Pytorch build from source help!

C++
1

Hi, I want build pytorch from source, first of all, my envrioment:

  • cuda10.1
  • g++5/6
  • Ubuntu16.04

I know already how to build pytorch from source, but now I got a problem and can no figure out why:

the purpose of build pytorch from source is simple: I want CUDA10.1 support and official does not have.

Currently, I can not build it since I am unable to disable mkldnn and gloo, seems my cmake config has no effect:

-- Found a library with LAPACK API (open).
disabling ROCM because NOT USE_ROCM is set
-- MIOpen not found. Compiling without MIOpen support
-- MKLDNN_THREADING = OMP:COMP
CMake Warning (dev) at third_party/ideep/mkl-dnn/cmake/options.cmake:33 (option):
  Policy CMP0077 is not set: option() honors normal variables.  Run "cmake
  --help-policy CMP0077" for policy details.  Use the cmake_policy command to
  set the policy and suppress this warning.

  For compatibility with older versions of CMake, option is clearing the
  normal variable 'MKLDNN_ENABLE_CONCURRENT_EXEC'.
Call Stack (most recent call first):
  third_party/ideep/mkl-dnn/cmake/utils.cmake:24 (include)
  third_party/ideep/mkl-dnn/CMakeLists.txt:74 (include)
This warning is for project developers.  Use -Wno-dev to suppress it.

-- This is a product build
-- Found OpenMP_C: -fopenmp  
-- Found OpenMP_CXX: -fopenmp  
-- Found OpenMP: TRUE   
-- OpenMP lib: provided by compiler
-- Could NOT find Doxygen (missing: DOXYGEN_EXECUTABLE) 
-- VTune profiling environment is unset
-- Found MKL-DNN: TRUE
-- Looking for clock_gettime in rt
-- Looking for clock_gettime in rt - found
-- Looking for mmap
-- Looking for mmap - found
-- Looking for shm_open
-- Looking for shm_open - found
-- Looking for shm_unlink
-- Looking for shm_unlink - found
-- Looking for malloc_usable_size
-- Looking for malloc_usable_size - found
-- Performing Test C_HAS_THREAD
-- Performing Test C_HAS_THREAD - Success
-- GCC 6.5.0: Adding gcc and gcc_s libs to link line
-- NUMA paths:
-- /usr/include
-- /usr/lib/x86_64-linux-gnu/libnuma.so
-- Performing Test COMPILER_SUPPORTS_NO_AVX256_SPLIT
-- Performing Test COMPILER_SUPPORTS_NO_AVX256_SPLIT - Success
-- Check size of long double
-- Check size of long double - done
-- Performing Test COMPILER_SUPPORTS_LONG_DOUBLE
-- Performing Test COMPILER_SUPPORTS_LONG_DOUBLE - Success
-- Performing Test COMPILER_SUPPORTS_FLOAT128
-- Performing Test COMPILER_SUPPORTS_FLOAT128 - Success
-- Performing Test COMPILER_SUPPORTS_SSE2
-- Performing Test COMPILER_SUPPORTS_SSE2 - Success
-- Performing Test COMPILER_SUPPORTS_SSE4
-- Performing Test COMPILER_SUPPORTS_SSE4 - Success
-- Performing Test COMPILER_SUPPORTS_AVX
-- Performing Test COMPILER_SUPPORTS_AVX - Success
-- Performing Test COMPILER_SUPPORTS_FMA4
-- Performing Test COMPILER_SUPPORTS_FMA4 - Success
-- Performing Test COMPILER_SUPPORTS_AVX2
-- Performing Test COMPILER_SUPPORTS_AVX2 - Success
-- Performing Test COMPILER_SUPPORTS_SVE
-- Performing Test COMPILER_SUPPORTS_SVE - Failed
-- Performing Test COMPILER_SUPPORTS_AVX512F
-- Performing Test COMPILER_SUPPORTS_AVX512F - Failed
-- Performing Test COMPILER_SUPPORTS_OPENMP
-- Performing Test COMPILER_SUPPORTS_OPENMP - Success
-- Performing Test COMPILER_SUPPORTS_WEAK_ALIASES
-- Performing Test COMPILER_SUPPORTS_WEAK_ALIASES - Success
-- Performing Test COMPILER_SUPPORTS_BUILTIN_MATH
-- Performing Test COMPILER_SUPPORTS_BUILTIN_MATH - Success
-- Configuring build for SLEEF-v3.2
   Target system: Linux-4.15.0-58-generic
   Target processor: x86_64
   Host system: Linux-4.15.0-58-generic
   Host processor: x86_64
   Detected C compiler: GNU @ /usr/bin/cc
-- Using option `-Wall -Wno-unused -Wno-attributes -Wno-unused-result -Wno-psabi -ffp-contract=off -fno-math-errno -fno-trapping-math` to compile libsleef
-- Building shared libs : OFF
-- MPFR : LIB_MPFR-NOTFOUND
-- GMP : /usr/lib/x86_64-linux-gnu/libgmp.so
-- RUNNING_ON_TRAVIS : 0
-- COMPILER_SUPPORTS_OPENMP : 1
-- Include NCCL operators
-- Including IDEEP operators
-- Excluding image processing operators due to no opencv
-- Excluding video processing operators due to no opencv
-- Include Observer library

command I runned is:

USE_MKLDNN=0 USE_DISTRIBUTED=0 sudo python3 setup.py install --cmake

what am wrong?

Anybody know why pls help!!

2

Hi,

I think the problem is that you use sudo. When you do this, it switch the environment to the root one, and so the env variable you define inline are not used.
In general, you don’t want to use sudo when you install pytorch. If you use linux, you can setup your rights to be in the staff group to be able to do global install without sudo. Or you can do a per-user install by adding a --user to the install command.

1 Like