Skip to content

Latest commit

 

History

History
132 lines (95 loc) · 4.59 KB

README.md

File metadata and controls

132 lines (95 loc) · 4.59 KB
Raw

PyTorch Bindings

We provide (so far experimental) bindings to PyTorch. The bindings are specified in bindings.cc, usage examples can be found in example.py.

We provide bindings for the main primitives of AdaPM: pull, push, intent, advance_clock, prepare_sample, and pull_sample. They further provide a set variant of the push primitive that sets parameter to specific values (instead of adding to the values, as push does).

PyTorch tensors

keys = torch.LongTensor([1,2,3,4])
values = torch.ones((len(keys)*10), dtype=torch.float32)

# pull
kv.pull(keys, values)

# push
kv.push(keys, values)

# set (variant of push)
kv.set(keys, values)

# signal intent and advance clock
kv.intent(keys, 2, 3)

kv.advance_clock()

NumPy arrays

The bindings also accept numpy arrays:

keys = np.array([1,2,3,4])
values = np.ones(len(keys)*10, dtype=np.float32)

# pull
kv.pull(keys, values)

# push
kv.push(keys, values)

# set (variant of push)
kv.set(keys, values)

# signal intent and advance clock
kv.intent(keys, 2, 3)

kv.advance_clock()

Synchronous and asynchronous operations

By default, operations run synchronously. To run asynchronously, pass async=True to the operation:

kv.pull(keys, values, True)
kv.push(keys, values, True)

In particular, push is often executed asynchronously.

kv.wait() explicitly waits for the execution of a specific operation:

timestamp = kv.pull(keys, values, True)

# do something else

kv.wait(timestamp) # wait for pull to finish

Parameter access operations (pull, push, set, and pull_sample) return a timestamp that can be used this way.

Installation

To compile working bindings, it is important that AdaPM is built with a key data type that matches PyTorch and NumPy default integer data types (int64) and that AdaPM is built with the the same C++ ABI as the installed PyTorch. If you installed PyTorch in a normal way (e.g., pip or conda), it is likely that this PyTorch installation uses the old pre-C++11 ABI (apparently, this is done to ensure compatibility to old Python versions). If you are on a recent OS, this will be incompatible with system-provided Protocol Buffers. Thus, you will probably have to install Protocol Buffers manually. To check the ABI of your installed PyTorch, run python bindings/lookup_torch_abi.py (0 means pre-C++11 ABI, 1 means C++ ABI). If that script returns 0 and you are on a recent OS with C++ ABI (e.g., compare echo '#include <string>' | g++ -x c++ -E -dM - | fgrep _GLIBCXX_USE_CXX11_ABI), you have to install Protocol Buffers manually. (Sorry :/)

To manually install protocol buffers with the pre-c++11 ABI, use:

# (in the root folder of this repository)
wget https://github.com/protocolbuffers/protobuf/releases/download/v3.6.1/protobuf-cpp-3.6.1.tar.gz
tar zxf protobuf-cpp-3.6.1.tar.gz
cd protobuf-3.6.1
mkdir release && cd release
cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=$(pwd)/../../deps/ -D CMAKE_CXX_FLAGS="-fPIC -D_GLIBCXX_USE_CXX11_ABI=0" -D protobuf_BUILD_TESTS=OFF ../cmake/
make -j  # this can take a while
make install

Then to compile AdaPM with the matching key type and the ABI of the installed PyTorch:

# (in the root folder of this repository)
cmake -S . -B build_bindings -D PS_KEY_TYPE=int64_t -DPROTOBUF_PATH=$(pwd)/deps/lib/cmake/protobuf/ -D CMAKE_CXX_FLAGS="-D_GLIBCXX_USE_CXX11_ABI=$(python bindings/lookup_torch_abi.py)"
cmake --build build_bindings --target adapm

Note Make sure that PyTorch is installed before you build AdaPM. Otherwise, the ABI read-out ($(python bindings/lookup_torch_abi.py) above) will fail. Also make sure that you run the read-out with the python that you are using (i.e., potentially python3 or python3.9 rather than python).

And finally, to compile the bindings:

cd bindings
python setup.py install --user

Note If you built AdaPM to another path than build_bindings/ or you installed Protocol Buffers to a folder other than deps/, you need to pass these paths to setup.py explicitly via environment variables BUILD_PATH and DEPS_PATH (using absolute paths). E.g.: BUILD_PATH=[ABSOLUTE_PATH_TO_BUILD] DEPS_PATH=[ABSOLUTE_PATH_TO_PROTOBUF_INSTALL] python setup.py install --user

If successful, you can now use AdaPM from Python

#!/usr/bin/python
import torch
import adapm

You should also be able to run python bindings/example.py without error messages.

Experimental status

These bindings have an experimental status. If you run into any problems or have questions, please don't hesitate to contact us.