PortChannel: add optional GPUNetIO backend with pinned FetchContent dependency by seagater · Pull Request #910 · microsoft/mscclpp

Qinghua Zhou (seagater) · 2026-09-22T19:29:23Z

Summary

Add an optional GPU-initiated RDMA backend (EndpointConfig::Ib::Mode::Device, backed by DOCA GPUNetIO / GDAKI) to PortChannel.
Fetch the official NVIDIA-DOCA/gpunetio dependency via CMake FetchContent, pinned to commit c5d97fd52346763bcfeaff593dd1d84215156218. Known upstream CQ error-path issues are tracked in NVIDIA-DOCA/gpunetio#23.
Template channels and device handles on PortChannelBackend (BasePortChannel<Backend>, PortChannel<Backend>, BasePortChannelDeviceHandleT<Backend, SmemSlots>, and PortChannelDeviceHandleT<Backend, SmemSlots> with Host and Device specializations, with BasePortChannel and PortChannel defaulting to PortChannelBackend::Host, plus non-templated aliases BasePortChannelDeviceHandle and PortChannelDeviceHandle).
Construct Device-mode PortChannel directly from std::shared_ptr<PortDevice2DeviceSemaphore> (or Semaphore) and RegisteredMemory (PortChannel channel(semaphore, dst, src)) without ProxyService, and obtain the device handle via channel.deviceHandle() (or channel.deviceHandle<SmemSlots>()).
Signal and accumulate with RDMA atomic fetch-and-add: the responder applies an atomic only after placing earlier RDMA writes on the same queue pair, so observing a signal guarantees the preceding data is visible. putWithAccumulate posts a write and an atomic together (putWithSignal uses it on the semaphore).
Support queue-pair concurrency scopes via PortChannelDeviceHandle::Concurrency (Device by default, Block, Exclusive). Exclusive operations submit through BlueFlame up to blueFlameLimit (default kDefaultBlueFlameLimit = 512 KiB) using internal 64-byte shared-memory slots (SmemSlots, default 32 per CTA); larger transfers and the shared Block/Device modes ring the regular doorbell (submit_db).

Build and Use

GPUNetIO is built automatically whenever MSCCLPP_USE_CUDA=ON, MSCCLPP_USE_ROCM=OFF, and MSCCLPP_USE_IB=ON (the default), and skipped for ROCm or -DMSCCLPP_USE_IB=OFF builds:

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DMSCCLPP_BUILD_TESTS=ON
cmake --build build -j$(nproc)

For an offline GPUNetIO checkout:

cmake -S . -B build -DFETCHCONTENT_SOURCE_DIR_GPUNETIO=/absolute/path/to/gpunetio

Usage Example

mscclpp::EndpointConfig cfg;
cfg.transport = mscclpp::Transport::IB0;
cfg.ib.mode = mscclpp::EndpointConfig::Ib::Mode::Device;
auto connection = communicator->connect(cfg, remoteRank).get();

auto localMem = communicator->registerMemory(buff, bytes, mscclpp::Transport::IB0);
communicator->sendMemory(localMem, remoteRank);
auto remoteMem = communicator->recvMemory(remoteRank).get();

auto semaphore = std::make_shared<mscclpp::PortDevice2DeviceSemaphore>(*communicator, connection);
mscclpp::PortChannel channel(semaphore, remoteMem, localMem);
auto deviceHandle = channel.deviceHandle();

In device code:

__global__ void send(mscclpp::PortChannelDeviceHandleT<mscclpp::PortChannelBackend::Device> channel, size_t bytes) {
  using Concurrency = mscclpp::PortChannelDeviceHandle::Concurrency;
  if (threadIdx.x != 0) return;
  channel.putWithSignal<Concurrency::Exclusive>(0, bytes);
  channel.flush<Concurrency::Exclusive>();
}

See docs/guide/gpunetio.md for full details.

Validation

# Unit tests
./build/bin/unit_tests --exclude-perf-tests

# Multi-process GPUNetIO (IbDeviceMode) unit tests
mpirun -np 2 ./build/bin/mp_unit_tests --filter=IbDeviceMode --exclude-perf-tests

# Multi-process GPUNetIO (IbDeviceMode) performance tests
mpirun -np 2 ./build/bin/mp_unit_tests --filter=IbDeviceMode --only-perf-tests

Performance: GPUNetIO (`IbMode::Device`) vs. IB CPU Proxy (`IbMode::Host` / `IbMode::HostNoAtomic`)

Measured with mp_unit_tests (PortChannelOneToOneTest) on 2x H100 GPUs over InfiniBand (kDefaultBlueFlameLimit = 512 KiB).

