nixl re-sync 4-23-2026#29
Merged
edgargabriel merged 60 commits intoApr 28, 2026
Merged
Conversation
* NIXL EP: Add torch version check Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Pin torch version to 2.11 during wheel build Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Pin torch exactly to 2.11.* Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Add linter ignore Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Add copyright Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> --------- Signed-off-by: Ovidiu Mara <ovidium@nvidia.com>
* NIXL EP: Add torch version check Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Pin torch version to 2.11 during wheel build Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Pin torch exactly to 2.11.* Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Add linter ignore Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Add copyright Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> --------- Signed-off-by: Ovidiu Mara <ovidium@nvidia.com>
`FI_THREAD_COMPLETION` semantics require locked access to all objects bound to a completion queue, particularly endpoints. Expand the CQ mutex to cover posting operations to the endpoint as well. * In case we need to progress the CQ in retry while posting an operation, only do so if the progress thread is not enabled. Otherwise, let the progress thread handle CQ polling. Signed-off-by: Eric Raut <eraut@amazon.com>
* use ucx_ref and ignore torch ver in module * Remove file * fix * fix
…BJ plugin. (#1441) * Add Dell ObjectScale accelerated engine support to plugin_gtest for OBJ plugin. Introduce NIXL_OBJ_ENDPOINT_OVERRIDE environment variable for setting the endpoint of the Dell ObjectScale cluster. Add support for VRAM_SEG tests. Signed-off-by: Jason Goldschmidt <jason.goldschmidt@dell.com> * Extract CUOBJ tests to their own plugin test file. Make endpoint override a universaly applied environment variable to allow all tests to run against a specific endpoint * Addressed code review comments * Address code review comment regarding guard naming in memory_handler.h Signed-off-by: Jason Goldschmidt <jason.goldschmidt@dell.com> * Address code review comment Signed-off-by: Jason Goldschmidt <jason.goldschmidt@dell.com> * Addressed Code review comments Signed-off-by: Jason Goldschmidt <jason.goldschmidt@dell.com> --------- Signed-off-by: Jason Goldschmidt <jason.goldschmidt@dell.com> Co-authored-by: Adit Ranadive <aranadive@nvidia.com>
Invoke artifactory-cleanup-by-time.sh in pipeline_stop (dry-run) to list Docker images that would be removed by age. Use REGISTRY_REPO_NAME, REGISTRY_REPO_PATH, ARTIFACTORY_CLEANUP_AGE (default 3m) and existing Artifactory credentials. Also: split REGISTRY_REPO into REGISTRY_REPO_NAME and REGISTRY_REPO_PATH, set ARTIFACTORY_REGISTRY_URL for push/echo, add containerSelector for podman on relevant steps, and simplify taskName (drop axis_index). Signed-off-by: Daniel Pressler <danielpr@nvidia.com>
* NIXL/EP: Use vmm API instead of cudaMalloc * NIXL/EP: Use vmm API instead of cudaMalloc * NIXL/EP: revert * NIXL/EP: Support gdr copy with vmm. * NIXL/EP: Improve. * NIXL/EP: Format. * NIXL/EP: check return val. * NIXL/EP: Format. * NIXL/EP:Improve. * NIXL/EP: Format. * NIXL/EP: Improve. * NIXL/EP: Format. * NIXL/EP: fallback to cudaMalloc if fabric not supported * NIXL/EP: set default vals * NIXL/EP: Fix. * NIXL/EP: Fix comments. * NIXL/EP: Fix. * NIXL/EP: Fix. * NIXL/EP: not needed. * NIXL/EP: new files terms. * NIXL/EP: Fix comments. * NIXL/EP: Fix comment. * NIXL/EP: Fix. * NIXL/EP: fix. * NIXL/EP: Fix comment. * NIXL/EP: Fix comment. * NIXL/EP: Fix comment. * NIXL/EP: Fix comment. * NIXL/EP: Fix comment. --------- Co-authored-by: Ofir Farjon <ofarjon@funk11.mec01.labs.mlnx>
Move the p2p_ptr_get() call inside the is_rank_masked guard in three locations in nixl_ep.cu (two in dispatch, one in combine). In elastic scenarios, ranks can be set up in non-consecutive order — for example, ranks 0 and 2 may start before rank 1 has been initialized. During this early phase, rank 1's P2P memory mapping is not yet established. Previously, p2p_ptr_get() was called unconditionally for all ranks before the is_rank_masked check, so it would attempt to access the uninitialized P2P mapping for rank 1 and fail, even though rank 1 is correctly masked and should be skipped. In each of the three locations, the p2p_ptr_get(dst_ptr, dst_rank) call is moved to inside the rank mask guard, ensuring the P2P pointer lookup only occurs for ranks that are actually participating and initialized.
