[Question] Install from Source - Pytorch


#1

Hi There

I am trying to build tvm from source with pytorch, cuda 10, cudnn 7.5.1 and when I run the command cmake .. , I see the following warning (listed below) . However, I dont see any build fail with make -j4 . Will it lead to any performance degradation ? Also, I was following this pull request and was wondering if Pytorch can run directly with TVM without ONNX support ?

-- Found CUDA_CUDA_LIBRARY=/usr/local/cuda/targets/x86_64-linux/lib/stubs/libcuda.so                                                                                                                                                               [337/990]
-- Found CUDA_CUDART_LIBRARY=/usr/local/cuda/lib64/libcudart.so
-- Found CUDA_NVRTC_LIBRARY=/usr/local/cuda/lib64/libnvrtc.so
-- Found CUDA_CUDNN_LIBRARY=/usr/lib/x86_64-linux-gnu/libcudnn.so
-- Found CUDA_CUBLAS_LIBRARY=/opt/conda/lib/libcublas.so
-- Build with CUDA support
-- Build with cuDNN support
-- Build with cuBLAS support
-- Use llvm-config=/clang+llvm-8.0.0-x86_64-linux-gnu-ubuntu-16.04/bin/llvm-config
-- /clang+llvm-8.0.0-x86_64-linux-gnu-ubuntu-16.04/include
-- Found LLVM_INCLUDE_DIRS=/clang+llvm-8.0.0-x86_64-linux-gnu-ubuntu-16.04/include
-- Found LLVM_DEFINITIONS= -D_GNU_SOURCE -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS
-- Found TVM_LLVM_VERSION=80
-- Build with LLVM 
-- Set TVM_LLVM_VERSION=80
-- Use BLAS library /usr/lib/libopenblas.so
-- Build with contrib.sort
-- Build with contrib.hybriddump
-- Configuring done
CMake Warning at CMakeLists.txt:197 (add_library):
  Cannot generate a safe runtime search path for target tvm because there is
  a cycle in the constraint graph:

    dir 0 is [/usr/local/cuda/lib64]
      dir 2 must precede it due to runtime library [libcublas.so.10.0]
    dir 1 is [/usr/local/cuda/targets/x86_64-linux/lib/stubs]
    dir 2 is [/opt/conda/lib]
      dir 0 must precede it due to runtime library [libnvrtc.so.10.0]

  Some of these libraries may not be found correctly.


CMake Warning at CMakeLists.txt:198 (add_library):
  Cannot generate a safe runtime search path for target tvm_topi because
  there is a cycle in the constraint graph:

    dir 0 is [/tvm/build]
    dir 1 is [/usr/local/cuda/lib64]
      dir 3 must precede it due to runtime library [libcublas.so.10.0]
    dir 2 is [/usr/local/cuda/targets/x86_64-linux/lib/stubs]
    dir 3 is [/opt/conda/lib]
      dir 1 must precede it due to runtime library [libnvrtc.so.10.0]

  Some of these libraries may not be found correctly.


CMake Warning at CMakeLists.txt:205 (add_library):
  Cannot generate a safe runtime search path for target nnvm_compiler because
  there is a cycle in the constraint graph:

    dir 0 is [/tvm/build]
    dir 1 is [/usr/local/cuda/lib64]
      dir 3 must precede it due to runtime library [libcublas.so.10.0]
    dir 2 is [/usr/local/cuda/targets/x86_64-linux/lib/stubs]
    dir 3 is [/opt/conda/lib]
      dir 1 must precede it due to runtime library [libnvrtc.so.10.0]

  Some of these libraries may not be found correctly.


CMake Warning at CMakeLists.txt:199 (add_library):
  Cannot generate a safe runtime search path for target tvm_runtime because
  there is a cycle in the constraint graph:

    dir 0 is [/usr/local/cuda/lib64]
      dir 2 must precede it due to runtime library [libcublas.so.10.0]
    dir 1 is [/usr/local/cuda/targets/x86_64-linux/lib/stubs]
    dir 2 is [/opt/conda/lib]
      dir 0 must precede it due to runtime library [libcudart.so.10.0]

  Some of these libraries may not be found correctly.


-- Generating done
-- Build files have been written to: /tvm/build