One Thread per QP (`Concurrency::Exclusive`, BlueFlame)

Benchmark	IB Proxy (`Host`)	IB Proxy (`HostNoAtomic`)	GPUNetIO (`Exclusive`)	Improvement vs. `Host`
PingPong latency (`putWithSignal` + `wait` + `flush`)	22.34 µs/iter	22.88 µs/iter	18.45 µs/iter	-17.4% latency
PacketPingPong (LL) latency	6.34 µs/iter	6.51 µs/iter	4.50 µs/iter	-29.1% latency
Bidirectional bandwidth (1 KB)	0.0605 GB/s	0.0573 GB/s	0.0698 GB/s	+15.3%
1 QP bandwidth (1 KB)	0.0598 GB/s	0.0582 GB/s	0.0715 GB/s	+19.6%
4 QPs bandwidth (4 KB total, 1 KB/QP)	0.2123 GB/s	0.1919 GB/s	0.2733 GB/s	+28.7%
4 QPs flush stress (1 KB/QP) latency	23.36 µs/iter	24.35 µs/iter	18.66 µs/iter	-20.1% latency
4 QPs flush stress (1 KB/QP) aggregate BW	0.1753 GB/s	0.1682 GB/s	0.2196 GB/s	+25.3%

Message Size	IB Proxy (`Host`)	IB Proxy (`HostNoAtomic`)	GPUNetIO (`Exclusive`)
1 QP, 64 KB	3.34 GB/s	3.26 GB/s	3.82 GB/s
1 QP, 1 MB	25.85 GB/s	25.49 GB/s	25.87 GB/s
1 QP, 4 MB	40.01 GB/s	39.77 GB/s	39.99 GB/s
1 QP, 128 MB	48.21 GB/s	48.22 GB/s	48.21 GB/s
4 QPs, 256 KB (64 KB/QP)	10.95 GB/s	10.81 GB/s	12.30 GB/s
4 QPs, 4 MB (1 MB/QP)	40.40 GB/s	40.15 GB/s	40.45 GB/s
4 QPs, 512 MB (128 MB/QP)	49.19 GB/s	49.18 GB/s	49.18 GB/s

Shared QPs (`Concurrency::Block` / `Device`, regular doorbell)

The proxy runs the same kernels with the same number of threads per QP.

Benchmark	IB Proxy (`Host`)	GPUNetIO (`Block`)	GPUNetIO (`Device`)
PingPong latency (1 thread)	22.34 µs/iter	23.99 µs/iter	24.60 µs/iter
PacketPingPong (LL) latency (1 thread)	6.34 µs/iter	7.26 µs/iter	7.49 µs/iter
Same-channel concurrent flush (4 threads, 1 QP)	29.35 µs/iter	29.09 µs/iter	29.00 µs/iter
4 QPs bandwidth, 2 threads/QP (4 KB total)	0.1686 GB/s	0.1588 GB/s	0.1549 GB/s
4 QPs bandwidth, 2 threads/QP (256 KB total)	9.31 GB/s	7.94 GB/s	7.80 GB/s
4 QPs bandwidth, 2 threads/QP (4 MB total)	39.08 GB/s	37.33 GB/s	37.11 GB/s
4 QPs bandwidth, 2 threads/QP (512 MB total)	49.16 GB/s	49.14 GB/s	49.14 GB/s
4 QPs flush stress, 2 threads/QP (1 KB/QP) latency	28.14 µs/iter	29.54 µs/iter	30.30 µs/iter
4 QPs flush stress, 2 threads/QP (256 KB/QP) latency	46.99 µs/iter	50.97 µs/iter	51.52 µs/iter

In Block and Device modes, each posting thread updates the shared queue pair with GPU atomics and fences, and submits without BlueFlame. For the lowest latency, give each posting thread its own QP in Exclusive mode.

