Enable training under spatial model parallelism (ModelTorchDistributed)

fme/ace/stepper/test_single_module_csfno.py

@@ -0,0 +1,245 @@
+"""
+Parallel regression tests for the SingleModuleStepper with NoiseConditionedSFNO.
+
+These tests verify that the forward pass and loss computation produce identical
+results regardless of spatial decomposition (nproc=1 vs model-parallel).
+"""
+
+import dataclasses
+import datetime
+import os
+from collections.abc import Mapping
+
+import numpy as np
+import pytest
+import torch
+import xarray as xr
+
+from fme.ace.data_loading.batch_data import BatchData
+from fme.ace.registry.stochastic_sfno import NoiseConditionedSFNOBuilder
+from fme.ace.stepper.single_module import (
+    StepperConfig,
+    TrainOutput,
+    TrainStepper,
+    TrainStepperConfig,
+)
+from fme.core.coordinates import HybridSigmaPressureCoordinate, LatLonCoordinates
+from fme.core.dataset_info import DatasetInfo
+from fme.core.device import get_device
+from fme.core.distributed.distributed import Distributed
+from fme.core.loss import StepLossConfig
+from fme.core.normalizer import NetworkAndLossNormalizationConfig, NormalizationConfig
+from fme.core.optimization import NullOptimization, OptimizationConfig
+from fme.core.registry.module import ModuleSelector
+from fme.core.step import SingleModuleStepConfig, StepSelector
+from fme.core.testing.regression import validate_tensor_dict
+from fme.core.typing_ import EnsembleTensorDict
+
+DIR = os.path.abspath(os.path.dirname(__file__))
+TIMESTEP = datetime.timedelta(hours=6)
+
+
+def get_dataset_info(
+    img_shape=(5, 5),
+) -> DatasetInfo:
+    horizontal_coordinate = LatLonCoordinates(
+        lat=torch.zeros(img_shape[-2]),
+        lon=torch.zeros(img_shape[-1]),
+    )
+    vertical_coordinate = HybridSigmaPressureCoordinate(
+        ak=torch.arange(7), bk=torch.arange(7)
+    )
+    return DatasetInfo(
+        horizontal_coordinates=horizontal_coordinate,
+        vertical_coordinate=vertical_coordinate,
+        timestep=TIMESTEP,
+    )
+
+
+def _get_train_stepper(
+    stepper_config: StepperConfig,
+    dataset_info: DatasetInfo,
+    **train_config_kwargs,
+) -> TrainStepper:
+    train_config = TrainStepperConfig(**train_config_kwargs)
+    return train_config.get_train_stepper(stepper_config, dataset_info)
+
+
+def get_regression_stepper_and_data() -> (
+    tuple[TrainStepper, BatchData, tuple[int, int]]
+):
+    in_names = ["a", "b"]
+    out_names = ["b", "c"]
+    n_forward_steps = 2
+    n_samples = 3
+    img_shape = (9, 18)
+    device = get_device()
+
+    all_names = list(set(in_names + out_names))
+
+    loss = StepLossConfig(type="AreaWeightedMSE")
+
+    config = StepperConfig(
+        step=StepSelector(
+            type="single_module",
+            config=dataclasses.asdict(
+                SingleModuleStepConfig(
+                    builder=ModuleSelector(
+                        type="NoiseConditionedSFNO",
+                        config=dataclasses.asdict(
+                            NoiseConditionedSFNOBuilder(
+                                embed_dim=16,
+                                num_layers=2,
+                                noise_embed_dim=16,
+                                noise_type="isotropic",
+                            )
+                        ),
+                    ),
+                    in_names=in_names,
+                    out_names=out_names,
+                    normalization=NetworkAndLossNormalizationConfig(
+                        network=NormalizationConfig(
+                            means={n: 0.1 for n in all_names},
+                            stds={n: 1.1 for n in all_names},
+                        ),
+                    ),
+                    ocean=None,
+                )
+            ),
+        ),
+    )
+
+    dataset_info = get_dataset_info(img_shape=img_shape)
+    train_stepper = _get_train_stepper(config, dataset_info, loss=loss)
+    data = BatchData.new_on_device(
+        data={
+            "a": torch.randn(n_samples, n_forward_steps + 1, *img_shape).to(device),
+            "b": torch.randn(n_samples, n_forward_steps + 1, *img_shape).to(device),
+            "c": torch.randn(n_samples, n_forward_steps + 1, *img_shape).to(device),
