclude list (every framework registers one). Given this common patten it is worthwhile to formalize an MCA include list v...
#12796 Created the part direct module mdosanjh 2024-09-05 (2y) Target: main - +2444 / -4 ⏳ Review required — commented only: AxelSchneewind Created the part direct module. This adds a module for partitioned communication that uses RMA similar to better match the original research. This has performance benefits but places additional constraints on the user...
#12725 opal_thread integration with libevent npe9 2024-08-01 (2y1m) Target: main - +187 / -2 ⏳ Review required — no reviewer assigned, no reviews yet This is an initial stab at changing the Open MPI libevent integration to use opal threads instead of pthreads for alternative threading libraries. Currently, although opal threads are now replaceable via an mca, opal_...
#12586 Rework the MPI connect/accept code bosilca 2024-05-29 (2y3m) Target: main - +26 / -29 ⏳ Review required — no reviewer assigned, no reviews yet I'm puzzled by this code working. It does all the right things, but it was calling PML->add_procs with zero new procs, ALWAYS (the ilist was empty by the time we got to the add_procs call). But somehow it still work...
#12570 Make datatype copy stream-aware devreal 2024-05-23 (2y4m) Target: main - +87 / -34 ⏳ Review required — awaiting: @bosilca This is the third chunk of #12318, which makes datatype copying stream-aware. This will be needed in the coll components to order data movement with kernel invocations. Adds `ompi_datatype_copy_content_same_ddt_stream...
#12569 Add CUDA/HIP implementations of reduction operators devreal 2024-05-23 (2y4m) Target: main - +8490 / -52 🔴 Changes requested — awaiting: @Akshay-Venkatesh This is the second part #12318, which provides the device-side reduction operators and adds stream semantics to ompi_op_reduce. As usual, the operators are generated from macros. Function pointers to kernel launch f...
#12453 mca/coll: Add any radix k for alltoall bruck algorithm jiaxiyan 2024-04-05 (2y5m) Target: main - +130 / -45 ⏳ Review required — awaiting: @juntangc, @lrbison This method extends ompi_coll_base_alltoall_intra_bruck to handle any radix k.
#12285 test/datatype: do not use installed components ggouaillardet 2024-01-28 (2y7m) bug, Target: main - +2 / -1 ✅ Approved — approved by jsquyres By default, the tests will use the components in the build directory but also in the installation directory if any. That can cause some bizarre issues when they are not compatible. Thanks Kook Jin Noh for the initial ...
#12259 Fix incorrect TCP connections. bosilca 2024-01-21 (2y8m) Target: main - +125 / -11 ⏳ Review required — awaiting: @devreal, @ggouaillardet, @jsquyres If nodes have the same IP addresses (for containers or other purposes) and these addresses get published as part of the modex, a remote peer might try to use one of the addresses to connect. As both nodes have the sam...
#12122 mpi-ext.h: include <mpi.h> jsquyres 2023-11-23 (2y10m) Target: main - +3 / -0 ✅ Approved — approved by devreal — awaiting: @bosilca Ensure to #include <mpi.h> at the top of mpi-ext.h, just so that it includes files that it depends on. Thanks to @eschnett for raising the issue. Refs #12111
#12119 Allow MPI_Comm_disconnect on intracoms. bosilca 2023-11-23 (2y10m) Target: main - +33 / -45 ⏳ Review required — awaiting: @devreal, @jsquyres Rework the logic to avoid creating temporary arrays of processes. Indeed, in order to call into the PMIX_Fence we only need an array of pmix_proc_t which is already inefficient enough to not require an intermediary st...
#11628 Topic/improved alltoallv bosilca 2023-04-27 (3y4m) Target: main - +228 / -17 ⏳ Review required — awaiting: @wzamazon Instead of a large number of persistent requests (for both sends and receives) work on a predefined size window. This allows us to control the number of concurrent communications per process.
#9934 tuned decision for single node communicators bosilca 2022-01-27 (4y7m) enhancement, Target: main - +478 / -18 ⏳ Review required — commented only: bosilca, wenduwan 1. Cache the peer location information in the group itself, doing the location discovery only once per group. 2. Split the collective decision function to allow for different behaviors if the communicator spans over a...
#8597 decide CONVERTOR_CUDA flag based on first byte of buffer markalle 2021-03-12 (5y6m) - - +70 / -18 ⏳ Review required — awaiting: @bosilca There's a cuda check in the converter that checks one address and caches a flag for CONVERTOR_CUDA. But it was just checking pUserBuf which isn't necessarily anywhere near the data, and doesn't even need to be a legal...