Port the generic integration and safety fixes from qinghuazhou/gpunetio_port_channel_merge_feature_ep onto main, preserving CPU-proxy defaults and accumulate semantics. Exclude all EP kernels and policy. Pin the unmodified NVIDIA-DOCA/gpunetio v4.0.1 submodule. Add public host channel construction, CPU regressions, two-rank smoke coverage and installation instructions. ON/OFF CUDA builds, six CPU checks, installed-header compilation and lint pass. GPU/NIC runtime remains unverified; documented upstream allocation and CPU-service shutdown issues block merging.

Port GpuNetIoMultiQpBandwidth from the EP integration branch with configurable single-HCA QPs, collective geometry validation, queue-indexed device operations and per-QP atomic scratch. Preserve single-QP defaults and add CPU queue-isolation regression coverage. Validated GPUNetIO ON/OFF builds, seven CPU test runs and repository lint. GPU/NIC benchmark execution remains pending; the upstream submodule is unchanged.

azure-pipelines · 2026-09-22T19:29:33Z

Azure Pipelines: There may be pipelines that require an authorized user to comment /azp run to run.

Replace the dense-only upstream flatten helper with a checked sparse table adapter that keeps null self-rank slots zero and preserves remote peer/QP indices. Own the uploaded CUDA table in GpuNetIoService without modifying the upstream submodule. Add actual-upstream-type regression coverage for 28 layouts and 296 rejected inputs. Commit the exact patch verified by the user; ON/OFF builds, CPU regressions and lint were validated separately.

Copilot

Copilot review overview

🔵 Needs a closer look

Device completion and peer-validation defects remain alongside acknowledged upstream blockers and unverified GPU/NIC runtime behavior.

Review effort: Balanced
Findings: 1 · 3 · 3

Open (7)

Reject self and out-of-range peer ranks during construction · New
Validate peer before indexing peer arrays · New
Release builds discard EBUSY instead of completing bounded flush · New
Avoid enabling backend with broken pinned upstream implementation · New
Add required MIT license header to public C++ header · New
Add required MIT license header to C++ header · New
Add required MIT license header to C++ test file · New

What changed in this PR

Adds an optional GPUNetIO/GDAKI backend to PortChannel while retaining CPU proxy behavior by default.

Changes:

Adds collective GPUNetIO service setup and multi-QP device operations.
Integrates the upstream GPUNetIO submodule and build/install configuration.
Adds CPU-stub, P2P, and bandwidth tests plus usage documentation.

File	Description
`.gitmodules`	Registers the GPUNetIO submodule.
`CMakeLists.txt`	Adds and validates the build option.
`docs/quickstart.md`	Documents optional backend setup.
`include/mscclpp/gpu_net_io_service.hpp`	Declares the host service API.
`include/mscclpp/internal/port_channel_gpunetio_device_impl.hpp`	Implements device RDMA operations.
`include/mscclpp/port_channel.hpp`	Adds GPUNetIO channel construction.
`include/mscclpp/port_channel_device.hpp`	Routes device operations by backend.
`include/mscclpp/port_channel_gpunetio_device.hpp`	Defines the device context API.
`src/CMakeLists.txt`	Includes the optional component.
`src/core/CMakeLists.txt`	Links GPUNetIO objects.
`src/core/include/ib.hpp`	Exposes internal verbs handles.
`src/core/port_channel.cc`	Constructs GPUNetIO channels.
`src/gpunetio/CMakeLists.txt`	Builds and installs the dependency.
`src/gpunetio/PROVENANCE.md`	Documents provenance and limitations.
`src/gpunetio/host/gpu_net_io_qp_table.hpp`	Builds sparse device QP tables.
`src/gpunetio/host/gpu_net_io_service.cpp`	Sets up QPs, memory, and device state.
`src/gpunetio/vendor`	Pins the upstream GPUNetIO dependency.
`test/mp_unit/CMakeLists.txt`	Enables GPUNetIO MP tests.
`test/mp_unit/port_channel_tests.cu`	Adds P2P and bandwidth coverage.
`test/unit/CMakeLists.txt`	Registers focused unit tests.
`test/unit/gpunetio_channel_test.cc`	Tests routing, QPs, and host API behavior.

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

Refactor flush logic to handle errors from tryFlush. Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

Construct GPUNetIO host channels through the service to validate peers against retained rank/world-size metadata and reject incomplete setup. Restrict QP creation to direct GPU doorbells, valid DBRs and GPU-resident CQs; remove CPU progress fallback and exchange QP failures before connection-info collectives. Add host peer-validation and direct-only policy regression coverage. Preserve the unmodified upstream submodule. ON/OFF builds, eight CPU test runs and repository lint pass; direct-GPU hardware qualification remains separate.

