Skip to content

Latest commit

 

History

History
315 lines (219 loc) · 16.3 KB

README.md

File metadata and controls

315 lines (219 loc) · 16.3 KB
Raw

TensorRT-LLM for Windows

Table of Contents

Overview

NOTE: The Windows release of TensorRT-LLM is currently in beta. We recommend using the rel branch for the most stable experience. The latest supported Windows release is 0.6.1. You are currently on main.

TensorRT-LLM is supported on bare-metal Windows for single-GPU inference. The release supports GeForce 40-series GPUs.

The release wheel for Windows can be installed with pip. Alternatively, you may build TensorRT-LLM for Windows from source. Building from source is an advanced option and is not necessary for building or running LLM engines. It is, however, required if you plan to use the C++ runtime directly or run C++ benchmarks.

Quick Start

You can clone this repository using Git for Windows.

We provide a Powershell script, setup_env.ps1, which installs Python, CUDA 12.2, and Microsoft MPI automatically with default settings. Be sure to run Powershell as Administrator to use the script. Usage:

./setup_env.ps1 [-skipCUDA] [-skipPython] [-skipMPI]

Close and reopen Powershell after running the script so that Path changes take effect. The script will install whichever components are not skipped. Any components may be installed manually instead of using the script. Further, cuDNN must be installed manually. For more details about manually installing prerequisites, check the Detailed Setup instructions below.

Prerequisites:

Once your prerequisites are installed, install TensorRT-LLM:

pip install tensorrt_llm --extra-index-url https://pypi.nvidia.com --extra-index-url https://download.pytorch.org/whl/cu121

You may now build and run models!

Detailed Setup

Python

Install Python 3.10. Select "Add python.exe to PATH" at the start of the installation. The installation may only add the python command, but not the python3 command. Navigate to the installation path, %USERPROFILE%\AppData\Local\Programs\Python\Python310 (note AppData is a hidden folder), and copy python.exe to python3.exe.

CUDA

Install the CUDA 12.2 Toolkit. You may use the Express Installation option. Installation may require a restart.

Microsoft MPI

Download and install Microsoft MPI. You will be prompted to choose between an exe, which installs the MPI executable, and an msi, which installs the MPI SDK. Download and install both.

TensorRT-LLM Repo

It may be useful to create a single folder for holding TensorRT-LLM and its dependencies, such as %USERPROFILE%\inference\. We will assume this folder structure in further steps.

Clone TensorRT-LLM:

git clone --branch rel https://github.com/NVIDIA/TensorRT-LLM.git
cd TensorRT-LLM
git submodule update --init --recursive

cuDNN

Download and unzip cuDNN. Move the folder to a location you can reference later, such as %USERPROFILE%\inference\cuDNN.

You'll need to add libraries and binaries for cuDNN to your system's Path environment variable. To do so, click the Windows button and search for "environment variables." Select "Edit the system environment variables." A "System Properties" window will open. Select the "Environment Variables" button at the bottom right, then in the new window under "System variables" click "Path" then the "Edit" button. Add "New" lines for the bin and lib dirs of cuDNN. Your Path should include lines like this:

%USERPROFILE%\inference\cuDNN\bin
%USERPROFILE%\inference\cuDNN\lib

Click "OK" on all the open dialogue windows. Be sure to close and re-open any existing Powershell or Git Bash windows so they pick up the new Path.

If you are using the pre-built TensorRT-LLM release wheel (recommended unless you need to directly invoke the C++ runtime), skip to Installation. If you are building your own wheel from source, proceed to Building from Source.

Building from Source

Advanced. Skip this section if you plan to use the pre-built TensorRT-LLM release wheel.

Building from source requires extra prerequisites:

We provide a Docker container with these prerequisites already installed. Building with Docker will require you to install Docker Desktop on Windows, build the container, build TensorRT-LLM, and copy files out of the Docker container for usage on your Windows host machine. Alternatively, you may install the prerequisites on a bare-metal machine and build there. See Docker Build Instructions or Bare-Metal Build Instructions to proceed.