+        },
+        time=xr.DataArray(
+            np.zeros((n_samples, n_forward_steps + 1)),
+            dims=["sample", "time"],
+        ),
+        labels=None,
+        epoch=0,
+        horizontal_dims=["lat", "lon"],
+    )
+    data = data.scatter_spatial(img_shape)
+    return train_stepper, data, img_shape
+
+
+def flatten_dict(
+    d: Mapping[str, Mapping[str, torch.Tensor]],
+) -> dict[str, torch.Tensor]:
+    return_dict = {}
+    for k, v in d.items():
+        for k2, v2 in v.items():
+            return_dict[f"{k}.{k2}"] = v2
+    return return_dict
+
+
+def _get_train_output_tensor_dict(data: TrainOutput) -> dict[str, torch.Tensor]:
+    return_dict = {}
+    for k, v in data.metrics.items():
+        return_dict[f"metrics.{k}"] = v
+    for k, v in data.gen_data.items():
+        return_dict[f"gen_data.{k}"] = v
+    for k, v in data.target_data.items():
+        assert v.shape[1] == 1
+        return_dict[f"target_data.{k}"] = v
+    return return_dict
+
+
+def get_train_outputs_tensor_dict(
+    step_1: TrainOutput, step_2: TrainOutput
+) -> dict[str, torch.Tensor]:
+    return flatten_dict(
+        {
+            "step_1": _get_train_output_tensor_dict(step_1),
+            "step_2": _get_train_output_tensor_dict(step_2),
+        }
+    )
+
+
+@pytest.mark.parallel
+def test_stepper_train_on_batch_regression():
+    torch.manual_seed(0)
+    train_stepper, data, img_shape = get_regression_stepper_and_data()
+    optimization = NullOptimization()
+    result1 = train_stepper.train_on_batch(data, optimization)
+    result2 = train_stepper.train_on_batch(data, optimization)
+    dist = Distributed.get_instance()
+    for result in [result1, result2]:
+        result.gen_data = EnsembleTensorDict(
+            dist.gather_spatial(dict(result.gen_data), img_shape)
+        )
+        result.target_data = EnsembleTensorDict(
+            dist.gather_spatial(dict(result.target_data), img_shape)
+        )
+    output_dict = get_train_outputs_tensor_dict(result1, result2)
+    validate_tensor_dict(
+        output_dict,
+        os.path.join(
+            DIR,
+            "testdata/csfno_stepper_train_on_batch_regression.pt",
+        ),
+        atol=1e-4,
+        rtol=1e-4,
+    )
+
+
+@pytest.mark.parallel
+def test_stepper_train_on_batch_with_optimization_regression():
+    torch.manual_seed(0)
+    train_stepper, data, img_shape = get_regression_stepper_and_data()
+    optimization = OptimizationConfig(
+        optimizer_type="Adam",
+        lr=0.0001,
+    ).build(train_stepper.modules, max_epochs=1)
+    result1 = train_stepper.train_on_batch(data, optimization)
+    result2 = train_stepper.train_on_batch(data, optimization)
+    dist = Distributed.get_instance()
+    for result in [result1, result2]:
+        result.gen_data = EnsembleTensorDict(
+            dist.gather_spatial(dict(result.gen_data), img_shape)
+        )
+        result.target_data = EnsembleTensorDict(
+            dist.gather_spatial(dict(result.target_data), img_shape)
+        )
+    output_dict = get_train_outputs_tensor_dict(result1, result2)
+    validate_tensor_dict(
+        output_dict,
+        os.path.join(
+            DIR,
+            "testdata/csfno_stepper_train_on_batch_with_optimization_regression.pt",
+        ),
+        atol=1e-2,
+        rtol=1e-2,
+    )
+
+
+@pytest.mark.parallel
+def test_stepper_predict_regression():
+    torch.manual_seed(0)
+    train_stepper, data, img_shape = get_regression_stepper_and_data()
+    stepper = train_stepper._stepper
+    initial_condition = data.get_start(
+        prognostic_names=["b"],
+        n_ic_timesteps=1,
+    )
+    output, next_state = stepper.predict(
+        initial_condition, data, compute_derived_variables=True
+    )
+    dist = Distributed.get_instance()
+    output_data = dist.gather_spatial(dict(output.data), img_shape)
+    next_state_data = dist.gather_spatial(
+        dict(next_state.as_batch_data().data), img_shape
+    )
+    output_dict = flatten_dict({"output": output_data, "next_state": next_state_data})
+    validate_tensor_dict(
+        output_dict,
+        os.path.join(DIR, "testdata/csfno_stepper_predict_regression.pt"),
+        atol=1e-4,
+        rtol=1e-4,
+    )

