xbtorch.optim

XBTorch-wrapped optimizers with support for WAGE quantization, device-aware updates, and decomposition algorithms

This module wraps standard PyTorch optimizers (SGD and Adam) with XBTorch-specific functionality for:

• Gradient quantization (WAGE)

• Device weight modeling

• Decomposition algorithms

• Weight clipping to specified ranges

Classes

SGD

XBTorch-wrapped stochastic gradient descent optimizer.

Adam

XBTorch-wrapped Adam optimizer.

Classes

Adam(*args, **kwargs)
SGD(*args, **kwargs)

class xbtorch.optim.Adam(*args, **kwargs)

Bases: Adam

step()

Custom step function implementing:

• Gradient decomposition

• Device weight updates

• WAGE quantization

• Optional SW clipping

• Calls original optimizer step at the end

class xbtorch.optim.SGD(*args, **kwargs)

Bases: SGD

step()

Custom step function implementing:

• Gradient decomposition

• Device weight updates

• WAGE quantization

• Optional SW clipping

• Calls original optimizer step at the end