Replace the Git submodule with ON-only FetchContent pinned to the full upstream v4.0.1 SHA. Share dependency include paths through an interface target and retain offline overrides and installed headers. Remove unused DO_PRAGMA and NVCC_PRAGMA_UNROLL compatibility definitions. Add CUDA compile probes for clean and caller-defined macro environments. Validate ON/OFF builds, eight CPU tests, disconnected overrides, installed headers, and lint.

Expose explicit connection handles for existing QPs, independently registered local/remote memory, and owned semaphore counters exchanged per connection. Preserve per-channel memory IDs and use the bound QP for writes, signals, atomics and flushes while retaining resource lifetimes. Migrate P2P and multi-QP bandwidth tests to the public binding API and add two-channel isolation coverage. ON/OFF builds, eight CPU tests, installed-header checks and lint pass. GPU/NIC execution remains pending; uniform collective QP setup is retained.

Rename gin-prefixed device members, constructor parameters and helpers to gpuNetIo and update their test references. Remove NCCL-specific comparisons from public backend comments. No behavior or layout changes. Validated naming-only equivalence, ON/OFF builds, eight CPU tests and repository lint.

Copilot

Copilot review overview

🔵 Needs a closer look

The GPU/NIC transport remains hardware-unverified, and the review found documentation, licensing, and build-diagnostic issues.

Review effort: Balanced
Findings: 2 · 3

Open (5)

Update PR metadata and build instructions for FetchContent · New
Restrict -w to vendored sources · New
Add required Microsoft MIT license header · New
Add required Microsoft MIT license header · New
Update comment for compact qpHl indexing · New

Resolved since last review (7)

Reject self and out-of-range peer ranks during construction
Avoid enabling backend with broken pinned upstream implementation
Release builds discard EBUSY instead of completing bounded flush
Validate peer before indexing peer arrays
Add required MIT license header to C++ test file
Add required MIT license header to C++ header
Add required MIT license header to public C++ header

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

Copilot

🟡 Changes recommended

Unresolved issues can silently disable Device-mode transfers and hang peer waits for installed-header consumers, while missing optional mlx5 headers can break otherwise supported IB builds.

5 open issues

5 resolved since last review

_{Review effort: Balanced}

GPUNetIO queue pairs are created on an IB device's protection domain, so the DOCA device and per-GPU handles opened on it now belong to that IbCtx (IbCtx::gpuNetIoDevice(), created on first use and released before the protection domain). Context no longer keeps a separate map of GPUNetIO contexts. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 95bbe000-7818-4c60-8184-773c93680b01

Copilot

🔵 Needs a closer look

The change breaks CMake generation when optional MLX5 headers are missing and allocates semaphore counters on the wrong GPU when the current device differs from the connection’s device, risking illegal memory access.

5 open issues

_{Review effort: Balanced}

Configure per-CTA WQE capacity through device-handle templates and accept runtime slot indices with doorbell fallback. Remove caller-owned shared-memory binding and update existing callers and documentation. Drop the local upstream CQ cleanup patch and track its error-handling issues upstream. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 55ef1b7e-0ccb-4b59-ad58-48a168608302

Copilot

🟡 Changes recommended

Unresolved critical defects include a double-free path, wrong-GPU memory allocation, silent communication no-ops for installed consumers, and a polling race that can consume the same signal twice.

9 open issues

_{Review effort: Balanced}

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 55ef1b7e-0ccb-4b59-ad58-48a168608302

Copilot

🟡 Changes recommended

Unresolved issues can break builds without optional MLX5 headers and silently turn GPUNetIO device operations into no-ops for installed-library consumers.

10 open issues

_{Review effort: Balanced}

…elDeviceHandle Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 55ef1b7e-0ccb-4b59-ad58-48a168608302

Copilot

🔵 Needs a closer look

Unresolved critical defects allocate semaphore state on the wrong GPU, risking illegal access or hangs, and break CUDA+IB configuration when optional MLX5 headers are absent.