* nixl_ep: Add high-throughput kernels Port the DeepEP high-throughput dispatch and combine kernels to use the NIXL device API V2 (memory views, nixlPut, nixlAtomicAdd). Rename nixl_ep.cu -> nixl_ep_ll.cu Co-authored-by: Roey Azran <roeya@nvidia.com> Co-authored-by: Micha Dery <mdery@nvidia.com> Signed-off-by: Yoray Zack <yorayz@nvidia.com> * fixups pre-commit * fix pre-commit 2 * fix pre-commit3 --------- Signed-off-by: Yoray Zack <yorayz@nvidia.com> Co-authored-by: Roey Azran <roeya@nvidia.com> Co-authored-by: Micha Dery <mdery@nvidia.com>
* chore: update attributions for the latest release * fix missing license * update attr * fix dup * more changes * add missing license * add new pckage
PR #1258 changed two independent if-statements into if/else in registerMem(), breaking the fallthrough that calls cudaSetDevice() when the CUDA address workaround is disabled for device N>0, in multi-GPU per process case. Without cudaSetDevice(), cuMemGetAddressRange() fails inside EFA's dmabuf path for large buffers, causing fi_mr_key to return FI_KEY_NOTAVAIL. Blackwell (B200) requires the correct CUDA context for cuMemGetAddressRange; Hopper (H200) continues to work based on tests. Restore the original two-if pattern and downgrade the expected multi-GPU detection log from WARN to INFO. Validated with nixlbench --scheme tp --mode MG --num-initiator-dev 4 --num-target-dev 8 --check-consistency on p6 B200 instances. Signed-off-by: Ye Xiang <yexiang@amazon.com>
Signed-off-by: Raul Akhmetshin <rakhmetshin@nvidia.com>
…resolution (#1402) * BINDINGS/PYTHON/NIXL_META/NIXL: fix static analysis issues for nixl_agent and nixl_logger --------- Signed-off-by: youkaichao <youkaichao@gmail.com> Signed-off-by: Roie Danino <rdanino@nvidia.com> Co-authored-by: Roie Danino <rdanino@nvidia.com> Co-authored-by: Mikhail Brinskiy <brminich@users.noreply.github.com>
* Remove redundant count buffer * nixl_ep: Fix internode destruction flows * Re-guard p2p_ptr_get with is_rank_masked * Merge two !low_latency_mode blocks * Remove redundant condition
* Fix unbounded growth of in-memory telemetry event buffer The events_ vector grew without limit between periodic flushes. Enforce the configured buffer size (NIXL_TELEMETRY_BUFFER_SIZE) by dropping new events when the buffer is full (drop-newest policy). Signed-off-by: Batsheva Black <bblack@nvidia.com> * Reduce naming redundancy for telemetry buffer size Remove the buffer_size local variable in initializeTelemetry() and assign directly to maxBufferedEvents_. Use maxBufferedEvents_ instead of exporter_->getMaxEventsBuffered() in writeEventHelper() since they hold the same value. --------- Signed-off-by: Batsheva Black <bblack@nvidia.com>
elastic.py intentionally sends SIGTERM to simulate rank failure during the elastic test, but PR #1427 treats those expected exits as worker errors. Ignore planned SIGTERM exits when checking worker exit codes after the test completes.
