Building ktorch¶
The ktorch library has been built on Linux and MacOS; it has not been tested on Windows. (Windows is beginning to get more support with Microsoft becoming the maintainer of the Windows version in 2020.)
The first step is to download the relevant zip file from PyTorch.
The k interface is built with the latest version of PyTorch as of Apr 2024, labeled Stable(2.3.0).
For Linux, the zip file contains all the necessary libraries and include files; there is no need to install CUDA or Intel MKL as these components are included. The zip file is large, around 2 gigabytes for versions which include libraries for working with GPU’s and around 200 megabytes for CPU-only. Each platform has additional choices for CPU-only/GPU versions (more here on pre c++11 and post c++11 ABI versions).
For MacOS, the libtorch zip file distributed by PyTorch is for Apple’s M1/M2/M3 chips. It is also possible to compile the libtorch libraries to incude a CUDA setup on Mac or use Intel CPU. See this section on how build the PyTorch c++ libraries from source before building the k interface.
Below is a brief summary of retrieval steps for a Linux machine with Nvidia GPU(s):
Download and unzip the file for CUDA 11.8 with the pre-c++11 ABI:
> wget -q https://download.pytorch.org/libtorch/cu118/libtorch-shared-with-deps-2.3.0%2Bcu118.zip
> ls -lh libtorch-shared-with-deps-2.3.0+cu118.zip
-rw-rw-r-- 1 t t 2.5G Apr 23 10:12 libtorch-shared-with-deps-2.3.0+cu118.zip
> unzip libtorch-shared-with-deps-2.3.0+cu118.zip
Archive: libtorch-shared-with-deps-2.3.0+cu118.zip
creating: libtorch/
creating: libtorch/lib/
inflating: libtorch/lib/libasmjit.a
inflating: libtorch/lib/libbackend_with_compiler.so
inflating: libtorch/lib/libbenchmark.a
inflating: libtorch/lib/libbenchmark_main.a
inflating: libtorch/lib/libc10_cuda.so
..
Download the ktorch source code via git clone:
> cd ~
> git clone https://github.com/ktorch/ktorch.git
Cloning into 'ktorch'...
Or download as a zip file:
> wget --quiet https://github.com/ktorch/ktorch/archive/refs/heads/master.zip
> unzip master.zip
Archive: master.zip
creating: ktorch-master/
inflating: ktorch-master/LICENSE
inflating: ktorch-master/Makefile
..
Assuming the libtorch libraries are in ~/libtorch and the ktorch source in ~/ktorch:
> make
(see sample builds for mac on intel, with M1/M2 chips and linux for more detail)
PyTorch zip files¶
The libtorch.zip files from PyTorch are saved by version: when version 2.3.0 is no longer the latest version, it is still possible to retrieve the version-specific files:
Makefile¶
The makefile can be changed to suit preferences. There are 3 main variables, CXX, TORCH and ABI, that may need to be changed in the file itself or specified on the command line.
CXX¶
The c++ compiler defaults to clang. To run with GCC:
make CXX=g++
TORCH¶
TORCH has the location of the c++ libraries and include files for PyTorch. Default is set to ~/libtorch.
make TORCH=/customdir/libtorch
ABI¶
In Linux, there’s a choice of ABI (application binary interface). Changes in the C++11 standard created a newer ABI. The supplied libtorch zip files from PyTorch come in two versions, one for the ABI prior to the changes for the C++11 standard, and one with the new ABI.
For example, for Linux, version 2.3.0, with support for CUDA 11.8, the zip files are listed as:
Download here (Pre-cxx11 ABI):
https://download.pytorch.org/libtorch/cu118/libtorch-shared-with-deps-2.3.0%2Bcu118.zip
Download here (cxx11 ABI):
https://download.pytorch.org/libtorch/cu118/libtorch-cxx11-abi-shared-with-deps-2.3.0%2Bcu118.zip
In their earlier versions, PyTorch only offered the older ABI with their zip files so users could maintain compatibility with older third-party libraries compiled under the old ABI, but now PyTorch offers the choice of old or new versions.
