diff --git a/include/mscclpp/utils.hpp b/include/mscclpp/utils.hpp index ffe269da..54cfa4a0 100644 --- a/include/mscclpp/utils.hpp +++ b/include/mscclpp/utils.hpp @@ -46,6 +46,15 @@ std::string getIBDeviceName(Transport ibTransport); /// @return The InfiniBand transport associated with the specified device name. Transport getIBTransportByDeviceName(const std::string& ibDeviceName); +/// Check whether this process can allocate/import CUDA memory with NVIDIA fabric handles +/// (`CU_MEM_HANDLE_TYPE_FABRIC`). Fabric handles enable cross-node `Transport::CudaIpc` on +/// MNNVL systems (e.g., GB200 NVL72) when the IMEX service is running. Returns `false` on +/// hardware/software stacks without MNNVL+IMEX, in which case `Transport::CudaIpc` is +/// restricted to ranks within the same node. +/// +/// @return `true` if fabric handles are usable from this process, `false` otherwise. +bool isFabricMemHandleAvailable(); + } // namespace mscclpp #endif // MSCCLPP_UTILS_HPP_ diff --git a/src/core/connection.cc b/src/core/connection.cc index 8b6c0afb..11ecd968 100644 --- a/src/core/connection.cc +++ b/src/core/connection.cc @@ -82,6 +82,17 @@ MSCCLPP_API_CPP int Connection::getMaxWriteQueueSize() const { return impl_->get CudaIpcConnection::CudaIpcConnection(std::shared_ptr context, const Endpoint& localEndpoint, const Endpoint& remoteEndpoint) : BaseConnection(context, localEndpoint) { + // Log fabric/MNNVL availability exactly once per process so any later cross-node CudaIpc failure + // is easy to triage. C++11 magic statics make this thread-safe without an explicit mutex. + // NOTE: assigning the message to a std::string first avoids the logger's pointer-formatting + // overload from kicking in on the const char* result of the ternary. + [[maybe_unused]] static const bool fabricAvailable_ = []() { + const bool avail = isFabricMemHandleAvailable(); + const std::string status = avail ? "available (cross-node CudaIpc via MNNVL/IMEX is supported)" + : "NOT available (CudaIpc is restricted to intra-node ranks on this system)"; + INFO(CONN, "CudaIpc transport selected: fabric handles ", status); + return avail; + }(); if (localEndpoint.transport() != Transport::CudaIpc || remoteEndpoint.transport() != Transport::CudaIpc) { THROW(CONN, Error, ErrorCode::InternalError, "CudaIpc transport is required for CudaIpcConnection"); } diff --git a/src/core/context.cc b/src/core/context.cc index aabe71df..b55939e3 100644 --- a/src/core/context.cc +++ b/src/core/context.cc @@ -4,12 +4,11 @@ #include "context.hpp" #include -#include #include "api.h" #include "connection.hpp" -#include "debug.h" #include "endpoint.hpp" +#include "logger.hpp" #include "registered_memory.hpp" namespace mscclpp { @@ -78,19 +77,17 @@ MSCCLPP_API_CPP Endpoint Context::createEndpoint(EndpointConfig config) { MSCCLPP_API_CPP Connection Context::connect(const Endpoint& localEndpoint, const Endpoint& remoteEndpoint) { if (localEndpoint.device().type == DeviceType::GPU && localEndpoint.device().id < 0) { - throw Error("No GPU device ID provided for local endpoint", ErrorCode::InvalidUsage); + THROW(CONN, Error, ErrorCode::InvalidUsage, "No GPU device ID provided for local endpoint"); } if (remoteEndpoint.device().type == DeviceType::GPU && remoteEndpoint.device().id < 0) { - throw Error("No GPU device ID provided for remote endpoint", ErrorCode::InvalidUsage); + THROW(CONN, Error, ErrorCode::InvalidUsage, "No GPU device ID provided for remote endpoint"); } auto localTransport = localEndpoint.transport(); auto remoteTransport = remoteEndpoint.transport(); if (localTransport != remoteTransport && !