fme/core/distributed/model_torch_distributed.py

@@ -23,8 +23,10 @@
 
 import torch
 import torch.distributed
+import torch.distributed as pt_dist
 import torch.nn as nn
 import torch_harmonics.distributed as thd
+from torch.amp import custom_bwd, custom_fwd
 from torch.nn import SyncBatchNorm
 from torch.nn.parallel import DistributedDataParallel
 
@@ -42,6 +44,35 @@
 T = TypeVar("T")
 
 
+class _AutogradAllReduce(torch.autograd.Function):
+    """Autograd-aware all-reduce (sum) for spatial parallelism.
+    Forward: all-reduce (sum) the input across the given process group.
+    Backward: identity — gradients pass through without communication.
+    This makes ``spatial_reduce_sum`` differentiable so that gradients
+    flow correctly through the loss computation path::
+        AreaWeightedMSELoss → area_weighted_mean → weighted_mean
+            → spatial_reduce_sum (uses this function)
+    Without this, the raw ``torch.distributed.all_reduce`` would break
+    the autograd graph because it is an in-place, non-differentiable op.
+    """
+
+    @staticmethod
+    @custom_fwd(device_type="cuda")
+    def forward(
+        ctx,
+        input: torch.Tensor,
+        group: torch.distributed.ProcessGroup,
+    ) -> torch.Tensor:
+        output = input.clone()
+        torch.distributed.all_reduce(output, group=group)
+        return output
+
+    @staticmethod
+    @custom_bwd(device_type="cuda")
+    def backward(ctx, grad_output: torch.Tensor):
+        return grad_output.clone(), None
+
+
 class ModelTorchDistributed(DistributedBackend):
     """Distributed backend with spatial model parallelism.
 
@@ -307,31 +338,73 @@ def _device_ids(self) -> list[int] | None:
     def wrap_module(self, module: torch.nn.Module) -> torch.nn.Module:
         """Wrap with DDP over the **data** process group.
 
-        For now, we assume spatial communication is expected to be handled
-        inside the model layers themselves. If we need to change course, we
-        can revisit...
+        Spatial model parallelism is handled by:
+        - Forward: communication inside model layers (distributed SHT/iSHT)
+        - Backward: gradient hooks registered here that all-reduce across
+          spatial ranks, so every rank sees the global-mean gradient.
+
+        ``broadcast_buffers=False`` is required because the SHT/iSHT layers
+        store precomputed Legendre polynomial buffers.  DDP's default
+        buffer broadcast modifies these in-place between forward calls,
+        which breaks autograd's tensor-version tracking.
         """
         if any(p.requires_grad for p in module.parameters()):
             if using_gpu():
                 output_device = [self._device_id]
             else:
                 output_device = None
-            return DistributedDataParallel(
+            wrapped = DistributedDataParallel(
                 SyncBatchNorm.convert_sync_batchnorm(module),
                 device_ids=self._device_ids,
                 output_device=output_device,
                 process_group=self._data_group,
+                broadcast_buffers=False,
             )
+            self._register_spatial_grad_hooks(wrapped)
+            return wrapped
         return DummyWrapper(module)
 
+    def _register_spatial_grad_hooks(self, module: torch.nn.Module) -> None:
+        """All-reduce gradients across spatial ranks after each backward.
+
+        Each spatial rank only sees its local slice of the input, so its
+        gradient is a partial sum.  This hook sums those partials so
+        that every rank applies the same weight update.
+
+        The hook fires via ``register_hook`` on each parameter, which is
+        invoked with the per-backward gradient tensor before it is
+        accumulated into ``.grad`` and before DDP's data-parallel
+        all-reduce. The two reductions commute (orthogonal groups), so
+        ordering does not matter.
+        """
+        if self._h_size <= 1 and self._w_size <= 1:
+            return
+        spatial_group = self._spatial_group
+
+        def _hook(grad: torch.Tensor) -> torch.Tensor:
+            if grad is None:
+                return grad
+
+            reduced = grad.contiguous().clone()
+            torch.distributed.all_reduce(reduced, group=spatial_group)
+
+            # If we want mean gradient instead of sum, we want:
+            # reduced /= (self._h_size * self._w_size)
+
+            return reduced
+
+        for p in module.parameters():
+            if p.requires_grad:
+                p.register_hook(_hook)
+
     def barrier(self):
         """Global barrier across all ranks."""
         logger.debug("Barrier on rank %d", self._rank)
         torch.distributed.barrier(device_ids=self._device_ids)
 
     def spatial_reduce_sum(self, tensor: torch.Tensor) -> torch.Tensor:
         if self._h_size > 1 or self._w_size > 1:
-            torch.distributed.all_reduce(tensor, group=self._spatial_group)
+            return _AutogradAllReduce.apply(tensor, self._spatial_group)
         return tensor
 
     def weighted_mean(
@@ -341,6 +414,7 @@ def weighted_mean(
         dim: tuple[int, ...],
         keepdim: bool = False,
     ) -> torch.Tensor:
+
         from fme.core.metrics import weighted_sum
 
         local_weighted_sum = weighted_sum(data, weights, dim=dim, keepdim=keepdim)