IMPLEMENTING.md

Implementing GRACE components

Compressor

The Compressor class has two methods, compress and decompress. As described in the paper, there are many ways to implement a compressor. According to the categorization of each compressor, additional utility functions can be used such as quantize, dequantize, sparisify, desparsify, pack, and unpack. Examples of these functions are implemented for each compressor and can be found bundled with each compressor.

• compress is the compression operator. It expects a single tensor as input, and returns a collection of tensors along with a context. These values should be passed to decompress as is.
• decompress takes the output of compress and returns a single tensor.

A developer should keep in mind that only the collection of tensors is communicated across workers. The context is kept locally for each worker.

Generally it is expected, but not enforced, that the result of decompress(**compress(tensor)) has the same shape and data type as tensor. A good compressor would try to keep these values very close, but also try to reduce the size of the collection of tensors in comparison with the original tensor.

Memory

The Memory class has two methods, update and compensate. As described in the paper, a typical residual memory implementation is provided by ResidualMemory.

• The update method would store the residual difference between the original tensor and the decompression of the compressed original tensor (i.e. residual[name] = tensor - decompress(compress(tensor))). Generally, an update method expects the original tensor, the name that uniquely identifies it, the compressor (used for decompression), and the output of compressor.compress(tensor). This method has no return value, with an exception for TensorFlow 1.15, in which case some Variable.assign operations may be needed.
• The compensate in this case has to apply the residual memory to the next iteration over this tensor. This is achieved by adding the new tensor with the residual memory stored for the tensor name which is used to uniquely identify each tensor (i.e. new_tensor = residula[name] + tensor). To make ResidualMemory a bit more general, optional parameters can be provided during construction. Generally, a compensate method expects the original tensor, and the name that uniquely identifies it. This method has returns the compensated tensor.

Other compression-specific implementations of Memory can be found in the same directory. There is no specific reason not to bundle a compression-specific Memory implementation with the appropriate Compressor implementation.

Communicator

Three implementations of Communicator classes are provided for each framework, each with a different collective method. A Communicator implementation needs to override at least the send_receive method, or in the case of Horovod PyTorch which splits this step to the async_send and wait_receive methods. These methods expect a collection of tensors, the context which is the result of compression.compress method, and the name of the compressed tensor. It has to be noted that not all communicator are compatible with every compressor. A notable example is the Allreduce communicator which can be used with RandomKCompressor but not with TopKCompressor.

Putting it all together

# Horovod PyTorch
from grace_dl.torch import Communicator, Compressor, Memory

class MyCommunicator(Communicator):
    ...

class MyCompressor(Compressor):
    ...

class MyMemory(Memory):
    ...

grc = MyCommunicator(MyCompressor(...), MyMemory(...), ...)