* Integration build scripts for vllm Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Add copyright Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Remove CUDA version parameter, we can use CUDA_VERSION from base image instead Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Address AI comments Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> --------- Signed-off-by: Ovidiu Mara <ovidium@nvidia.com>
* Count all matches in LogProblemCounter * Fix version check for WarnWhenIbPresentButRdmaNotSupported * Use constexpr variable for ucp version * Exit after the loop if it matched any regex
* build: use Ninja + memory-aware parallelism for Mooncake build Upgrade Mooncake from v0.3.9 to v0.3.10.post1 and switch from Make to Ninja generator. v0.3.10.post1 includes limit_jobs.cmake (Mooncake PR#1718) which auto-detects available memory and creates Ninja job pools to cap compile and link parallelism separately. Previously the build hardcoded `make -j4` as a workaround for OOM during linking. With Ninja job pools, compilation can use all cores while memory-heavy link steps are automatically throttled — no manual -j tuning needed. Signed-off-by: staryxchen <staryxchen@tencent.com> * docs: update Mooncake version in ATTRIBUTIONS-CPP.md Bump Mooncake attribution from v0.3.9 to v0.3.10.post1 to match the version upgrade in .gitlab/build.sh. Signed-off-by: staryxchen <staryxchen@tencent.com> * build: disable Mooncake store module (nixl only needs transfer engine) Add -DWITH_STORE=OFF to skip building mooncake-store, which nixl does not use (the Mooncake plugin only links libtransfer_engine). This also works around a Pimpl compilation error in S3SnapshotObjectStore introduced in v0.3.10.post1 (see kvcache-ai/Mooncake#1796). Signed-off-by: staryxchen <staryxchen@tencent.com> * ci: bump CI_IMAGE_TAG to 20260402-1 Mooncake version and build flags changed (v0.3.10.post1, Ninja, -DWITH_STORE=OFF), so the CI image must be rebuilt. Signed-off-by: staryxchen <staryxchen@tencent.com> --------- Signed-off-by: staryxchen <staryxchen@tencent.com> Signed-off-by: ovidiusm <ovidium@nvidia.com> Co-authored-by: Mikhail Brinskiy <brminich@users.noreply.github.com> Co-authored-by: ovidiusm <ovidium@nvidia.com>
Signed-off-by: Raul Akhmetshin <rakhmetshin@nvidia.com>
Use raw pointers instead of dereferencing std::optional in ternary expressions, following the same pattern as recv_topk_idx_ptr and recv_topk_weights_ptr. This prevents GCC's -Wmaybe-uninitialized warning, which becomes a build error with -Werror. Made-with: Cursor
* nixlbench: fix poll() infinite loop, add SIGTERM handling
- Register SIGTERM handler in main.cpp alongside existing SIGINT handler
- Add !signaled() check to both do-while conditions in poll() so the loop
exits immediately when terminate is set
- Add checkLiveness() lambda in poll() that queries arePeersAlive() every 5s;
sets terminate=1 if peer's etcd key is gone, triggering the !signaled() break
- Thread &terminate through execTransfer() and execTransferIterations() so the
hot transfer loop breaks early on signal instead of running all N iterations
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
* nixlbench: add etcd lease-based peer liveness detection
Each rank attaches its presence key to an etcd v3 lease (TTL=15s). A
standalone etcd::KeepAlive object (independent gRPC channel) sends periodic
renewals. If the process dies (any signal including SIGKILL), the KeepAlive
thread dies, the gRPC connection closes, and etcd auto-deletes the key after
the TTL.
- runtime.h: add virtual arePeersAlive() default (returns true) to base class
- etcd_rt.h: add KeepAlive member + arePeersAlive() override declaration
- etcd_rt.cpp:
- setup(): attach rank key to lease via standalone KeepAlive constructor
(avoids sharing the main SyncClient gRPC channel)
- ~xferBenchEtcdRT(): Cancel()+reset() keepalive before rmdir() on clean exit
- arePeersAlive(): query each peer's rank key; detect absence via
resp.value().key().empty() (is_ok() returns true even for missing keys)
- Store etcd endpoints string for standalone KeepAlive constructor
- worker.cpp: use std::_Exit() instead of exit() in synchronize() to bypass
gRPC atexit handlers that deadlock with the KeepAlive background thread
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
* Fix clang
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
* Fix copyright
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
* nixlbench: address style review comments
- arePeersAlive(): add [[nodiscard]] attribute
- keepalive: change shared_ptr to unique_ptr (ownership is exclusive)
- LEASE_TTL_S: move from class header to anonymous namespace in .cpp,
rename to lease_ttl_s (ALL_CAPS reserved for macros)
- checkLiveness: use chrono_literals (5s) and direct duration comparison
instead of duration_cast + count(); add const to local 'now'
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
* nixlbench: address safety review comments
- terminate flag: change to std::atomic<int> to prevent data races between
the signal handler, poll() loop, and OpenMP transfer threads that all read
or write it concurrently
- destructor null guard: check client \!= nullptr before calling rmdir() in
~xferBenchEtcdRT(); setup() may fail after partial init, leaving client null
- cleanupForExit(): add best-effort namespace cleanup before _Exit() so the
non-leased 'size' key (and any others) are removed; a stale size key would
corrupt rank allocation in the next run of the same benchmark_group
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
* nixlbench: fix python_bindings.cpp for atomic<int> terminate
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
* nixlbench: remove redundant keepalive.reset() from destructor
unique_ptr destroys the held object automatically when the destructor
returns; the explicit reset() was a no-op. cleanupForExit() retains
reset() because it runs before _Exit() which skips destructors.