Target branch: v6.0.x (8 open PRs)

# Title Author Created (age) Labels Merge/Backport Status LOC (+/-) Review Status Description
#14500 [v6.0.x] Drop recommonmark dependency app/ompi-backport-bot 2026-09-23 (0d) Target: v6.0.x, target:v6.0.x Backport of merged main PR #14215 (merged 2026-09-23). +1 / -3 ✅ Approved — approved by jsquyres Backport of #14215 to v6.0.x. --- Fixes #14214
#14497 [v6.0.x] Fix fortran double complex support app/ompi-backport-bot 2026-09-23 (0d) Target: v6.0.x, target:v6.0.x Backport of merged main PR #14478 (merged 2026-09-23). +180 / -3 ✅ Approved — approved by hppritcha Backport of #14478 to v6.0.x. --- Fixes #14471
#14491 [v6.0.x] Fix MPI_Request_get_status() so that the empty-status path preserves MPI_ERROR = MPI_SUCCESS app/ompi-backport-bot 2026-09-22 (1d) NEWS, Target: v6.0.x, target:v6.0.x Backport of merged main PR #14486 (merged 2026-09-22). +1 / -1 ⏳ Review required — no reviewer assigned, no reviews yet Backport of #14486 to v6.0.x. --- The MPI Standard allows MPI_Request_get_status() to be called with MPI_REQUEST_NULL or an inactive request. In that case, the routine returns flag = true and an empty status.
#14490 [v6.0.x] Deal with the MPI_ERROR field for empty statuses app/ompi-backport-bot 2026-09-22 (1d) Target: v6.0.x, target:v6.0.x Backport of merged main PR #14488 (merged 2026-09-22). +14 / -14 ⏳ Review required — no reviewer assigned, no reviews yet Backport of #14488 to v6.0.x. --- The practical restatement for an implementer: only write MPI_ERROR when you are (a) stamping an empty status, or (b) inside a multi-status completion that is returning MPI_ERR_IN_ST...
#14489 [v6.0.x] PRRTE/PMIx: advance shas for 6.0.0RC2 hppritcha 2026-09-22 (1d) Target: v6.0.x standalone v6.0.x-only change (RC shas bump), bot:notacherrypick +2 / -2 ⏳ Review required — awaiting: @edgargabriel bot:notacherrypick
#14485 [v6.0.x] flang: add a patch file to extra dist app/ompi-backport-bot 2026-09-22 (1d) Target: v6.0.x, target:v6.0.x Backport of merged main PR #14440 (merged 2026-09-15). +1 / -0 ⏳ Review required — no reviewer assigned, no reviews yet Backport of #14440 to v6.0.x. --- that was missing
#14473 [v6.0.x] Rename the MPI Forum ABI pkg-config files to ompi-forum-abi*.pc app/ompi-backport-bot 2026-09-21 (2d) Target: v6.0.x, target:v6.0.x Backport of merged main PR #14467 (merged 2026-09-21). +89 / -72 ⏳ Review required — no reviewer assigned, no reviews yet Backport of #14467 to v6.0.x. --- Stacked on #14466 — please review/merge that one first.
#14459 [v6.0.x] Make OpenMPI request FI_PROGRESS_CONTROL_UNIFIED app/ompi-backport-bot 2026-09-17 (6d) Target: v6.0.x, target:v6.0.x Backport of merged main PR #14368 (merged 2026-09-17). +40 / -6 ⏳ Review required — no reviewer assigned, no reviews yet Backport of #14368 to v6.0.x. --- When OpenMPI uses FI_THREAD_DOMAIN, it already serializes libfabric control

Target branch: jsq/rename-standard-abi-switch (1 open PRs)

# Title Author Created (age) Labels Merge/Backport Status LOC (+/-) Review Status Description
#14475 Rename internal "standard ABI" identifiers to "forum" jsquyres 2026-09-21 (2d) Target: main, backport: v6.0.x - +628 / -629 ⏳ Review required — no reviewer assigned, no reviews yet Stacked on #14474 — review/merge that first; this PR's diff is only the last two commits. ompi/mpi/c/standard_*.c was added with the first ABI bindings in 2023; nothing generates files by that name anymore (gene...

Draft / WIP PRs (18)

Pulled out of the tables above because they are marked as a GitHub draft, or carry a WIP/DNM label, and are not realistically mergeable as-is. Same columns, same newest-to-oldest ordering within each target branch.