10 open issues

_{Review effort: Balanced}

…eHandle Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 55ef1b7e-0ccb-4b59-ad58-48a168608302

Copilot

🟡 Changes recommended

Unresolved critical issues can break CMake generation, allocate semaphore counters on the wrong GPU, and cause installed consumers to compile assertion-only GPUNetIO stubs.

11 open issues

_{Review effort: Balanced}

…_IBVERBS Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 55ef1b7e-0ccb-4b59-ad58-48a168608302

Copilot

🟡 Changes recommended

Unresolved defects route downstream device kernels to disabled stubs, allow concurrent pollers to consume the same signal twice, and can double-free CQ buffers on initialization failure.

2 open issues

11 resolved since last review

_{Review effort: Balanced}

+#include "atomic_device.hpp"
+#endif
+
+#if defined(MSCCLPP_DEVICE_CUDA) && defined(USE_IBVERBS)


+  MSCCLPP_DOCA_THROW(doca_verbs_qp_attr_set_ack_timeout(attr, 18));
+  MSCCLPP_DOCA_THROW(doca_verbs_qp_attr_set_retry_cnt(attr, 7));
+  MSCCLPP_DOCA_THROW(doca_verbs_qp_attr_set_rnr_retry(attr, 7));
+  MSCCLPP_DOCA_THROW(doca_verbs_qp_attr_set_max_rd_atomic(attr, 1));


…hannel Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 55ef1b7e-0ccb-4b59-ad58-48a168608302

Qinghua Zhou (seagater) added 2 commits September 22, 2026 01:38

Qinghua Zhou (seagater) requested review from Binyang Li (Binyang2014) and a balanced review from Copilot September 23, 2026 00:28

Copilot started reviewing on behalf of Qinghua Zhou (seagater) September 23, 2026 00:30

View session

Copilot AI reviewed Sep 23, 2026

View reviewed changes

Binyang Li (Binyang2014) reviewed Sep 23, 2026

View reviewed changes

Comment thread .gitmodules Resolved

Comment thread include/mscclpp/internal/port_channel_gpunetio_device_impl.hpp Resolved

Comment thread include/mscclpp/port_channel_device.hpp Resolved

Comment thread include/mscclpp/port_channel.hpp Resolved

Binyang Li (Binyang2014) reviewed Sep 23, 2026

View reviewed changes

Comment thread src/gpunetio/host/gpu_net_io_service.cpp Resolved

Qinghua Zhou (seagater) and others added 10 commits September 23, 2026 12:20

Add licensing comments to gpu_net_io_service.hpp

c30012e

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

Add license comments to gpu_net_io_qp_table.hpp

746fe7b

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

Update copyright notice in gpunetio_channel_test.cc

945e16b

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

Refactor put method to use local qp variable

ac206ed

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

Improve error handling in flush operation

eba4933

Refactor flush logic to handle errors from tryFlush. Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

test(gpunetio): add LL ping-pong correctness and latency tests

b11bccc

Binyang Li (Binyang2014) reviewed Sep 25, 2026

View reviewed changes

Comment thread src/gpunetio/host/gpu_net_io_service.cpp Resolved

feat(gpunetio): support sparse per-peer connection plans

55c40d8

Copilot AI review requested due to automatic review settings September 25, 2026 21:23

Copilot started reviewing on behalf of Qinghua Zhou (seagater) September 25, 2026 21:23

View session

Mahdieh Ghazi (mahdiehghazim) reviewed Sep 25, 2026

View reviewed changes

Comment thread src/gpunetio/host/gpu_net_io_service.cpp Resolved

Comment thread include/mscclpp/gpunetio_device.hpp Resolved

Comment thread src/gpunetio/host/gpu_net_io_service.cpp Resolved

Copilot AI reviewed Sep 25, 2026

View reviewed changes

Update compile options and definitions for gpunetio

d6d9090

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

Copilot AI review requested due to automatic review settings September 25, 2026 23:04

Qinghua Zhou (seagater) and others added 2 commits September 25, 2026 16:04

Add license comments to gpu_net_io_binding.hpp

abc8559

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

Add license comments to gpu_net_io_policy.hpp

45477a9

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

Copilot started reviewing on behalf of Binyang Li (Binyang2014) November 4, 2026 17:53

