Integration with PyTorch about k2 HOT 6 CLOSED

RE 1): Is the build system supposed to work with any installation of PyTorch? e.g. if users build PyTorch from source, or install using Conda
RE 2): Why do we need to replace CudaContext? I thought we can just create a new class PyTorchContext which inherits from Context, then we can operate PyTorch tensor (no matter GPU or CPU, exposed with DLPack) with current interfaces (e.g. Array1) and algorithms. Correct me if I'm wrong.

from k2.

It takes several hours to build PyTorch from source. Another problem to build from source is that there is a good chance the PyTorch version K2 uses is different from the one Python uses.

The current CudaContext is a naive implementation. What I want to do is to change GetCudaContext so that
it returns a pytorch context.

I have no experience with conda. But I guess it will work.

from k2.

from k2.

To the best of my knowledge, there are two approaches to build against PyTorch:

(1) Build PyTorch from source. I think only PyTorch developers use this method. The building process can take several hours.
It's very likely that the source version is different from the one installed with pip.

(2) Build against LibTorch. https://pytorch.org/ provides links to download libtorch-shared-with-deps-xxx.zip, which
contains a set of header files and shared/static libraries. These files are also contained in the PyTorch package installed via pip install torch. pip install torch is chosen so that we use the same PyTorch version with Python.

cd build/bin
readelf -d cu_array_test

prints

 0x0000000000000001 (NEEDED)             Shared library: [libfsa.so]
 0x0000000000000001 (NEEDED)             Shared library: [libgtest_maind.so]
 0x0000000000000001 (NEEDED)             Shared library: [libtorch.so]
 0x0000000000000001 (NEEDED)             Shared library: [libtorch_cpu.so]
 0x0000000000000001 (NEEDED)             Shared library: [libtorch_cuda.so]
 0x0000000000000001 (NEEDED)             Shared library: [libc10_cuda.so]
 0x0000000000000001 (NEEDED)             Shared library: [libc10.so]
 0x0000000000000001 (NEEDED)             Shared library: [libnvToolsExt.so.1]
 0x0000000000000001 (NEEDED)             Shared library: [libcudart.so.10.0]
 0x0000000000000001 (NEEDED)             Shared library: [libgtestd.so]
 0x0000000000000001 (NEEDED)             Shared library: [libstdc++.so.6]
 0x0000000000000001 (NEEDED)             Shared library: [libgcc_s.so.1]
 0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]
 0x000000000000000f (RPATH)              Library rpath: [/xxx/pyenv/versions/3.7.8/lib/python3.7/site-packages/torch/lib:/xxx/cuda/lib64:/xxx/cuda/lib64/stubs]

Part of the linking commands for cu_array_test is given below:

/usr/bin/g++ xxx/array_test.cu.o -o ../../bin/cu_array_test  
-Wl,-rpath,/xxx/k2/build/lib:/xxx/pyenv/versions/3.7.8/lib/python3.7/site-packages/torch/lib
/xxx/pyenv/versions/3.7.8/lib/python3.7/site-packages/torch/lib/libtorch.so 
-Wl,--no-as-needed,/xxx/pyenv/versions/3.7.8/lib/python3.7/site-packages/torch/lib/libtorch_cpu.so 
-Wl,--no-as-needed,/xxx/pyenv/versions/3.7.8/lib/python3.7/site-packages/torch/lib/libtorch_cuda.so 
-Wl,--as-needed /xxx/pyenv/versions/3.7.8/lib/python3.7/site-packages/torch/lib/libc10_cuda.so 
/xxx/pyenv/versions/3.7.8/lib/python3.7/site-packages/torch/lib/libc10.so

We can see that k2 uses the same PyTorch libraries from the torch package.

Note that there are no restrictions on the PyTorch version. Users can choose which PyTorch version to
use by pip install torch==x.x.x and K2 will pick the corresponding version correctly.

from k2.

how does your pull request relate to this, in terms of PyTorch versions and compatibility?

K2 uses the PyTorch installation information from the pip package, so it's guaranteed that k2 uses whichever version Python
is currently using.

from k2.

from k2.