Target branch: main (12 draft/WIP PRs)

# Title Author Created (age) Labels Merge/Backport Status LOC (+/-) Review Status Description
#14461 dpm: connect the two sides with one PMIx group, not publish/lookup rhc54 2026-09-18 (5d) DRAFT, Target: main - +992 / -195 ⏳ Review required — no reviewer assigned, no reviews yet > CI will fail until the PMIx and PRRTE submodules are advanced to their current masters. That update is being done separately; this is up so the changes can be reviewed. See Dependencies. `ompi_d...
#14448 UBCL: PML/UBCL fixes and allow truncated messages vanman-nguyen 2026-09-16 (7d) DRAFT, Target: main, backport: v6.0.x Not yet backported - pending merge to main first. +32 / -8 ⏳ Review required — no reviewer assigned, no reviews yet This PR enables the support of truncated pt2pt communications by default in the PML/UBCL. This was previously disabled because the behaviour of UBCL regarding truncation was not entirely compliant with the MPI standar...
#14263 OSC/UCX: Defer nonblocking-accumulate finalize to top-level progress bwestheimer 2026-08-02 (1m) DRAFT, Target: main - +628 / -163 ⏳ Review required — no reviewer assigned, no reviews yet With osc_ucx_enable_nonblocking_accumulate=1, accumulate finalize issues ucp_worker_fence() and then calls ucp_worker_progress() from inside a UCX send completion callback, re-entering ucp_worker_progress. Som...
#14195 btl/tcp: fix the multi-module wire-up race (replace the #5917 workaround) jsquyres 2026-07-21 (2m) DRAFT, Target: main, backport: v6.0.x Not yet backported - pending merge to main first. +142 / -110 ⏳ Review required — awaiting: @bosilca, @bwbarrett, @hppritcha While reviewing BTL framework code/comments recently, I came across the MCA_BTL_FLAGS_SINGLE_ADD_PROCS workaround that #5917 added to the TCP BTL back in 2018. That PR's description explicitly says it is a workaroun...
#13943 accelerator/cuda: skip VMM/mpool fallback for unrecognized pointers and add tunables to disable them entirely bosilca 2026-06-05 (3m) DRAFT, Target: main - +604 / -11 ⏳ Review required — awaiting: @devreal, @edgargabriel Builds on a312180b43 ("defer VMM/mpool checks in check_addr fast path"), which short-circuits classification of cudaMalloc pointers reported as CU_MEMORYTYPE_DEVICE with a valid context. Two further reductions to the ...
#13767 coll/ucc: fix ucc resource leak on MPI_Finalize janjust 2026-03-17 (6m) ⚠️ WIP-DNM!, Target: main - +26 / -45 ⏳ Review required — awaiting: @bosilca All UCC team teardown and context/library finalization was driven by a communicator attribute delete callback on MPI_COMM_WORLD. OMPI commit https://github.com/open-mpi/ompi/commit/b79004e77f1b5e247a1bd7a6cfc34d1a80ae...
#13216 OSC/UCX: explicitly flush default endpoints MamziB 2025-04-28 (1y4m) DRAFT, Target: main - +26 / -0 🔴 Changes requested — commented only: MamziB, bosilca, devreal This PR fixes a hang in osc/ucx observed during MPI_Finalize with Graph500 by ensuring no outstanding operations remain on default endpoints before they are destroyed.
#13188 configury: enhance short float detection with clang ggouaillardet 2025-04-10 (1y5m) DRAFT, Target: main - +13 / -5 🔴 Changes requested — changes requested by bwbarrett With clang, short float support requires --rtlib=compiler-rt to be passed to the linker. In order to avoid a false positive, it is necessary have the runtime called in the test program. The test program was hence upda...
#12991 coll/accelerator: allow to select functions to register edgargabriel 2024-12-19 (1y9m) ⚠️ WIP-DNM!, Target: main - +123 / -12 ✅ Approved — approved by devreal This PR introduces the ability to register the component only to the select functions specified by an MCA parameter string. The idea and the code is based on the UCC component, and some of the bits might be moved late...
#12745 configury: fix IGNORE_TKR check ggouaillardet 2024-08-12 (2y1m) DRAFT, Target: main - +1 / -1 ✅ Approved — approved by jsquyres As reported by Chris Parrott in #12681, the enhanced test was buggy. Pass the expected argument to make it Fortran compliant and LLVM 17 and above happy compilers
#9867 Remove all ops with 3 buffers bosilca 2022-01-12 (4y8m) ⚠️ WIP-DNM!, Target: main - +2 / -856 ✅ Approved — approved by jsquyres We planned to remove support for all the MPI_Op with 3 buffer arguments, but this feature might be useful for the MPI extension described in https://github.com/mpi-forum/mpi-issues/issues/530. Full discussion in #9719.
#6426 Add MCA parameters to define the size of memcpy chunks. bosilca 2019-02-23 (7y7m) ⚠️ WIP-DNM!, performance, Target: main, Target: v6.0.x - +152 / -58 ⏳ Review required — no reviewer assigned, no reviews yet CUDA memcpys are divided in many very small chunks (a fixed size of 128k), leading to terrible performance. To improve performance we need to increase this limit (maybe even remove it), and add support for vector CUDA...