By default, the Makefile builds code with -D_GLIBCXX_USE_CXX11_ABI=0 for the older API.
The Makefile variable ABI is set to 0, but can be overwritten with the command-line call ABI=1 if the newer ABI zip file is used.
Source files¶
LICENSE - MIT license
docs/ - reStructuredText files for documentation at ktorch.readthedocs.io.
ktorch.h - main header file, which, in turn includes headers from PyTorch.
ktorch.cpp - contains the code used by the rest of system dealing with tensors, modules, optimizers, etc.
ktensor.cpp - code for operating on tensors
kmath.cpp - PyTorch math routines
knn.h - include k-api fns for PyTorch modules and custom module definitions
knn.cpp - code for building modules and sequences of modules
knn/ - custom modules and code to parse k args defined here
kloss.h - redefine binary cross entropy loss functions, add smooth cross entropy
kloss.cpp - code relating to loss functions and modules
kopt.h - include custom optimizer definitions
kopt.cpp - optimizer code
kopt/ - custom optimizers not found in PyTorch release
kmodel.cpp - code for building models (module + optimizer + loss function)
ktest.cpp - contains temporary tests, samples, etc. – nothing essential to the interface library
private.h - macros to gain access to private class elements, from martong.
stb_image_write.h - minimal code to write .png files, from stb.
Build libtorch from source¶
To customize the PyTorch c++ libraries or to build to a different setup than supplied by PyTorch, it is also possible to build the libraries from source.
For a mac with an Intel chip, since PyTorch no longer distributes binaries for x86_64, it’s necessary to build libtorch from source.
Installing some prerequisites for macOS:
brew install cmake
pip install pyyaml
Then download the source code for PyTorch as of release 2.3.0:
> git clone -b v2.3.0 --recurse-submodule https://github.com/pytorch/pytorch.git
> mkdir pytorch-build
> cd pytorch-build
> cmake -DUSE_MPS=ON -DBUILD_SHARED_LIBS:BOOL=ON -DCMAKE_BUILD_TYPE:STRING=Release -DPYTHON_EXECUTABLE:PATH=`which python3` -DCMAKE_INSTALL_PREFIX:PATH=../pytorch-install ../pytorch
> cmake --build . --target install
Once the build completes, the libraries are in the directory: pytorch-install and can be used to build the k interface:
> cd ~/ktorch
> make TORCH=/Users/t/pytorch-install
More notes on building the PyTorch c++ libraries via cmake as well as via python are available here.
Sample builds¶
MacOS, Intel CPU¶
First step, get the CPU-only version of libtorch 2.2.2 for MacOS:
> cd ~
> wget -q https://download.pytorch.org/libtorch/cpu/libtorch-macos-arm64-2.3.0.zip
> ls -lh libtorch-macos-arm64-2.3.0.zip
-rw-r--r-- 1 t staff 57M Apr 23 10:09 libtorch-macos-arm64-2.3.0.zip
> rm -rf ~/libtorch # erase any previous version
> unzip libtorch-macos-arm64-2.3.0.zip
Archive: libtorch-macos-arm64-2.3.0.zip
creating: libtorch/
creating: libtorch/bin/
inflating: libtorch/build-hash
creating: libtorch/include/
..
> ls libtorch
bin/ build-hash build-version include/ lib/ share/
Next, clone the ktorch repository:
> rm -rf ~/ktorch # remove any previous dir named ktorch
> git clone https://github.com/ktorch/ktorch.git
Cloning into 'ktorch'...
