TensorRT-LLMs/cpp/tensorrt_llm/runtime/mcastDeviceMemory.h
2025-09-17 07:43:20 +08:00

123 lines
3.7 KiB
C++

/*
* Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#pragma once
#include "tensorrt_llm/common/mcastDevMemUtils.h"
#include "tensorrt_llm/runtime/ipcNvlsMemory.h"
#include "tensorrt_llm/runtime/utils/mpiUtils.h"
#include <cstddef>
#include <cstdint>
#include <cuda.h>
#include <memory>
#include <vector>
namespace tensorrt_llm::runtime
{
//! \brief A class that manages multicast device memory for efficient communication between GPUs.
//!
//! This class uses IPC-based allocation if mnNvlink is true, otherwise it uses fabric allocation.
//! The fabric allocation can also be used for single-node/intra-node-only communication, but the machine
//! must properly configure IMEX services. See:
//! https://docs.nvidia.com/multi-node-nvlink-systems/imex-guide/gettingstarted.html
//!
//! The class manages both unicast pointers (one per rank) and a single multicast pointer,
//! along with signal pads used for synchronization between devices.
class McastDeviceMemory
{
public:
// Disallow copy construction
McastDeviceMemory(McastDeviceMemory const&) = delete;
McastDeviceMemory& operator=(McastDeviceMemory const&) = delete;
McastDeviceMemory(
size_t bufSize, uint32_t groupSize, uint32_t groupRank, uint32_t splitColor, int deviceIdx, bool mnNvlink);
// We don't register the pointer in these two functions since we don't expect any python-level code would call
// to obtain the raw pointers.
//! Get the raw array of signal pad pointers to all ranks (including self)
void** getSignalPadPtrsDev()
{
return mSignalPadsDev;
}
//! Get the raw array of unicast pointers to all ranks (including self)
void** getBufferPtrsDev()
{
return mUcPtrsDev;
}
//! Get the raw unicast pointer to a given rank
void* getUnicastPtr(uint32_t rank)
{
auto* data_ptr = reinterpret_cast<void*>(mUcPtrs[rank]);
tensorrt_llm::common::registerMcastDevMemBuffer(data_ptr, this);
return data_ptr;
}
//! Get the raw multicast pointer
void* getMulticastPtr()
{
auto* data_ptr = reinterpret_cast<void*>(mMcPtr);
tensorrt_llm::common::registerMcastDevMemBuffer(data_ptr, this);
return data_ptr;
}
[[nodiscard]] size_t getRank() const
{
return mGroupRank;
}
[[nodiscard]] size_t getWorldSize() const
{
return mGroupSize;
}
~McastDeviceMemory();
private:
bool mIsMNNvlink;
int mDeviceIdx;
uint32_t mGroupSize, mGroupRank;
size_t mBufSize;
size_t mSignalPadOffset;
size_t mAllocationSize;
CUdeviceptr mMcPtr;
CUmemGenericAllocationHandle mMcHandle;
std::vector<CUmemGenericAllocationHandle> mUcHandles;
tensorrt_llm::mpi::MpiComm mGroupComm; //!< The MPI communicator for the group
// Host array of pointers
std::vector<CUdeviceptr> mUcPtrs;
std::vector<CUdeviceptr> mSignalPads;
// Device array of pointers
void** mUcPtrsDev;
void** mSignalPadsDev;
// For intra-node mcast
tensorrt_llm::runtime::IpcNvlsHandle* mNvlsHandle;
void allocMnMcastMem(size_t bufSize);
void allocNvlsMcastMem(size_t bufSize);
};
constexpr size_t kSIGNAL_PAD_SIZE = 2048;
} // namespace tensorrt_llm::runtime