(AllIBTransports.has(localTransport) && AllIBTransports.has(remoteTransport))) { - std::stringstream ss; - ss << "Transport mismatch between local (" << localTransport << ") and remote (" << remoteEndpoint.transport() - << ") endpoints"; - throw Error(ss.str(), ErrorCode::InvalidUsage); + THROW(CONN, Error, ErrorCode::InvalidUsage, "Transport mismatch between local (", localTransport, ") and remote (", + remoteTransport, ") endpoints"); } std::shared_ptr conn; if (localTransport == Transport::CudaIpc) { @@ -100,7 +97,9 @@ MSCCLPP_API_CPP Connection Context::connect(const Endpoint& localEndpoint, const } else if (localTransport == Transport::Ethernet) { conn = std::make_shared(shared_from_this(), localEndpoint, remoteEndpoint); } else { - throw Error("Unsupported transport", ErrorCode::InternalError); + THROW(CONN, Error, ErrorCode::InternalError, "Unsupported transport: ", localTransport, + " (this usually means EndpointConfig.transport was left at Transport::Unknown — " + "set it explicitly to CudaIpc, an IB transport, or Ethernet)"); } return Connection(conn); } diff --git a/src/core/gpu_ipc_mem.cc b/src/core/gpu_ipc_mem.cc index c863ecdd..0f58ed20 100644 --- a/src/core/gpu_ipc_mem.cc +++ b/src/core/gpu_ipc_mem.cc @@ -7,6 +7,7 @@ #include #include +#include #include "logger.hpp" #include "unix_socket.hpp" @@ -35,7 +36,7 @@ std::ostream& operator<<(std::ostream& os, const GpuIpcMemHandle::TypeFlags& typ return os; } -[[maybe_unused]] static bool isFabricMemHandleAvailable() { +bool isFabricMemHandleAvailable() { #if (CUDA_NVLS_API_AVAILABLE) static int resultCache = -1; // -1: uninitialized, 0: not available, 1: available if (resultCache != -1) { @@ -283,11 +284,19 @@ GpuIpcMem::GpuIpcMem(const GpuIpcMemHandle& handle) THROW(GPU, Error, ErrorCode::InvalidUsage, "GpuIpcMemHandle type is None, cannot create GpuIpcMem"); } if ((type_ == GpuIpcMemHandle::Type::None) && (handle_.typeFlags & GpuIpcMemHandle::Type::Fabric)) { - if (cuMemImportFromShareableHandle(&allocHandle_, (void*)handle_.fabric.handle, CU_MEM_HANDLE_TYPE_FABRIC) == - CUDA_SUCCESS) { + CUresult res = + cuMemImportFromShareableHandle(&allocHandle_, (void*)handle_.fabric.handle, CU_MEM_HANDLE_TYPE_FABRIC); + if (res == CUDA_SUCCESS) { // Ignore allocHandle in the handle struct since it is process-local and not transferable across processes. handle_.fabric.allocHandle = {}; type_ = GpuIpcMemHandle::Type::Fabric; + } else { + const char* errStr = nullptr; + (void)cuGetErrorString(res, &errStr); + const std::string errMsg = errStr ? std::string(errStr) : std::string("unknown CUDA error"); + WARN(GPU, "Fabric IPC handle import failed (", errMsg, + "); cross-node CudaIpc requires NVIDIA MNNVL hardware and a running IMEX service. ", + "Falling back to other handle types if available."); } } if ((type_ == GpuIpcMemHandle::Type::None) && (handle_.typeFlags & GpuIpcMemHandle::Type::PosixFd)) { @@ -303,7 +312,17 @@ GpuIpcMem::GpuIpcMem(const GpuIpcMemHandle& handle) type_ = GpuIpcMemHandle::Type::RuntimeIpc; } if (type_ == GpuIpcMemHandle::Type::None) { - THROW(GPU, Error, ErrorCode::Aborted, "Failed to open GpuIpcMemHandle (type: ", handle_.typeFlags, ")"); + const bool fabricOnly = (handle_.typeFlags == GpuIpcMemHandle::Type::Fabric); + const std::string hint = fabricOnly + ? std::string( + "The remote rank sent only a Fabric (MNNVL) handle, but this rank could not " + "import it. Check that the IMEX daemon is running on both nodes and that the " + "GPUs share an NVLink fabric.") + : std::string( + "All handle types failed to import; check IMEX service and POSIX FD socket " + "availability."); + THROW(GPU, Error, ErrorCode::Aborted, "Failed to open GpuIpcMemHandle (offered types: ", handle_.typeFlags, "). ", + hint); } } diff --git a/test/mp_unit/port_channel_tests.cu b/test/mp_unit/port_channel_tests.cu index 3b14ed31..4b1b0cfb 100644 --- a/test/mp_unit/port_channel_tests.cu +++ b/test/mp_unit/port_channel_tests.cu @@ -36,6 +36,18 @@ inline void requireGdrForIbMode(IbMode mode, mscclpp::Transport ibTransport) { #define REQUIRE_GDR_FOR_IB_MODE(mode) // No extra requirements on non-CUDA platforms. #endif +// Skip an IPC-only PortChannel test (useIPC=true, useIB=false, useEthernet=false) when CudaIpc +// cannot connect this rank pair. CudaIpc works intra-node always, and cross-node only on MNNVL +// systems (GB200 NVL72 + IMEX). The combined check is "at least 2 ranks per node" OR "fabric +// (MNNVL) handles are usable on this system". +#define REQUIRE_CUDA_IPC_AVAILABLE \ + do { \ + if (gEnv->nRanksPerNode < 2 && !mscclpp::isFabricMemHandleAvailable()) { \ + SKIP_TEST() << "CudaIpc requires intra-node ranks (nRanksPerNode>=2) or MNNVL fabric handles, \ +both unavailable here."; \ + } \ + } while (0) + void PortChannelOneToOneTest::SetUp() { // Use only two ranks setNumRanksToUse(2); @@ -71,7 +83,10 @@ void PortChannelOneToOneTest::setupMeshConnections(std::vectorrank)) && useIPC) { + if (useIPC) { + // CudaIpc works intra-node always, and cross-node on MNNVL systems (GB200 NVL72 + IMEX) + // via fabric handles. Tests that exercise CudaIpc across nodes on non-MNNVL hardware should + // gate themselves with REQUIRE_CUDA_IPC_AVAILABLE; we always request CudaIpc here when asked. cfg.transport = mscclpp::Transport::CudaIpc; } else if (useIb) { cfg.transport = ibTransport; @@ -262,6 +277,7 @@ void PortChannelOneToOneTest::testPingPongPerf(PingPongTestParams params) { } TEST(PortChannelOneToOneTest, PingPong) { + REQUIRE_CUDA_IPC_AVAILABLE; testPingPong(PingPongTestParams{ .useIPC = true, .useIB = false, .useEthernet = false, .waitWithPoll = false, .ibMode = IbMode::Default}); } @@ -279,6 +295,7 @@ TEST(PortChannelOneToOneTest, PingPongEthernet) { } TEST(PortChannelOneToOneTest, PingPongWithPoll) { + REQUIRE_CUDA_IPC_AVAILABLE; testPingPong(PingPongTestParams{ .useIPC = true, .useIB = false, .useEthernet = false, .waitWithPoll = true, .ibMode = IbMode::Default}); } @@ -291,6 +308,7 @@ TEST(PortChannelOneToOneTest, PingPongIbHostModeWithPoll) { } PERF_TEST(PortChannelOneToOneTest, PingPongPerf) { + REQUIRE_CUDA_IPC_AVAILABLE; testPingPongPerf(PingPongTestParams{ .useIPC = true, .useIB = false, .useEthernet = false, .waitWithPoll = false, .ibMode = IbMode::Default}); } @@ -482,7 +500,10 @@ void PortChannelOneToOneTest::testPacketPingPongPerf(bool useIb, IbMode ibMode) proxyService->stopProxy(); } -TEST(PortChannelOneToOneTest, PacketPingPong) { testPacketPingPong(false, IbMode::Default); } +TEST(PortChannelOneToOneTest, PacketPingPong) { + REQUIRE_CUDA_IPC_AVAILABLE; + testPacketPingPong(false, IbMode::Default); +} TEST(PortChannelOneToOneTest, PacketPingPongIbHostMode) { REQUIRE_IBVERBS; @@ -490,7 +511,10 @@ TEST(PortChannelOneToOneTest, PacketPingPongIbHostMode) { testPacketPingPong(true, IbMode::Host); } -PERF_TEST(PortChannelOneToOneTest, PacketPingPongPerf) { testPacketPingPongPerf(false, IbMode::Default); } +PERF_TEST(PortChannelOneToOneTest, PacketPingPongPerf) { + REQUIRE_CUDA_IPC_AVAILABLE; + testPacketPingPongPerf(false, IbMode::Default); +} PERF_TEST(PortChannelOneToOneTest, PacketPingPongPerfIbHostMode) { REQUIRE_IBVERBS; @@ -583,6 +607,7 @@ void PortChannelOneToOneTest::testBandwidth(PingPongTestParams params) { } PERF_TEST(PortChannelOneToOneTest, Bandwidth) { + REQUIRE_CUDA_IPC_AVAILABLE; testBandwidth(PingPongTestParams{ .useIPC = true, .useIB = false, .useEthernet = false, .waitWithPoll = false, .ibMode = IbMode::Default}); }