Docker Build Instructions

Docker Desktop

Install Docker Desktop on Windows. You may need to change the following configurations:

  • Right click the Docker icon in the Windows system tray (bottom right of your taskbar) and select "Switch to Windows containers..."
  • In Docker Desktop settings on the General tab, uncheck "Use the WSL 2 based image"
  • On the Docker Engine tab, set you configuration file to
{
  "experimental": true
}

Note: After building, you'll need to copy files out of your container. docker cp is not supported on Windows for Hyper-V based images. Unless you are using WSL 2 based images, be sure to mount a folder, e.g. trt-llm-build, to your container when you run it for moving files between the container and host system.

Acquiring an Image

The Docker container will be hosted for public download in a future release. At this time, it must be built manually. See windows/docker/README.md for image build instructions.

Running the Container

Run the container in interactive mode with your build folder mounted. Be sure to specify a memory limit with the -m flag - by default the limit is 2GB, which is not sufficient to build TensorRT-LLM.

docker run -it -m 12g -v .\trt-llm-build:C:\workspace\trt-llm-build tensorrt-llm-windows-build:latest

Build and Extract Files

Clone and setup the TensorRT-LLM repository within the container:

git clone https://github.com/NVIDIA/TensorRT-LLM.git
cd TensorRT-LLM
git submodule update --init --recursive

Build TensorRT-LLM

python .\scripts\build_wheel.py -a "89-real" --trt_root C:\workspace\TensorRT-9.3.0.1\

The above command will generate build\tensorrt_llm-*.whl. Copy or move this into your mounted folder so it can be accessed on your host machine. If you intend to use the C++ runtime, you'll also need to gather various DLLs from the build into your mounted folder. Complete information about these files can be found below in Extra Steps for C++ Runtime Usage.

Once you've gathered your files into the mounted folder, you may exit the container and continue on to Installation.

Bare-Metal Build Instructions

We provide a second Powershell script, setup_build_env.ps1, which installs CMake, Microsoft Visual Studio Build Tools, and TensorRT automatically with default settings. Be sure to run Powershell as Administrator to use the script. Usage:

./setup_build_env.ps1 -TRTPath <TRT-containing-folder> [-skipCMake] [-skipVSBuildTools] [-skipTRT]

Close and reopen Powershell after running the script so that Path changes take effect. Note that you should supply to -TRTPath a directory that already exists to contain TensorRT - e.g. -TRTPath ~/inference may be valid, but -TRTPath ~/inference/TensorRT will not be valid if TensorRT does not exist. -TRTPath isn't required if -skipTRT is supplied.

The script will install whichever components are not skipped. Any components may be installed manually instead of using the script. Note that for Visual Studio, the script just installs the command-line Build Tools. You may prefer a full Visual Studio 2022 IDE installation, which is linked below.

Nsight NVTX must be installed manually. For more details about manually installing individual prerequisites, including NVTX, check the instructions below.

CMake

Install CMake (version 3.27.7 recommended) and select the option to add it to the system path.

Visual Studio

Download and install Visual Studio 2022. When prompted to select more Workloads, check "Desktop development with C++."

TensorRT

Download and unzip TensorRT 9.3.0.1 for TensorRT-LLM. Move the folder to a location you can reference later, such as %USERPROFILE%\inference\TensorRT.

You'll need to add libraries for TensorRT to your system's Path environment variable. Follow the same instructions used for cuDNN. Your Path should include a line like this:

%USERPROFILE%\inference\TensorRT\lib

Be sure to close and re-open any existing Powershell or Git Bash windows so they pick up the new Path.

Now, to install the TensorRT core libraries, run Powershell and use pip to install the Python wheel:

pip install %USERPROFILE%\inference\TensorRT\python\tensorrt-*.whl

You may run the following command to verify that your TensorRT installation is working properly:

python -c "import tensorrt as trt; print(trt.__version__)"

Nsight NVTX

TensorRT-LLM on Windows currently depends on NVTX assets that do not come packaged with the CUDA12.2 installer. To install these assets, download the CUDA11.8 Toolkit. During installation, select "Advanced installation." Nsight NVTX is located in the CUDA drop down. Deselect all packages, and then select Nsight NVTX.

64-bit Developer Powershell

In order to build, you'll need to launch a 64-bit Developer Powershell. From your usual Powershell terminal, run one of the following two commands.