View session

Copilot AI reviewed Nov 4, 2026

View reviewed changes

Comment thread cmake/GpuNetIo.cmake Resolved

Comment thread include/mscclpp/gpunetio_device.hpp Resolved

Comment thread src/core/gpunetio.cc Resolved

Comment thread src/core/port_channel.cc Resolved

Comment thread src/core/semaphore.cc Resolved

Copilot AI balanced review requested due to automatic review settings November 4, 2026 19:04

Copilot started reviewing on behalf of Qinghua Zhou (seagater) November 4, 2026 19:05

View session

Copilot AI reviewed Nov 4, 2026

View reviewed changes

Changho Hwang (chhwang) reviewed Nov 6, 2026

View reviewed changes

Copilot AI balanced review requested due to automatic review settings November 7, 2026 00:09

Copilot started reviewing on behalf of Qinghua Zhou (seagater) November 7, 2026 00:09

View session

Copilot AI reviewed Nov 7, 2026

View reviewed changes

Comment thread include/mscclpp/gpunetio_device.hpp Resolved

Comment thread include/mscclpp/semaphore_device.hpp Resolved

Comment thread src/core/CMakeLists.txt Resolved

Comment thread src/core/gpunetio.cc Resolved

Nest PortChannel concurrency and align host backend APIs

b9d7439

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 55ef1b7e-0ccb-4b59-ad58-48a168608302

Copilot AI balanced review requested due to automatic review settings November 7, 2026 01:23

Copilot started reviewing on behalf of Qinghua Zhou (seagater) November 7, 2026 01:24

View session

Copilot AI reviewed Nov 7, 2026

View reviewed changes

Comment thread cmake/GpuNetIo.cmake Resolved

Template PortChannel on backend and move Concurrency to BasePortChann…

af6460a

…elDeviceHandle Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 55ef1b7e-0ccb-4b59-ad58-48a168608302

Copilot AI balanced review requested due to automatic review settings November 7, 2026 02:54

Copilot started reviewing on behalf of Qinghua Zhou (seagater) November 7, 2026 02:55

View session

Copilot AI reviewed Nov 7, 2026

View reviewed changes

Rename MaxWqeSlots/wqeSlotIdx to SmemSlots/slotIdx and simplify devic…

b050f90

…eHandle Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 55ef1b7e-0ccb-4b59-ad58-48a168608302

Copilot AI balanced review requested due to automatic review settings November 7, 2026 06:31

Copilot started reviewing on behalf of Qinghua Zhou (seagater) November 7, 2026 06:31

View session

Copilot AI reviewed Nov 7, 2026

View reviewed changes

Comment thread include/mscclpp/port_channel.hpp Resolved

Validate GPUNetIO memory transport, guard MLX5 include, and reuse USE…

d8b6fae

…_IBVERBS Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 55ef1b7e-0ccb-4b59-ad58-48a168608302

Copilot AI balanced review requested due to automatic review settings November 9, 2026 02:00

Copilot started reviewing on behalf of Qinghua Zhou (seagater) November 9, 2026 02:01

View session

Copilot AI reviewed Nov 9, 2026

View reviewed changes

Rename BasePortChannelT and PortChannelT to BasePortChannel and PortC…

1cd9671

…hannel Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 55ef1b7e-0ccb-4b59-ad58-48a168608302

Copilot AI balanced review requested due to automatic review settings November 9, 2026 02:29

Copilot started reviewing on behalf of Qinghua Zhou (seagater) November 9, 2026 02:30

View session

Conversation

Qinghua Zhou (seagater) commented Sep 22, 2026 • edited by Binyang2014 Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Build and Use

Usage Example

Validation

Performance: GPUNetIO (IbMode::Device) vs. IB CPU Proxy (IbMode::Host / IbMode::HostNoAtomic)

One Thread per QP (Concurrency::Exclusive, BlueFlame)

Shared QPs (Concurrency::Block / Device, regular doorbell)

Uh oh!

azure-pipelines Bot commented Sep 22, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🔵 Needs a closer look

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🔵 Needs a closer look

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🔵 Needs a closer look

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🔵 Needs a closer look

Uh oh!

Qinghua Zhou (seagater) commented Sep 22, 2026 •

edited by Binyang2014

Loading

Performance: GPUNetIO (`IbMode::Device`) vs. IB CPU Proxy (`IbMode::Host` / `IbMode::HostNoAtomic`)

One Thread per QP (`Concurrency::Exclusive`, BlueFlame)

Shared QPs (`Concurrency::Block` / `Device`, regular doorbell)