pytorch_backend build fails with `cannot find -ltorch`

Description I’m building the Triton server with ./build.py --no-container-build --target-platform ubuntu --cmake-dir=$(pwd)/build --build-dir=pwd/citritonbuild --enable-logging --enable-stats --enable-tracing --enable-metrics --enable-gpu-metrics --enable-gpu --filesystem=gcs --filesystem=azure_storage --filesystem=s3 --endpoint=http --endpoint=grpc --repo-tag=common:r21.05 --repo-tag=core:r21.05 --repo-tag=backend:r21.05 --repo-tag=thirdparty:r21.05 --backend=ensemble --backend=tensorrt --backend=python --backend=tensorflow1 --backend=tensorflow2 --backend=pytorch --repoagent=checksum --image tensorflow1,nvcr.io/nvidia/tensorflow:21.05-tf1-py3 --image tensorflow2,nvcr.io/nvidia/tensorflow:21.05-tf2-py3 --image pytorch,nvcr.io/nvidia/pytorch:21.05-py3

it fails with

[  9%] Built target kernel-library-new
[ 13%] Built target ptlib_target
[ 27%] Built target triton-common-async-work-queue
[ 36%] Built target triton-core-serverstub
[ 68%] Built target triton-backend-utils
[ 72%] Linking CXX shared library libtriton_pytorch.so
/usr/bin/ld: cannot find -ltorch
/usr/bin/ld: cannot find -ltorchvision
collect2: error: ld returned 1 exit status
make[2]: *** [CMakeFiles/triton-pytorch-backend.dir/build.make:121: libtriton_pytorch.so] Error 1
make[1]: *** [CMakeFiles/Makefile2:173: CMakeFiles/triton-pytorch-backend.dir/all] Error 2
make: *** [Makefile:136: all] Error 2

I notice the libtorch.so and libtorchvision.so present in the build folder.

Triton Information What version of Triton are you using? r21.05

Are you using the Triton container or did you build it yourself? I’m building it myself

[GowthamKudupudi]

I’m referring to the build folder citritonbuild/pytorch/build. I see the Extracting pytorch and torchvision libraries and includes from in the build log at [ 13%]. Below is the complete build log:

gowtham@t-trt-ubuntu:~/server/citritonbuild/pytorch/build$ make clean
gowtham@t-trt-ubuntu:~/server/citritonbuild/pytorch/build$ make
[  4%] Building NVCC (Device) object _deps/repo-backend-build/CMakeFiles/kernel-library-new.dir/src/kernel-library-new_generated_kernel.cu.o
[  9%] Linking CXX static library libkernel-library-new.a
[  9%] Built target kernel-library-new
[ 13%] Extracting pytorch and torchvision libraries and includes from nvcr.io/nvidia/pytorch:21.05-py3
21.05-py3: Pulling from nvidia/pytorch
Digest: sha256:a5986639e4cf01eb35c0c0a9ca9fb9c6f905cc1b546966b78de4f69d15b894cf
Status: Image is up to date for nvcr.io/nvidia/pytorch:21.05-py3
nvcr.io/nvidia/pytorch:21.05-py3
Error: No such container: pytorch_backend_ptlib
error ignored...
bcf92ea4bbbd76853d665e06a3c401dc8319a04ae2db5ebb356aa0862629edec
pytorch_backend_ptlib
[ 13%] Built target ptlib_target
[ 18%] Building CXX object _deps/repo-common-build/CMakeFiles/triton-common-async-work-queue.dir/src/async_work_queue.cc.o
[ 22%] Building CXX object _deps/repo-common-build/CMakeFiles/triton-common-async-work-queue.dir/src/error.cc.o
[ 27%] Linking CXX static library libtritonasyncworkqueue.a
[ 27%] Built target triton-common-async-work-queue
[ 31%] Building CXX object _deps/repo-core-build/CMakeFiles/triton-core-serverstub.dir/src/tritonserver_stub.cc.o
[ 36%] Linking CXX shared library libtritonserver_stub.so
[ 36%] Built target triton-core-serverstub
[ 40%] Building CXX object _deps/repo-backend-build/CMakeFiles/triton-backend-utils.dir/src/backend_common.cc.o
[ 45%] Building CXX object _deps/repo-backend-build/CMakeFiles/triton-backend-utils.dir/src/backend_input_collector.cc.o
[ 50%] Building CXX object _deps/repo-backend-build/CMakeFiles/triton-backend-utils.dir/src/backend_memory.cc.o
[ 54%] Building CXX object _deps/repo-backend-build/CMakeFiles/triton-backend-utils.dir/src/backend_model_instance.cc.o
[ 59%] Building CXX object _deps/repo-backend-build/CMakeFiles/triton-backend-utils.dir/src/backend_model.cc.o
[ 63%] Building CXX object _deps/repo-backend-build/CMakeFiles/triton-backend-utils.dir/src/backend_output_responder.cc.o
[ 68%] Linking CXX static library libtritonbackendutils.a
[ 68%] Built target triton-backend-utils
[ 72%] Building CXX object CMakeFiles/triton-pytorch-backend.dir/src/libtorch.cc.o
In file included from /home/gowtham/server/citritonbuild/pytorch/src/libtorch.cc:43:
/home/gowtham/server/citritonbuild/pytorch/build/include/torchvision/torchvision/vision.h:10:40: warning: ‘_register_ops’ initialized and declared ‘extern’
   10 | extern "C" VISION_INLINE_VARIABLE auto _register_ops = &cuda_version;
      |                                        ^~~~~~~~~~~~~
[ 77%] Building CXX object CMakeFiles/triton-pytorch-backend.dir/src/libtorch_utils.cc.o
[ 81%] Linking CXX shared library libtriton_pytorch.so
/usr/bin/ld: cannot find -ltorch
/usr/bin/ld: cannot find -ltorchvision
collect2: error: ld returned 1 exit status
make[2]: *** [CMakeFiles/triton-pytorch-backend.dir/build.make:121: libtriton_pytorch.so] Error 1
make[1]: *** [CMakeFiles/Makefile2:173: CMakeFiles/triton-pytorch-backend.dir/all] Error 2
make: *** [Makefile:136: all] Error 2

[GowthamKudupudi]

I fixed the issue by replacing -ltorch and -ltorchvision with ${CMAKE_CURRENT_BINARY_DIR}/libtorch.so and ${CMAKE_CURRENT_BINARY_DIR}/libtorchvision.so in CMakeLists.txt