/nim_torch_serialization_experiment

Experiment for loading a traced PyTorch model in Nim

Primary LanguageNim

Using Nim for serving PyTorch models

Trying to replicate the Loading a PyTorch Model in C++ example using Nim.

Currently compiles with:

nim cpp serialization.nim

and runs with:

LD_LIBRARY_PATH="./libtorch/lib" ./serialization

Think have to set -Wl,-rpath,$(pwd)/libtorch/lib in order to not set LD_LIBRARY_PATH.

Analyzing Compilation and Linking

C++ Sample program

When compile the sample program, get the following from the cmake compilation steps.

Compiling:

/usr/bin/c++ \
  -isystem /home/foobar/programming/torch_serialization_experiments/libtorch/include \
  -isystem /home/foobar/programming/torch_serialization_experiments/libtorch/include/torch/csrc/api/include \
  -D_GLIBCXX_USE_CXX11_ABI=0 \
  -std=c++11 \
  -std=gnu++11 \
  -o CMakeFiles/example-app.dir/example-app.cpp.o \
  -c /home/foobar/programming/torch_serialization_experiments/example-app.cpp

Linking:

/usr/bin/c++ -rdynamic CMakeFiles/example-app.dir/example-app.cpp.o \
  -o example-app -Wl,-rpath,/home/foobar/programming/torch_serialization_experiments/libtorch/lib \
  ../libtorch/lib/libtorch.so ../libtorch/lib/libc10.so \
  -Wl,--no-as-needed,/home/foobar/programming/torch_serialization_experiments/libtorch/lib/libcaffe2.so \
  -Wl,--as-needed ../libtorch/lib/libc10.so \
  -lpthread

Nim Program

Compiling:

g++ -c -w -w -fpermissive -I./libtorch/include/ \
  -I'/home/foobar/.choosenim/toolchains/nim-#devel/lib' \
  -I/home/foobar/programming/torch_serialization_experiments \
  -o /home/foobar/.cache/nim/serialization_d/serialization.nim.cpp.o \
  /home/foobar/.cache/nim/serialization_d/serialization.nim.cpp

Linking:

g++ -o /home/foobar/programming/torch_serialization_experiments/serialization  \
  /home/foobar/.cache/nim/serialization_d/stdlib_io.nim.cpp.o \
  /home/foobar/.cache/nim/serialization_d/stdlib_system.nim.cpp.o \
  /home/foobar/.cache/nim/serialization_d/stdlib_times.nim.cpp.o \
  /home/foobar/.cache/nim/serialization_d/stdlib_os.nim.cpp.o \
  '/home/foobar/.cache/nim/serialization_d/_7_7.nimble7pkgs7nimline-#HEAD7nimline.nim.cpp.o' \
  /home/foobar/.cache/nim/serialization_d/serialization.nim.cpp.o  \
  -lm -lrt -L/home/foobar/programming/torch_serialization_experiments/libtorch/lib \
  -ltorch -lcaffe2 -lcaffe2_detectron_ops -lc10 -lfoxi -ldl

Was able to get this working by making sure had the same ABI flag as above when compiling Nim:

nimline.cppdefines("_GLIBCXX_USE_CXX11_ABI=0")

In order to do more than just load the model, need to be able to get data in (Nim sequence to tensor) and get data out (model.forward to Nim sequence). Ideally want to be able to interop with the existing nimtorch library, but don't quite know where C++ types and nimtorch types interact yet (looking at torch/torch_cpp.nim for some of the wrappers around ATen types).

Reminder: File 2 bug reports for nimline

  • converter toStdString() looks to be buggy (generating code with an extra comma).
  • When creating shared pointer for type with SharedPointer[T], tries to dereference the type std::shared_pointer<'*0>, not std::shared_pointer<'0>, which tries to get the base of the type passed in *T -> T. This somehow works with older versions of Nim.