Build using make:
> cd ktorch
> time make
clang -std=c++14 -std=gnu++14 -pedantic -Wall -Wfatal-errors -fPIC -O3 -I /Users/t/libtorch/include -I /Users/t/libtorch/include/torch/csrc/api/include -c -o ktorch.o ktorch.cpp
clang -std=c++14 -std=gnu++14 -pedantic -Wall -Wfatal-errors -fPIC -O3 -I /Users/t/libtorch/include -I /Users/t/libtorch/include/torch/csrc/api/include -c -o ktensor.o ktensor.cpp
..
clang -o ktorch.so ktorch.o ktensor.o kmath.o knn.o kloss.o kopt.o kmodel.o ktest.o knn/act.o knn/attention.o knn/callback.o knn/conv.o knn/distance.o knn/drop.o knn/embed.o knn/fns.o knn/fold.o knn/fork.o knn/linear.o knn/nbeats.o knn/norm.o knn/onehot.o knn/pad.o knn/recur.o knn/reshape.o knn/residual.o knn/select.o knn/seq.o knn/squeeze.o knn/transform.o knn/transformer.o knn/upsample.o knn/util.o kopt/lamb.o -undefined dynamic_lookup -shared -L/Users/t/libtorch/lib -l torch -Wl,-rpath /Users/t/libtorch/lib
real 6m32.023s
user 6m18.607s
sys 0m12.748s
Faster compile (1-2 minutes) is possible with the -j option:
> make -s clean
> time make -sj
real 1m42.412s
user 10m11.067s
sys 0m22.923s
> ls -lh ./ktorch.so
-rwxr-xr-x 1 t staff 4.3M May 22 15:11 ktorch.so*
Check if the ktorch.so library can be loaded from within a k session:
> q
KDB+ 4.0 2022.10.26 Copyright (C) 1993-2022 Kx Systems
m64/ 8(16)core 32768MB
q).nn:(`ktorch 2:`fns,1)[] / define interface functions in .nn
q).nn.setting[]
mkl | 1b /Intel's MKL libraries are available
openmp | 0b
threads | 4
interopthreads | 4
mps | 0b
cuda | 0b
magma | 0b
cudnn | 0b
cudnnversion | 0N
cudadevices | 0
benchmark | 0b
deterministic | 0
cudnndeterministic| 0b
stackframe | 0b
alloptions | 1b
complexfirst | 1b
Checking the configuration:
q).nn.config[]
PyTorch built with:
- GCC 4.2
- C++ Version: 201703
- clang 13.1.6
- Intel(R) oneAPI Math Kernel Library Version 2022.2-Product Build 20220801 for Intel(R) 64 architecture applications
- Intel(R) MKL-DNN v2.7.3 (Git Hash 6dbeffbae1f23cbbeae17adb7b5b13f1f37c080e)
- LAPACK is enabled (usually provided by MKL)
- NNPACK is enabled
- CPU capability usage: AVX2
- Build settings: BLAS_INFO=mkl, BUILD_TYPE=Release, CXX_COMPILER=/Applications/Xcode_13.3.1.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/c++, CXX_FLAGS= -Wno-deprecated -fvisibility-inlines-hidden -Wno-deprecated-declarations -DUSE_PTHREADPOOL -DNDEBUG -DUSE_KINETO -DLIBKINETO_NOCUPTI -DLIBKINETO_NOROCTRACER -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DUSE_PYTORCH_METAL_EXPORT -DSYMBOLICATE_MOBILE_DEBUG_HANDLE -DUSE_COREML_DELEGATE -O2 -fPIC -Wall -Wextra -Werror=return-type -Werror=non-virtual-dtor -Werror=braced-scalar-init -Werror=range-loop-construct -Werror=bool-operation -Winconsistent-missing-override -Wnarrowing -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wunused-local-typedefs -Wno-unused-parameter -Wno-unused-function -Wno-unused-result -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wvla-extension -Wno-range-loop-analysis -Wno-pass-failed -Wsuggest-override -Wno-error=pedantic -Wno-error=redundant-decls -Wno-error=old-style-cast -Wconstant-conversion -Wno-invalid-partial-specialization -Wno-typedef-redefinition -Wno-unused-private-field -Wno-inconsistent-missing-override -Wno-constexpr-not-const -Wno-missing-braces -Wunused-lambda-capture -Wunused-local-typedef -Qunused-arguments -fcolor-diagnostics -fdiagnostics-color=always -fno-math-errno -fno-trapping-math -Werror=format -Werror=cast-function-type -DUSE_MPS -fno-objc-arc -Wno-unguarded-availability-new -Wno-unused-private-field -Wno-missing-braces -Wno-constexpr-not-const, LAPACK_INFO=mkl, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, TORCH_DISABLE_GPU_ASSERTS=OFF, TORCH_VERSION=2.0.1, USE_CUDA=OFF, USE_CUDNN=OFF, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=OFF, USE_NNPACK=ON, USE_OPENMP=OFF, USE_ROCM=OFF,
ATen/Parallel:
at::get_num_threads() : 4
at::get_num_interop_threads() : 4
OpenMP not found
Intel(R) oneAPI Math Kernel Library Version 2022.2-Product Build 20220801 for Intel(R) 64 architecture applications
mkl_get_max_threads() : 1
Intel(R) MKL-DNN v2.7.3 (Git Hash 6dbeffbae1f23cbbeae17adb7b5b13f1f37c080e)
std::thread::hardware_concurrency() : 8
Environment variables:
OMP_NUM_THREADS : [not set]
MKL_NUM_THREADS : [not set]
ATen parallel backend: native thread pool
To make the ktorch.so library available to q sessions without specifying a path, can do something like the following:
ln -s $(pwd)/ktorch.so ~/q/m64
Once the library is built, it can be tested with some examples:
> cd
> rm -rf examples
> git clone https://github.com/ktorch/examples.git
> q examples/start/spirals.q
KDB+ 4.0 2022.10.26 Copyright (C) 1993-2022 Kx Systems
m64/ 8(16)core 32768MB
0 2 2 2 2 2 2
0 0 2 2 2 2 2 2 2 2
0 0 2 2 2 2 2 2 2 2 2 2 2
0 0 0 2 2 2 2 2 2 2
0 0 0 2 2 2 1 2 2 2
0 0 0 2 2 2 2 1 1 1 2 2
0 0 0 2 2 2 1 1 1 1 1 1 2 2
0 0 0 2 2 2 2 1 1 1 1 1 1 1 1 2
0 0 0 2 2 2 1 1 1 1 1 1 1
0 0 0 2 2 2 2 2 0 0 1 1 1
0 0 0 2 2 2 2 2 0 0 1 1 1
0 0 0 0 2 2 2 0 0 0 0 1 1 1
0 0 0 0 0 0 0 0 0 0 1 1 1
0 0 0 0 0 0 0 0 0 1 1 1
0 0 0 0 0 0 0 1 1 1 1
0 1 1 1
1 1 1 1
1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1
1 1 1 1 1
1456 1360
Accuracy on training data: 99.93333%
Accuracy using new sample: 99.9%
MacOS, M1/M2¶
As of release 2.0.1, PyTorch does not release the c++ libraries compiled for Apple’s new M1 & M2 chips.
It is necessary to build these libraries from source or link to the c++ libraries that are part of a python installation for Mac.
Once the libtorch c++ libraries are built, clone the k api repository.
> cd ~
> git clone https://github.com/ktorch/ktorch.git
Cloning into 'ktorch'...