Update header guard.
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
* nixlbench: handle keepalive and rank-key registration failures in setup()
Wrap KeepAlive construction in try/catch and check the put() result.
Both paths unlock the etcd lock and return -1 on failure, preventing
setup() from reporting success with the rank unregistered.
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
* nixlbench: address coderabbit review comments
- python_bindings: replace atomic<int>* constructor binding with a
3-arg lambda factory; use raw pointer return so it works with all
pybind11 versions (unique_ptr return requires pybind11 >= 2.6)
- nixl_worker: add terminate check in recreate_per_iteration loop to
match the fast-exit behaviour already present in the standard path
- worker: add static_assert for lock-free atomic<int> and use explicit
fetch_add(1, relaxed) in signalHandler instead of prefix ++
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
* nixlbench: register leased rank key before bumping size counter
A crash between put("size") and put("rank/N", lease) left a stale
non-leased size key that poisoned subsequent runs. Register the
leased rank key first so a mid-setup crash leaves size unchanged;
only bump size once rank/<n> is successfully attached to the lease.
Also add terminate_ptr check inside execSingleTransfer() polling loop
so threads don't spin in getXferStatus() after SIGTERM is received.
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
* nixlbench: rename arePeersAlive and use consistent atomic ordering
- Rename arePeersAlive() -> areAllPeersAlive() for clarity
- Use load(memory_order_relaxed) consistently at all terminate_ptr
read sites instead of implicit seq_cst via operator*; use
store(memory_order_relaxed) for the liveness-check write
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
* nixlbench: use default memory ordering for terminate flag
Drop explicit memory_order_relaxed from all load()/store()/fetch_add()
calls on the terminate flag. The default seq_cst is fine for a flag
checked a few times per second; cleaner and more readable.
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
---------
Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
* libfabric: Fix Neuron mem type support regression Signed-off-by: Ye Xiang <yexiang@amazon.com> * feat(nixlbench): add Neuron VRAM support for non-CUDA builds Enable nixlbench VRAM segment allocation, deallocation, and consistency checks on Neuron (Trainium/Inferentia) devices when CUDA is not available. Signed-off-by: Ye Xiang <yexiang@amazon.com> * address review: do-while(0) wrap macro, add #else guard in cleanupBasicDescVram * address review: refactor nixlbench VRAM helpers --------- Signed-off-by: Ye Xiang <yexiang@amazon.com>
The gtest subdir was gated on ucx_dep.found() which checks for the system UCX library, not whether the UCX plugin is actually being built. On machines with UCX installed, this enters the gtest subdir even when UCX is not in enable_plugins, hitting an undefined ucx_backend_lib variable in the mock backend build. The mock backend has no UCX source dependency -- the link_with was unnecessary. Remove it and gate the gtest subdir on the UCX plugin being enabled rather than the system library being present. Signed-off-by: Ben Walker <ben@nvidia.com> Co-authored-by: Adit Ranadive <aranadive@nvidia.com>
…1523) * Build UCX and NIXL at the end to allow easy bisection with cached builds Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Remove pip install step Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Cleanup Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Add git log and clean whitespace Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Update help string Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Quote UCX_REF Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> --------- Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> Co-authored-by: Adit Ranadive <aranadive@nvidia.com>
Initialize status = NIXL_ERR_NOT_FOUND before the backend creation loop to silence g++ -Werror=maybe-uninitialized. Signed-off-by: Ben Walker <ben@nvidia.com>
… more instance types) (#1461)
Signed-off-by: Daniel Pressler <danielpr@nvidia.com>