If you installed Visual Studio Build Tools (e.g. using the setup_build_env.ps1 script):

& 'C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\Common7\Tools\Launch-VsDevShell.ps1' -Arch amd64

If you installed Visual Studio Community (e.g. via manual GUI setup):

& 'C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\Tools\Launch-VsDevShell.ps1' -Arch amd64

Build

In Powershell, from the TensorRT-LLM root folder, run:

python .\scripts\build_wheel.py -a "89-real" --trt_root <path_to_trt_root>

The -a flag specifies the device architecture. "89-real" supports GeForce 40-series cards.

Note that the flag -D "ENABLE_MULTI_DEVICE=0", while not specified here, is implied on Windows. Multi-device inference is supported on Linux, but not on Windows.

The above command will generate build\tensorrt_llm-*.whl.

Installation

To download and install the wheel, in Powershell, run:

pip install tensorrt_llm --extra-index-url https://pypi.nvidia.com --extra-index-url https://download.pytorch.org/whl/cu121

Alternatively, if you built the wheel from source, locate your wheel (either in build\ or in the folder you mounted to your Docker contailer) and run:

pip install tensorrt_llm-*.whl

You may run the following command to verify that your TensorRT-LLM installation is working properly:

python -c "import tensorrt_llm; print(tensorrt_llm._utils.trt_version())"

Extra Steps for C++ Runtime Usage

Advanced. Skip this section if you do not intend to use the TensorRT-LLM C++ runtime directly. Note that you have to have built from source to use the C++ runtime.

Building from source creates libraries that can be used if you wish to directly link against the C++ runtime for TensorRT-LLM. These libraries are also required if you wish to run C++ unit tests and some benchmarks.

Building from source will produce the following library files:

  • tensorrt_llm libraries located in cpp\build\tensorrt_llm\Release
    • tensorrt_llm.dll - Shared library
    • tensorrt_llm.exp - Export file
    • tensorrt_llm.lib - Stub for linking to tensorrt_llm.dll
    • tensorrt_llm_static.lib - Static library
  • Dependency libraries (These get copied to tensorrt_llm\libs\)
    • nvinfer_plugin_tensorrt_llm libraries located in cpp\build\tensorrt_llm\plugins\
      • nvinfer_plugin_tensorrt_llm.dll
      • nvinfer_plugin_tensorrt_llm.exp
      • nvinfer_plugin_tensorrt_llm.lib
    • th_common libraries located in cpp\build\tensorrt_llm\thop\
      • th_common.dll
      • th_common.exp
      • th_common.lib

The locations of the DLLs, in addition to some torch DLLs, must be added to the Windows Path in order to us the TensorRT-LLM C++ runtime. As in Detailed Setup, append the locations of these libraries to your Path. When complete, your Path should include lines similar to these:

%USERPROFILE%\inference\TensorRT-LLM\cpp\build\tensorrt_llm\Release
%USERPROFILE%\AppData\Local\Programs\Python\Python310\Lib\site-packages\tensorrt_llm\libs
%USERPROFILE%\AppData\Local\Programs\Python\Python310\Lib\site-packages\torch\lib

Your Path additions may differ, particularly if you used the Docker method and copied all the relevant DLLs into a single folder.

For examples of how to use the C++ runtime, see the unit tests in gptSessionTest.cpp and the related CMakeLists.txt file.

Next Steps

See examples/llama for a showcase of how to run a quick benchmark on LLaMa.

Limitations

openai-triton examples are not supported on Windows.

Troubleshooting Common Errors

Many build errors can be resolved by simply deleting the build tree. Try running the build script with --clean or running rm -r cpp/build.

cuDNN Linking Errors

If you encounter errors such as "Entry Point Not Found" (see for example #1062) the issue might be a mismatch in the cuDNN libraries shipped from torch and tensorrt. To rectify this, please try the following steps

python -m pip uninstall -y tensorrt_llm
python -m pip install --upgrade pip
python -m pip install nvidia-cudnn-cu11==8.9.4.25 --no-cache-dir
python -m pip install --pre --extra-index-url https://pypi.nvidia.com/ tensorrt==9.2.0.post12.dev5 --no-cache-dir
python -m pip uninstall -y nvidia-cudnn-cu11
python -m pip install tensorrt_llm  --extra-index-url https://pypi.nvidia.com/ --extra-index-url https://pypi.nvidia.com/ --extra-index-url https://download.pytorch.org/whl/cu121