Assuming the libtorch c++ libraries and include files are in the default install dir, ~/pytorch-install:
> make TORCH=/Users/t/pytorch-install
clang -std=c++14 -std=gnu++14 -pedantic -Wall -Wfatal-errors -fPIC -O3 -I /Users/t/pytorch-install/include -I /Users/t/pytorch-install/include/torch/csrc/api/include -c -o ktorch.o ktorch.cpp
clang -std=c++14 -std=gnu++14 -pedantic -Wall -Wfatal-errors -fPIC -O3 -I /Users/t/pytorch-install/include -I /Users/t/pytorch-install/include/torch/csrc/api/include -c -o ktensor.o ktensor.cpp
..
clang -o ktorch.so ktorch.o ktensor.o kmath.o knn.o kloss.o kopt.o kmodel.o ktest.o knn/act.o knn/attention.o knn/callback.o knn/conv.o knn/distance.o knn/drop.o knn/embed.o knn/fns.o knn/fold.o knn/fork.o knn/linear.o knn/nbeats.o knn/norm.o knn/onehot.o knn/pad.o knn/recur.o knn/reshape.o knn/residual.o knn/select.o knn/seq.o knn/squeeze.o knn/transform.o knn/transformer.o knn/upsample.o knn/util.o kopt/lamb.o -undefined dynamic_lookup -shared -L/Users/t/pytorch-install/lib -l torch -Wl,-rpath /Users/t/pytorch-install/lib
Instead of building the libtorch c++ libraries, an alternate method is to use the libraries distributed as part of the PyTorch installation of the python interface. For example, using the conda install from PyTorch:
> conda install pytorch torchvision torchaudio -c pytorch
After the install, search for the main library:
> find ~/miniconda3/lib -name libtorch.dylib
/Users/t/miniconda3/lib/python3.10/site-packages/torch/lib/libtorch.dylib
Using the c++ libraries and include files that were included as part of the python install via conda:
> make TORCH=/Users/t/miniconda3/lib/python3.10/site-packages/torch
clang -std=c++14 -std=gnu++14 -pedantic -Wall -Wfatal-errors -fPIC -O3 -I /Users/t/miniconda3/lib/python3.10/site-packages/torch/include -I /Users/t/miniconda3/lib/python3.10/site-packages/torch/include/torch/csrc/api/include -c -o ktorch.o ktorch.cpp
..
Build time is around 2 minutes on a macbook pro with the M2 max chip; down to about 25 seconds with the parallel compilation flag, e.g. make -j ..
Using pip instead of conda:
> pip3 install torch torchvision torchaudio
> pip3 show torch|grep ^Location:
Location: /opt/homebrew/lib/python3.11/site-packages
> make -sj TORCH=/opt/homebrew/lib/python3.11/site-packages/torch
> ls -lh ktorch.so
-rwxr-xr-x 1 t staff 4.0M May 25 08:55 ktorch.so*
> otool -L ktorch.so
ktorch.so:
ktorch.so (compatibility version 0.0.0, current version 0.0.0)
@rpath/libtorch.dylib (compatibility version 0.0.0, current version 0.0.0)
/usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 1319.100.3)
> otool -l ktorch.so |grep LC_RPATH -A 2
cmd LC_RPATH
cmdsize 72
path /opt/homebrew/lib/python3.11/site-packages/torch/lib (offset 12)
Loading the api functions into q:
q){key[x]set'get x}(`ktorch 2:`fns,1)[]; /define interface fns in root
q)setting[]
mkl | 0b
openmp | 0b
threads | 12
interopthreads | 12
mps | 1b
cuda | 0b
magma | 0b
cudnn | 0b
cudnnversion | 0N
cudadevices | 0
..
q)help`device /display devices and initial random seed
cpu | 7993899427782217949
mps | 6228276533503624572
mps:0| 6228276533503624572
q)x:tensor(`randn; 4096 1024)
q)y:tensor(`randn; 1024 4096)
q)r:tensor()
q)\ts:10 use[r]mm(x;y)
196 1184
q){to(x;`mps)}'[(x;y;r)]; /move to metal performance shaders
q)\ts use[r]mm(x;y) /first use overhead
55 1184
q)\ts:10 use[r]mm(x;y)
3 1184
Linux, CUDA 11.7¶
Build in /tmp, using the libtorch zip file for Linux, version 2.0.1, CUDA 11.7 with the newer c++ ABI.