* CI: added a new slurm cluster Signed-off-by: Iaroslav Sydoruk <isydoruk@nvidia.com> * Fixed python tests partial_md_example.py:119 unconditionally creates agents with enable_listen_thread=True and hardcoded default ports (5555 for target, 7777 for initiator), regardless of whether --etcd or socket mode is used ``` [2026-03-02T17:34:37.516Z] + python3 partial_md_example.py --etcd [2026-03-02T17:34:38.149Z] 2026-03-02 09:34:37 NIXL INFO partial_md_example.py:94 Using NIXL Plugins from: /opt/nixl/lib/aarch64-linux-gnu/plugins [2026-03-02T17:34:38.149Z] 2026-03-02 09:34:37 NIXL INFO partial_md_example.py:102 NIXL_ETCD_ENDPOINTS is set, using endpoints: http://127.0.0.1:59501 [2026-03-02T17:34:38.149Z] E0302 09:34:37.704217 464337 metadata_stream.cpp:84] Socket Bind failed while setting up listener for MD: Address already in use [98] ``` Signed-off-by: Iaroslav Sydoruk <isydoruk@nvidia.com> * restore options Signed-off-by: Iaroslav Sydoruk <isydoruk@nvidia.com> --------- Signed-off-by: Iaroslav Sydoruk <isydoruk@nvidia.com>
This reverts commit 7c87ff5.
* NIXLBENCH: Runtime without ETCD. * NIXLBENCH: Fix formatting. * NIXLBENCH: Review comments. * NIXLBENCH: Bump tag. * Apply suggestion from @guy-ealey-morag Co-authored-by: Guy Ealey Morag <guymorag142@gmail.com> Signed-off-by: Colin Hirsch <chirsch@nvidia.com> * NIXLBENCH: Review comments. * NIXLBENCH: Fix formatting. * Factor out constants. * Bump image tags. * Add ASIO fallback. * Fix test run. * Fix test run. * Fix formatting. * Remove libasio-dev install again. * Catch listen error. * Fix image tag. * NIXLBENCH: Review comments. * Restore CI_IMAGE_TAG. --------- Signed-off-by: Colin Hirsch <chirsch@nvidia.com>
…470) * telemetry: share Prometheus Exposer across agents in same process Each agent tried to create its own prometheus::Exposer HTTP server on the same port (default 9090), causing the second agent to fail with EADDRINUSE. The second agent's telemetry was silently disabled. Share a single Exposer via a process-wide weak_ptr so the first agent creates the HTTP server and subsequent agents reuse it. Each agent still owns its own Registry with per-agent labels, so metrics remain correctly attributed. The Exposer is destroyed when the last agent's exporter is released. Signed-off-by: Batsheva Black <bblack@nvidia.com> * telemetry: share single Prometheus Registry across all agents Refactor from per-agent registries to a single shared Registry so that all metric families live in one place, producing cleaner /metrics output. Each agent's metrics are still distinguished by the agent_name label. A destructor removes agent-specific metrics from the shared Registry when an agent is destroyed. * fix clang * fix duplicate agent names * telemetry: fix header guard naming and reject duplicate agent names Rename the include guard to NIXL_SRC_PLUGINS_TELEMETRY_PROMETHEUS_EXPORTER_H to follow the repo convention. Change duplicate agent_name handling from a warning to throwing std::runtime_error so the exporter is never constructed for a duplicate, preventing use-after-free on destruction. * resolve comment * telemetry: address PR review feedback * telemetry: use RAII destructors for CounterEntry/GaugeEntry cleanup Move family->Remove(metric) into ~CounterEntry and ~GaugeEntry, then replace explicit loops in the destructor with counters_.clear() and gauges_.clear() under the lock. * fix clang --------- Signed-off-by: Batsheva Black <bblack@nvidia.com> Co-authored-by: ovidiusm <ovidium@nvidia.com>
Signed-off-by: Ovidiu Mara <ovidium@nvidia.com>
* telemetry: replace string event names with enum type Replace per-event char[32] eventName_ with a uint8_t enum nixl_telemetry_event_type_t, avoiding string copies on every telemetry event and enabling compile-time type checking. - Add nixl_telemetry_event_type_t enum with all known event types - Add constexpr telemetryEventTypeStr() returning string_view - Rename eventName_ to eventType_ for clarity - Update all callers to use enum values instead of string literals - Update Python/C++ readers and Prometheus exporter accordingly - Move Python lookup dicts to module scope - Remove unused cstring include Signed-off-by: Batsheva Black <bblack@nvidia.com> * fix clang * telemetry: add TODO for C++20 string_view map lookup in Prometheus exporter. * Update src/core/telemetry/telemetry.cpp Co-authored-by: Guy Ealey Morag <guymorag142@gmail.com> Signed-off-by: BatshevaBlack <132911331+BatshevaBlack@users.noreply.github.com> * Update