|
1 | | -# Important notice |
2 | | -***In order to uniformize and simplify the build system we had to make choices. TC is currently only officially supported on Ubuntu 16.04 with gcc 5.4.0, cuda 9.0 and cudnn 7.*** |
3 | | -Other configurations may work too but are not yet officially supported. |
4 | | -For more information about setting up the config that we use to build the conda dependencies see the following [Dockerfile](conda_recipes/Dockerfile). |
5 | | - |
6 | | -Our main goal with this decision is to make the build procedure extremely simple, both reproducible internally and extensible to new targets in the future. |
7 | | -In particular the gcc-4 / gcc-5 ABI switch is not something we want to concern ourselves with at this point, we go for gcc-5.4.0. |
8 | | - |
9 | | -# Conda from scratch (first time configuration) |
10 | | -Choose and set an INSTALLATION_PATH then run the following: |
11 | | - |
12 | | -``` |
13 | | -wget https://repo.anaconda.com/archive/Anaconda3-5.1.0-Linux-x86_64.sh -O anaconda.sh && \ |
14 | | - chmod +x anaconda.sh && \ |
15 | | - ./anaconda.sh -b -p ${INSTALLATION_PATH} && \ |
16 | | - rm anaconda.sh |
17 | | -
|
18 | | -. ${INSTALLATION_PATH}/bin/activate |
19 | | -conda update -y -n base conda |
20 | | -``` |
21 | | - |
22 | | -Create a new environment in which TC will be built and install core dependencies: |
23 | | -``` |
24 | | -conda create -y --name tc_build python=3.6 |
25 | | -conda activate tc_build |
26 | | -conda install -y pyyaml mkl-include pytest |
27 | | -conda install -y -c pytorch pytorch torchvision cuda90 |
28 | | -conda remove -y cudatoolkit --force |
29 | | -conda install -y -c nicolasvasilache llvm-tapir50 halide |
30 | | -``` |
31 | | - |
32 | | -***Note*** As of PyTorch 0.4, PyTorch links cuda libraries dynamically and it |
33 | | -pulls cudatoolkit. However cudatoolkit can never replace a system installation |
34 | | -because it cannot package libcuda.so (which comes with the driver, not the toolkit). |
35 | | -As a consequence cudatoolkit only contains redundant libraries and we remove it |
36 | | -explicitly. In a near future, the unified PyTorch + Caffe2 build system will link |
37 | | -everything statically and stop pulling the cudatoolkit dependency. |
38 | | - |
39 | | -# Optional dependencies |
40 | | -Optionally if you want to use Caffe2 (which is necessary for building the C++ benchmarks |
41 | | -since Caffe2 is our baseline): |
42 | | -``` |
43 | | -conda install -y -c conda-forge eigen |
44 | | -conda install -y -c nicolasvasilache caffe2 |
45 | | -``` |
46 | | - |
47 | | -# Activate preinstalled conda in your current terminal |
48 | | - |
49 | | -Once the first time configuration above has been completed, one should activate conda in |
50 | | -each new terminal window explicitly (it is discouraged to add this to your `.bashrc` or |
51 | | -equivalent) |
52 | | -``` |
53 | | -. ${CONDA_PATH}/bin/activate |
54 | | -conda activate tc_build |
55 | | -``` |
56 | | - |
57 | | -# Cudnn version |
58 | | -***Note*** As of PyTorch 0.4, we need to package our own Caffe2. The curent PyTorch + Caffe2 |
59 | | -build system links cudnn dynamically. The version of cudnn that is linked dynamically |
60 | | -is imposed on us by the docker image supported by NVIDIA |
61 | | -[Dockerfile](conda_recipes/docker-images/tc-cuda9.0-cudnn7.1-ubuntu16.04-devel/Dockerfile). |
62 | | -For now this cudnn version is cudnn 7.1. |
63 | | -If for some reason, one cannot install cudnn 7.1 system-wide, one may resort to the |
64 | | -following: |
65 | | -``` |
66 | | -conda install -c anaconda cudnn |
67 | | -conda remove -y cudatoolkit --force |
68 | | -``` |
69 | | - |
70 | | -***Note*** cudnn pulls a cudatoolkit dependencey but this can never replace a system |
71 | | -installation because it cannot package libcuda.so (which comes with the driver, |
72 | | -not the toolkit). |
73 | | -As a consequence cudatoolkit only contains redundant libraries and we remove it |
74 | | -explicitly. In a near future, the unified PyTorch + Caffe2 build system will link |
75 | | -everything statically and we will not need to worry about cudnn anymore. |
76 | | - |
77 | | -# Build TC with dependencies supplied by conda (including cudnn 7.1) |
78 | | -``` |
79 | | -CLANG_PREFIX=$(${CONDA_PREFIX}/bin/llvm-config --prefix) ./build.sh |
80 | | -``` |
81 | | -You may need to pass the environment variable `CUDA_TOOLKIT_ROOT_DIR` pointing |
82 | | -to your cuda installation (this is required for `FindCUDA.cmake` to find your cuda installation |
83 | | -and can be omitted on most systems). |
84 | | - |
85 | | -# Test locally |
86 | | -Run C++ tests: |
87 | | -``` |
88 | | -./test.sh |
89 | | -``` |
90 | | - |
91 | | -Install the TC Python package locally to `/tmp`: |
92 | | -``` |
93 | | -python setup.py install --prefix=/tmp |
94 | | -export PYTHONPATH=${PYTHONPATH}:$(find /tmp/lib -name site-packages) |
95 | | -``` |
96 | | - |
97 | | -Run Python smoke checks: |
98 | | -``` |
99 | | -python -c 'import torch' |
100 | | -python -c 'import tensor_comprehensions' |
101 | | -``` |
102 | | - |
103 | | -Run Python tests: |
104 | | -``` |
105 | | -./test_python/run_test.sh |
106 | | -``` |
| 1 | +see the [instructions](docs/source/installation.rst). |
0 commit comments