> cd /tmp
> rm -rf libtorch
> wget -q https://download.pytorch.org/libtorch/cu117/libtorch-cxx11-abi-shared-with-deps-2.0.1%2Bcu117.zip
> ls -lh libtorch-cxx11-abi-shared-with-deps-2.0.1+cu117.zip
-rw-rw-r-- 1 t t 1.9G May 5 09:45 libtorch-cxx11-abi-shared-with-deps-2.0.1+cu117.zip
> unzip -q libtorch-cxx11-abi-shared-with-deps-2.0.1+cu117.zip
> ls libtorch
bin/ build-hash build-version include/ lib/ share/
Get the ktorch repository as a zip file:
> wget -q https://github.com/ktorch/ktorch/archive/refs/heads/master.zip
> unzip -q master.zip
Build with the ABI flag set on and the TORCH location pointing to the /tmp/torchlib directory, using clang, the default compiler:
> cd ktorch-master
> time make ABI=1 TORCH=/tmp/libtorch
clang -std=c++14 -std=gnu++14 -pedantic -Wall -Wfatal-errors -fPIC -O3 -D_GLIBCXX_USE_CXX11_ABI=1 -I /tmp/libtorch/include -I /tmp/libtorch/include/torch/csrc/api/include -c -o ktorch.o ktorch.cpp
clang -std=c++14 -std=gnu++14 -pedantic -Wall -Wfatal-errors -fPIC -O3 -D_GLIBCXX_USE_CXX11_ABI=1 -I /tmp/libtorch/include -I /tmp/libtorch/include/torch/csrc/api/include -c -o ktensor.o ktensor.cpp
..
clang -o ktorch.so ktorch.o ktensor.o kmath.o knn.o kloss.o kopt.o kmodel.o ktest.o knn/act.o knn/attention.o knn/callback.o knn/conv.o knn/distance.o knn/drop.o knn/embed.o knn/fns.o knn/fold.o knn/fork.o knn/linear.o knn/nbeats.o knn/norm.o knn/onehot.o knn/pad.o knn/recur.o knn/reshape.o knn/residual.o knn/select.o knn/seq.o knn/squeeze.o knn/transform.o knn/transformer.o knn/upsample.o knn/util.o kopt/lamb.o -shared -L/tmp/libtorch/lib -l torch -Wl,-rpath /tmp/libtorch/lib
real 5m55.095s
user 5m42.950s
sys 0m12.081s
The build can be faster with parallel compilation if ordered output isn’t required:
> make -s clean
> time make -sj ABI=1 TORCH=/tmp/libtorch
real 1m9.112s
user 9m50.856s
sys 0m17.043s
Load in a k session, check version and settings:
> pwd
/tmp/ktorch-master
> ls -lh ktorch.so
-rwxrwxr-x 1 t t 5.6M May 23 06:40 ktorch.so*
> mv ktorch.so ktorchtmp.so #avoid confusion w'any other existing ktorch.so
> q
KDB+ 4.0 2022.10.26 Copyright (C) 1993-2022 Kx Systems
l64/ 12(24)core 64025MB
q){key[x]set'x}(`ktorchtmp 2:`fns,1)[]; /define api fns in root
q)version[]
2.0001
q)version()
"2.0.1"
q)setting[]
mkl | 1b
openmp | 1b
threads | 6
interopthreads | 6
mps | 0b
cuda | 1b
magma | 1b
cudnn | 1b
cudnnversion | 8500
cudadevices | 2
..
q)config[]
PyTorch built with:
- GCC 9.3
- C++ Version: 201703
- Intel(R) oneAPI Math Kernel Library Version 2021.4-Product Build 20210904 for Intel(R) 64 architecture applications
- Intel(R) MKL-DNN v2.7.3 (Git Hash 6dbeffbae1f23cbbeae17adb7b5b13f1f37c080e)
- OpenMP 201511 (a.k.a. OpenMP 4.5)
- LAPACK is enabled (usually provided by MKL)
- NNPACK is enabled
- CPU capability usage: AVX2
- CUDA Runtime 11.7
..
