mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
16d2467ea8
* Update TensorRT-LLM --------- Co-authored-by: Denis Kayshev <topenkoff@gmail.com> Co-authored-by: akhoroshev <arthoroshev@gmail.com> Co-authored-by: Patrick Reiter Horn <patrick.horn@gmail.com> Update
71 lines
2.5 KiB
C++
71 lines
2.5 KiB
C++
/*
|
|
* Copyright (c) 2022-2024, 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 <memory>
|
|
|
|
/// @brief Forward declaration of cudaMemPool_t to avoid including "driver_types.h"
|
|
struct CUmemPoolHandle_st;
|
|
using cudaMemPool_t = CUmemPoolHandle_st*;
|
|
|
|
namespace tensorrt_llm::runtime
|
|
{
|
|
|
|
class CudaMemPool
|
|
{
|
|
public:
|
|
explicit CudaMemPool(cudaMemPool_t pool);
|
|
|
|
/// @brief Gets the amount of reserved memory in the memory pool stream, WITHOUT synchronizing.
|
|
[[nodiscard]] std::size_t memoryPoolReserved() const;
|
|
|
|
/// @brief Gets the amount of used memory in the memory pool, WITHOUT synchronizing.
|
|
[[nodiscard]] std::size_t memoryPoolUsed() const;
|
|
|
|
/// @brief Gets the amount of free memory in the memory pool, WITHOUT synchronizing.
|
|
[[nodiscard]] std::size_t memoryPoolFree() const;
|
|
|
|
/// @brief Hints the driver to trim the pool. Does not guarantee that the amount of reserved memory will actually
|
|
/// decrease, only guarantees that this amount after trimming will be larger than the provided size.
|
|
void memoryPoolTrimTo(std::size_t size);
|
|
|
|
/// @brief Returns the underlying cudaMemPool_t for usage by CUDA APIs.
|
|
[[nodiscard]] cudaMemPool_t getPool() const;
|
|
|
|
/// @brief Gets or initializes and gets the primary memory pool for the provided device ID if it was successfully
|
|
/// initialized, nullptr otherwise.
|
|
static std::shared_ptr<tensorrt_llm::runtime::CudaMemPool> getPrimaryPoolForDevice(int deviceId);
|
|
|
|
/// @brief Returns a value indicating whether memory pools are supported on the device.
|
|
/// @details Memory pools depend on the presence of the UVM driver. On some systems, the UVM driver is explicitly
|
|
/// disabled.
|
|
static bool supportsMemoryPool(int deviceId);
|
|
|
|
private:
|
|
class Deleter
|
|
{
|
|
public:
|
|
void operator()(cudaMemPool_t pool) const;
|
|
};
|
|
|
|
using PoolPtr = std::unique_ptr<std::remove_pointer_t<cudaMemPool_t>, Deleter>;
|
|
|
|
PoolPtr mPool;
|
|
};
|
|
|
|
} // namespace tensorrt_llm::runtime
|