credit.domain_parallel.halo_exchange

credit.domain_parallel.halo_exchange#

Halo exchange for domain-parallel operations.

Implements differentiable halo exchange along the sharding dimension using point-to-point communication (isend/irecv). The backward pass performs the reverse exchange so gradients flow correctly across domain boundaries.

For latitude sharding of weather data: - The “previous” neighbor is the rank holding the region just north. - The “next” neighbor is the rank holding the region just south. - Edge ranks (poles) get zero-padded halos on their outer boundary,

since TensorPadding already handled pole reflection before sharding.

Classes#

_HaloExchangeFunction

Differentiable halo exchange.

HaloExchange

Halo exchange layer for domain-parallel operations.

Module Contents#

class credit.domain_parallel.halo_exchange._HaloExchangeFunction(*args, **kwargs)#

Bases: torch.autograd.Function

Differentiable halo exchange.

Forward: pads the tensor with halo rows received from neighbors. Backward: sends gradient halos back to the ranks that contributed them.

static forward(ctx, x, halo_width, dim, manager)#

Define the forward of the custom autograd Function.

This function is to be overridden by all subclasses. There are two ways to define forward:

Usage 1 (Combined forward and ctx):

@staticmethod
def forward(ctx: Any, *args: Any, **kwargs: Any) -> Any:
    pass

  • It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).

  • See combining-forward-context for more details

Usage 2 (Separate forward and ctx):

@staticmethod
def forward(*args: Any, **kwargs: Any) -> Any:
    pass


@staticmethod
def setup_context(ctx: Any, inputs: Tuple[Any, ...], output: Any) -> None:
    pass

  • The forward no longer accepts a ctx argument.

  • Instead, you must also override the torch.autograd.Function.setup_context() staticmethod to handle setting up the ctx object. output is the output of the forward, inputs are a Tuple of inputs to the forward.

  • See extending-autograd for more details

The context can be used to store arbitrary data that can be then retrieved during the backward pass. Tensors should not be stored directly on ctx (though this is not currently enforced for backward compatibility). Instead, tensors should be saved either with ctx.save_for_backward() if they are intended to be used in backward (equivalently, vjp) or ctx.save_for_forward() if they are intended to be used for in jvp.

static backward(ctx, grad_output)#

Define a formula for differentiating the operation with backward mode automatic differentiation.

This function is to be overridden by all subclasses. (Defining this function is equivalent to defining the vjp function.)

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computed w.r.t. the output.

class credit.domain_parallel.halo_exchange.HaloExchange(halo_width, dim=-2)#

Bases: torch.nn.Module

Halo exchange layer for domain-parallel operations.

Pads the input tensor with halo rows from neighboring ranks along the sharding dimension. The operation is differentiable.

Parameters:

  • halo_width – Number of rows to exchange on each side.

  • dim – Tensor dimension to exchange along (default: -2 for lat in BCHW).

halo_width#
dim = -2#
forward(x)#
static trim(x, halo_before, halo_after, dim=-2)#

Trim halo rows from the output after a convolution.

Parameters:

  • x – Tensor with extra halo rows.

  • halo_before – Number of rows to trim from the start.

  • halo_after – Number of rows to trim from the end.

  • dim – Dimension to trim along.