Check matrix multiply on GPU if available:
q)setting`cuda
1b
q)a:tensor(`randn;4096 1024;`cuda`double)
q)b:tensor(`randn;1024 4096;`cuda`double)
q)\ts r:mm(a;b)
208 1200
q)\ts use[r]mm(a;b)
1 1184
q)to(a;`cpu) /move tensors to cpu
q)to(b;`cpu)
q)\ts use[r]mm(a;b)
112 1184
q)x:tensor a /run q's matrix multiply
q)y:tensor b
q)\ts z:x$y
3421 268501328
q)equal(z;r)
0b
q)allclose(z;r)
1b
q)(avg;max)@\:abs raze over z-tensor r
2.603255e-14 4.831691e-13
Linked libraries¶
During the link stage of the build, the path of the PyTorch libraries are added via -rpath so that the same libraries can be located at runtime.
From the above Linux build example in /tmp:
clang -o ktorch.so ktorch.o ktensor.o kmath.o knn.o .. kopt/lamb.o -shared -L/tmp/libtorch/lib -l torch -Wl,-rpath /tmp/libtorch/lib
> ldd ktorchtmp.so
linux-vdso.so.1 (0x00007ffd33fda000)
libtorch.so => /tmp/libtorch/lib/libtorch.so (0x00007ff5a5600000)
..
If the location of the libtorch/lib subdirectory is changed or in a different place on the deployment machine,
then the environment variable LD_LIBRARY_PATH can be used to point to a new location for the PyTorch shared libraries.
> mv /tmp/libtorch /tmp/torch
> ldd ktorchtmp.so
linux-vdso.so.1 (0x00007ffc48b03000)
libtorch.so => not found
..
> export LD_LIBRARY_PATH=/tmp/torch/lib
> ldd ktorchtmp.so
linux-vdso.so.1 (0x00007ffe67bbf000)
libtorch.so => /tmp/torch/lib/libtorch.so (0x00007f3a14200000)
..
Location of ktorch.so¶
In most of the examples in this documentation, the k api functions in the shared library, typically named ktorch.so, are loaded via 2: without any path.
q)(`ktorch 2:`options,1)[] / show default options
device | cpu
dtype | float
layout | strided
gradient| nograd
pin | unpinned
memory | contiguous
This will work if the ktorch.so file is placed in, for 64-bit linux, ~/q/l64 or ${QHOME}/l64 or a symbolic link is placed there to the actual location.
> ls -l ~/q/l64/ktorch.so
lrwxrwxrwx 1 t t 24 Dec 2 14:07 /home/t/q/l64/ktorch.so -> /home/t/ktorch/ktorch.so*
An alternative is to use the full path directly or via some agreed upon environment variable.
> cd /tmp
> q
q)(`:/home/t/ktorch/ktorch 2:`options,1)[]
device | cpu
dtype | float
..
q)`KTORCH setenv "/home/t/ktorch/ktorch"
q)((`$getenv`KTORCH)2:`options,1)[]
device | cpu
dtype | float
..
Defining api functions in k¶
The api function fns, when called with an empty or dummy argument, returns a dictionary of function name and code.
q)(`ktorch 2:`fns,1)[]
dv | code
tree | code
addref | code
free | code
..
The result of this function can be assigned to a to a namespace:
q).nn:(`ktorch 2:`fns,1)[]
q)t:.nn.tensor 1 2 3
q).nn.tensor t
1 2 3
or defined in the root namespace:
q){key[x]set'x}(`ktorch 2:`fns,1)[];
q)t:tensor 1 2 3
q)tensor t
1 2 3