Fork Sync: Update from parent repository#175
Open
github-actions[bot] wants to merge 10 commits into
Open
Conversation
Claude seems to want it - port of open-mpi/ompi#14065 Signed-off-by: Ralph Castain <rhc@pmix.org> (cherry picked from commit 4a2e55e)
Add a Contributor Covenant v1.4 Code of Conduct for the PRRTE project, following the same pattern used by OpenPMIx. The authoritative source is .github/CODE_OF_CONDUCT.md (Markdown); docs/Makefile.am drives a Python converter script that regenerates docs/code-of-conduct.rst at build time so the content appears in the Sphinx-built HTML documentation. The new page is wired into the docs/index.rst toctree between the Contributing and License sections. Signed-off-by: Ralph Castain <rhc@pmix.org> (cherry picked from commit dfe14c6)
ReadTheDocs does not run Automake, so the pre_build job in .readthedocs.yaml must invoke the generator script directly before Sphinx runs. Also add docs/code-of-conduct.rst to .gitignore since it is a generated file and should not be tracked. Signed-off-by: Ralph Castain <rhc@pmix.org> (cherry picked from commit 3b5598d)
LSF sets LSB_AFFINITY_HOSTFILE for every job, even when the user has not requested process pinning, in which case the file is empty. The new rmaps/lsf component introduced in v3.0.13 fires whenever that environment variable is present, calls file_parse(), which returns PRTE_SUCCESS immediately on an empty file leaving num_ranks at zero, and then incorrectly triggers a fatal "bad-syntax" error instead of stepping aside. In v3.0.12 the equivalent code path returned from the rank_file mapper and let round_robin take over. Five issues are addressed: - When file_parse() returns with num_ranks == 0 (empty affinity file), return PRTE_ERR_TAKE_NEXT_OPTION so normal mapping proceeds. This fixes the reported regression (GitHub issue #2449). - Reset the static num_ranks counter to 0 at the top of lsf_map() and in the error path. The counter was never cleared between jobs, so the second and subsequent jobs in a DVM session would store rankmap entries at wrong indices and read back NULLs, silently producing incorrect process placement. - Move "options->map = PRTE_MAPPING_BYUSER" to just after last_mapper is set, before the per-app loop. It was inside the per-process inner loop, so prte_rmaps_base_get_target_nodes() was called on the first app with the wrong map type. The rank_file mapper sets this flag in the same early position. - Fix the error: path to destruct rankmap and release its rfmap entries, matching what the success path already does. Previously only node_list was cleaned up on error. - Remove the dead assigned_ranks_array from file_parse(). It was allocated and initialized (copied from the rank_file parser) but never written to, read, or freed. Also initialize sep = NULL to prevent use of an indeterminate value if a malformed affinity-file line contains no space separator. Signed-off-by: Ralph Castain <rhc@pmix.org> (cherry picked from commit a01985a)
All RAS components (lsf, slurm, pbs, flux, gridengine, hosts) create prte_node_t objects and hand them to prte_ras_base_node_insert, but none of them applied the FQDN-to-short-name normalization that hostfile.c and dash_host.c had been doing independently. This meant nodes sourced directly from a scheduler API (e.g. lsb_getalloc()) were inserted into prte_node_pool with whatever hostname the scheduler returned, with no short-name alias, so prte_quickmatch and prte_nptr_match could fail when the affinity file or -host list used a different form of the same name. Add a static normalize_node() helper in ras_base_node.c. When keep_fqdn_hostnames is not set and a node name contains a dot, it copies the full FQDN, truncates node->name in place to the short form, stores the FQDN in node->rawname (or in aliases if rawname is already populated), and appends rawname to node->aliases. normalize_node() is called for every new node just before it is inserted into prte_node_pool, after the existing FQDN-detection check so that prte_have_fqdn_allocation is still set correctly. Remove the equivalent per-site normalization code from hostfile.c (STRING, RELATIVE, and RANK token branches) and dash_host.c. Both utilities now create node objects with the raw name from the file or command line; the RAS base normalizes on insertion. The matching functions (prte_quickmatch, prte_node_match, prte_nptr_match) all check node->aliases, so nodes produced by these utilities in the rmaps filter path continue to match against pool nodes correctly because the pool node's FQDN alias satisfies the alias check even when the filter node carries only the original unstripped name. Signed-off-by: Ralph Castain <rhc@pmix.org> (cherry picked from commit ea2b8e6)
…actor Commit 8ddaa19 removed the alias variable from hostfile_parse_line when centralizing FQDN/short-name normalization into the RAS base, but left behind the alias %s format specifier and its argument in the PMIX_OUTPUT_VERBOSE call at the top of the include-node path. Drop both so the code compiles. Signed-off-by: Ralph Castain <rhc@pmix.org> (cherry picked from commit 38c7eb9)
Update them to PRRTE Signed-off-by: Ralph Castain <rhc@pmix.org> (cherry picked from commit cbbf009)
Display directives (map, bindings, allocation, ...) are job-level: there is no way to scope a display to a single app context. When a command line carries more than one - most easily by attaching one to a second app in an MPMD line - silently applying just one would be surprising. Detect this in prte_prun_parse_common_cli(), which both prun and the prte HNP launcher route through, and reject it with the existing "multi-instances" diagnostic and PRTE_ERR_BAD_PARAM rather than promoting an app-level directive to the job (which has no sensible meaning for display). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Signed-off-by: Ralph Castain <rhc@pmix.org> (cherry picked from commit 63754e3)
Add a golden-rule bullet to AGENTS.md pointing AI coding agents at the portable compiler attribute wrappers defined in src/include/prte_config_bottom.h (pulled in transitively by prte_config.h). These macros — __prte_attribute_unused__, __prte_attribute_noreturn__, __prte_attribute_format__, __prte_attribute_deprecated__, and many more — expand to the appropriate __attribute__((...)) on compilers that support it and to nothing elsewhere. The new guidance tells agents to reach for them (for example, to mark an unused function parameter) rather than writing a bare __attribute__ or leaving a compiler warning unaddressed. Docs-only change to the agent guidance file; no code or build impact. Signed-off-by: Ralph Castain <rhc@pmix.org> (cherry picked from commit 15f1c27)
Signed-off-by: Ralph Castain <rhc@pmix.org>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.