PR #39797: [PD] Bump Nixl version

• Repository: vllm-project/vllm
• Author: NickLucche
• State: open | merged: False
• Link: https://github.com/vllm-project/vllm/pull/39797

Description (problem / solution / changelog)

Address https://github.com/vllm-project/vllm/issues/39521.

Mind that we're pinning cu13 on vllm right now.

Changed files

• requirements/kv_connectors.txt (modified, +3/-3)

PR #39851: [CI][NIXL] Fix PD CI breakage: pin nixl-cu{12,13} versions

• Repository: vllm-project/vllm
• Author: ZhanqiuHu
• State: closed | merged: True
• Link: https://github.com/vllm-project/vllm/pull/39851

Description (problem / solution / changelog)

nixl-cu12==1.0.1 dropped on PyPI today (19:38 UTC) and ships nixl_ep compiled against libcudart.so.12 — crashes on CUDA 13 CI runners. Our < 0.10.0 constraint only pins the meta-package, not the backends:

[2026-04-14T21:21:43Z]  + nixl==0.9.0
[2026-04-14T21:21:43Z]  + nixl-cu12==1.0.1
[2026-04-14T21:21:43Z]  + nixl-cu13==1.0.1

• Passing: https://buildkite.com/vllm/ci/builds/61159 (nixl-cu12==1.0.0, no nixl_ep)
• Failing: https://buildkite.com/vllm/ci/builds/61308 (nixl-cu12==1.0.1, nixl_ep import crashes)

Temp fix: pin nixl-cu12 and nixl-cu13 to < 0.10.0. @NickLucche is working on the proper version bump in #39797 (tracking #39521).

Changed files

• requirements/kv_connectors.txt (modified, +2/-0)

PR #1546: PLUGINMANAGER: Defer plugin load to the point where they are actually used

• Repository: ai-dynamo/nixl
• Author: ovidiusm
• State: open | merged: False
• Link: https://github.com/ai-dynamo/nixl/pull/1546

Description (problem / solution / changelog)

What?

Defer plugin dlopen from discovery time to first use. Plugins found on disk are now recorded by name only; the actual dlopen happens when createBackend() or getPluginParams() is called for that specific plugin.

Why?

https://github.com/vllm-project/vllm/issues/39521

Creating a nixl_agent crashes with a segfault on some systems (RHEL 9 / Rocky 9 bare-metal DGX H100) due to eager plugin loading.

During nixlPluginManager construction, every plugin .so in the plugins directory is dlopen'd, including plugins the user never requested (e.g. LIBFABRIC on InfiniBand systems). The dlopen triggers transitive dependency initialization (libnuma -> numa_hook.so) which crashes before any NIXL code even executes. The Python nixl_agent.__init__ further aggravates this by calling getPluginParams() on every discovered plugin, forcing a second dlopen attempt for each.

How?

C++ (plugin_manager.h, nixl_plugin_manager.cpp):

• Add discovered_backend_plugins_ and discovered_telemetry_plugins_ sets to track plugin names found on disk without loading them.
• Add explicit_plugin_paths_ map to preserve exact .so paths from the plugin list file, so loadBackendPlugin() tries those first before scanning directories.
• discoverBackendPlugin() / discoverTelemetryPlugin(): record the plugin name instead of calling loadBackendPlugin().
• loadPluginsFromList(): same, record names and paths instead of dlopen'ing.
• getLoadedBackendPluginNames() / getLoadedTelemetryPluginNames(): include discovered names so getAvailPlugins() still returns them.
• loadBackendPlugin() / loadTelemetryPlugin(): remove from the discovered set on load (success or failure).
• unloadBackendPlugin() / unloadTelemetryPlugin(): put the name back into the discovered set so getAvailPlugins() still reports it after unload.

Python (_api.py):

• Remove the for plugin in self.plugin_list: getPluginParams(plugin) loop from nixl_agent.__init__.
• Add _load_plugin_params() helper that lazily populates the cache on first access.
• get_plugin_params() and get_plugin_mem_types() call the helper before returning.

No API changes. get_plugin_list() returns the same names. create_backend(), getPluginParams(), and instantiate_all=True all work identically. The dlopen just happens on demand instead of unconditionally at init.

Changed files

• src/api/python/_api.py (modified, +14/-4)
• src/core/nixl_plugin_manager.cpp (modified, +46/-11)
• src/core/plugin_manager.h (modified, +6/-0)