src/core/telemetry/telemetry_event.h Co-authored-by: Guy Ealey Morag <guymorag142@gmail.com> Signed-off-by: BatshevaBlack <132911331+BatshevaBlack@users.noreply.github.com> * Update src/core/telemetry/telemetry_event.h Co-authored-by: Guy Ealey Morag <guymorag142@gmail.com> Signed-off-by: BatshevaBlack <132911331+BatshevaBlack@users.noreply.github.com> * telemetry: per-status error event types * resolve comments * telemetry: assert error paths; move status→event mapping to .cpp - Declare nixlTelemetryEventTypeForStatus in telemetry_event.h; define in telemetry.cpp with explicit NIXL_SUCCESS/NIXL_IN_PROG and NIXL_ASSERT_ALWAYS on unhandled values (addresses review: do not silently drop new error codes). - updateErrorCount: NIXL_ASSERT_ALWAYS for negative status and for missing mapping instead of ignoring empty optional. * fix_clang * Update src/core/telemetry/telemetry.cpp Co-authored-by: Guy Ealey Morag <guymorag142@gmail.com> Signed-off-by: BatshevaBlack <132911331+BatshevaBlack@users.noreply.github.com> * Update src/core/telemetry/telemetry.cpp Co-authored-by: Guy Ealey Morag <guymorag142@gmail.com> Signed-off-by: BatshevaBlack <132911331+BatshevaBlack@users.noreply.github.com> * resolve comment * telemetry: handle NIXL_SUCCESS/IN_PROG in status→event switch Explicit cases satisfy -Wswitch under -Werror and assert if nixlTelemetryEventTypeForStatus is called with a non-error status. * fix clang --------- Signed-off-by: Batsheva Black <bblack@nvidia.com> Signed-off-by: BatshevaBlack <132911331+BatshevaBlack@users.noreply.github.com> Co-authored-by: Guy Ealey Morag <guymorag142@gmail.com> Co-authored-by: ovidiusm <ovidium@nvidia.com>
…559) * Use correct synchronization with etcd metadata comm thread Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Address style comments Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Refactor to reduce code duplication Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Style changes Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * clang-format Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> --------- Signed-off-by: Ovidiu Mara <ovidium@nvidia.com>
…y used (#1546) * Defer plugin load to the point where they are actually used Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Fix unit tests Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Simplify logic Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Fix race due to incorrect sync mode with etcd metadata Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Fix clang format Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Try RW to see perf difference Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Switch back to strict locking Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Simplify code Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Split etcd fix into separate PR Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Revert unrelated change Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Code cleanup Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Clean up logs Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Clang-format Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Simplify python code Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Remove dead code from telemetry Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Const locks Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Add unit test for deferred load/unload flow Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> * Fix copyright Signed-off-by: Ovidiu Mara <ovidium@nvidia.com> --------- Signed-off-by: Ovidiu Mara <ovidium@nvidia.com>
* PLUGINS/POSIX: Remove dangling pointer. * PLUGINS/POSIX: Fix includes.
Signed-off-by: Yoray Zack <yorayz@nvidia.com>
Signed-off-by: Guy Ealey Morag <gealeymorag@nvidia.com>
gaoikawa
approved these changes
Apr 25, 2026
edgargabriel
force-pushed
the
topic/nixl-sync-4-23-2026
branch
2 times, most recently
from
April 27, 2026 16:35
583a29b to
5a475b9
Compare
add a cleaner solution for some of the CUDA/HIP incompatilibities, not just disable -Werror
edgargabriel
force-pushed
the
topic/nixl-sync-4-23-2026
branch
from
April 27, 2026 16:38
5a475b9 to
39be1de
Compare
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
20 participants
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.
What?
Periodic resync with upstream repo