Target branch: v6.0.x (2 draft/WIP PRs)

# Title Author Created (age) Labels Merge/Backport Status LOC (+/-) Review Status Description
#13768 coll/ucc: fix ucc finalize leak on MPI_Finalize janjust 2026-03-17 (6m) ⚠️ WIP-DNM!, Target: v6.0.x direct cherry-pick of a main commit, not tied to an open/tracked main PR +26 / -45 ⏳ Review required — awaiting: @bosilca, @edgargabriel, @hppritcha commit b79004e77f intentionally skips user-defined attribute delete callbacks on MPI_COMM_WORLD during MPI_Finalize to fix an issue where PETSc attribute callbacks invoked MPI after finalize had started (PR #12072 / i...
#13658 v6.0.x: coll/ucc: enable asymmetric datatype check in UCC without OMPI fallback QiaoK 2026-01-16 (8m) DRAFT, bug, Target: v6.0.x parallel direct submission to v6.0.x (no separate main PR); twin of open PR #13657 on v5.0.x +52 / -0 🔴 Changes requested — changes requested by jsquyres — awaiting: @janjust This change enables UCC to perform asymmetric datatype checks internally for gather, gatherv, scatter, and scatterv operations. This feature is only available in UCC 1.8 or greater. The motivation for this change is t...

Target branch: v5.0.x (4 draft/WIP PRs)

# Title Author Created (age) Labels Merge/Backport Status LOC (+/-) Review Status Description
#14163 [v5.0.x] Fix file descriptor leaks and teardown ordering in MPI instance finalize app/ompi-backport-bot 2026-07-13 (2m) DRAFT, Target: v5.0.x, target:v5.0.x Backport of merged main PR #14154 (merged 2026-07-13). MERGED sibling backport [v6.0.x] #14162 (merged 2026-07-14). +40 / -7 ⏳ Review required — awaiting: @bwbarrett Backport of #14154 to v5.0.x. --- Three teardown bugs in the common ompi_mpi_instance_* / OPAL finalize path. They are not sessions-specific -- MPI_Init() goes through the same ompi_mpi_instance_init() -- but ...
#13953 v5.0.x: docs: add MPI-4.0 and MPI-4.1 conformance gap analyses jsquyres 2026-06-10 (3m) DRAFT, Target: v5.0.x standalone v5.0.x-only docs change, bot:notacherrypick +382 / -34 ⏳ Review required — no reviewer assigned, no reviews yet Similar to #13954, I had Claude analyze the v5.0.x branch vs. the MPI Forum 4.0 and 4.1 PDFs, and update our RST docs based on the actual conformance in the git branch. This PR represents the results. I have not (ye...
#13657 v5.0.x: coll/ucc: enable asymmetric datatype check in UCC without OMPI fallback QiaoK 2026-01-16 (8m) DRAFT, bug, ⚠️ WIP-DNM!, Target: v5.0.x parallel direct submission to v5.0.x (no separate main PR); twin of open PR #13658 on v6.0.x +52 / -0 🔴 Changes requested — changes requested by jsquyres — approved by janjust This change enables UCC to perform asymmetric datatype checks internally for gather, gatherv, scatter, and scatterv operations. This feature is only available in UCC 1.8 or greater. The motivation for this change is t...
#12253 v5.0.x: Propagate the error up to the user. jsquyres 2024-01-19 (2y8m) bug, ⚠️ WIP-DNM!, Target: v5.0.x Backport of merged main PR #11683 (merged 2024-01-19). no separate v6.0.x backport PR found (source PR predates current v6.0.x branch point). +47 / -20 ⏳ Review required — awaiting: @bosilca Make sure to reset the generalized request to guarantee not to call the free callback a second time. (cherry picked from commit ac3647e16d2b9765fec2e60a61c92e9b8661efc4) This is the v5.0.x PR corresponding to main PR ...

Clone this wiki locally