mirror of
https://github.com/microsoft/mscclpp.git
synced 2026-05-12 01:10:22 +00:00
34945fb107
* Renamed and moved mem alloc functions into the `mscclpp::detail::` namespace (now `mscclpp::detail::gpuCalloc*<T>()`) * Deprecated constructor-calling mem alloc functions (`mscclpp::makeShared*<T>()` and `mscclpp::makeUnique*<T>()`) * Added a new `mscclpp::GpuBuffer<T>()` class that should be used in general for allocating communication buffers * Added a new `mscclpp.utils.GpuBuffer` Python class that inherits `cupy.ndarray` and allocates using `mscclpp::gpuMemAlloc` * Renamed `mscclpp::memcpyCuda*<T>()` functions into `mscclpp::gpuMemcpy*<T>()` for name consistency * A few fixes in NVLS memory allocation * Tackled minor compiler warnings
66 lines
2.2 KiB
C++
66 lines
2.2 KiB
C++
// Copyright (c) Microsoft Corporation.
|
|
// Licensed under the MIT license.
|
|
|
|
#include <mpi.h>
|
|
|
|
#include <filesystem>
|
|
#include <mscclpp/npkit/npkit.hpp>
|
|
|
|
#include "mp_unit_tests.hpp"
|
|
|
|
namespace {
|
|
std::string getExecutablePath() {
|
|
char result[PATH_MAX];
|
|
ssize_t count = readlink("/proc/self/exe", result, PATH_MAX);
|
|
if (count == -1) {
|
|
throw std::runtime_error("Failed to get executable path");
|
|
}
|
|
return std::string(result, count);
|
|
}
|
|
} // namespace
|
|
|
|
void ExecutorTest::SetUp() {
|
|
MultiProcessTest::SetUp();
|
|
|
|
MSCCLPP_CUDATHROW(cudaSetDevice(rankToLocalRank(gEnv->rank)));
|
|
std::shared_ptr<mscclpp::TcpBootstrap> bootstrap;
|
|
mscclpp::UniqueId id;
|
|
bootstrap = std::make_shared<mscclpp::TcpBootstrap>(gEnv->rank, gEnv->worldSize);
|
|
if (gEnv->rank == 0) id = bootstrap->createUniqueId();
|
|
MPI_Bcast(&id, sizeof(id), MPI_BYTE, 0, MPI_COMM_WORLD);
|
|
bootstrap->initialize(id);
|
|
std::shared_ptr<mscclpp::Communicator> communicator = std::make_shared<mscclpp::Communicator>(bootstrap);
|
|
executor = std::make_shared<mscclpp::Executor>(communicator);
|
|
npkitDumpDir = getenv("NPKIT_DUMP_DIR");
|
|
if (npkitDumpDir != nullptr) {
|
|
NpKit::Init(gEnv->rank);
|
|
}
|
|
}
|
|
|
|
void ExecutorTest::TearDown() {
|
|
if (npkitDumpDir != nullptr) {
|
|
NpKit::Dump(npkitDumpDir);
|
|
NpKit::Shutdown();
|
|
}
|
|
executor.reset();
|
|
MultiProcessTest::TearDown();
|
|
}
|
|
|
|
TEST_F(ExecutorTest, TwoNodesAllreduce) {
|
|
if (gEnv->worldSize != 2 || gEnv->nRanksPerNode != 2) {
|
|
GTEST_SKIP() << "This test requires world size to be 2 and ranks per node to be 2";
|
|
return;
|
|
}
|
|
std::string executablePath = getExecutablePath();
|
|
std::filesystem::path path = executablePath;
|
|
std::filesystem::path executionFilesPath =
|
|
path.parent_path().parent_path().parent_path() / "test/execution-files/allreduce.json";
|
|
mscclpp::ExecutionPlan plan(executionFilesPath.string());
|
|
const int bufferSize = 1024 * 1024;
|
|
std::shared_ptr<char> sendbuff = mscclpp::GpuBuffer(bufferSize).memory();
|
|
mscclpp::CudaStreamWithFlags stream(cudaStreamNonBlocking);
|
|
executor->execute(gEnv->rank, sendbuff.get(), sendbuff.get(), bufferSize, bufferSize, mscclpp::DataType::FLOAT16,
|
|
plan, stream);
|
|
MSCCLPP_CUDATHROW(cudaStreamSynchronize(stream));
|
|
}
|