lightning/tests/tests_pytorch/utilities/test_all_gather_grad.py

134 lines
4.9 KiB
Python

# Copyright The Lightning AI team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
import pytest
import torch
from lightning.pytorch import Trainer
from lightning.pytorch.demos.boring_classes import BoringModel
from tests_pytorch.core.test_results import spawn_launch
from tests_pytorch.helpers.runif import RunIf
def all_gather_ddp_spawn_fn(strategy):
rank = strategy.local_rank
world_size = strategy.num_processes
tensor1 = torch.ones(8, requires_grad=True)
tensor2 = torch.ones((8, 16, 32), requires_grad=True)
tensor1_gathered = strategy.all_gather(tensor1, sync_grads=True)
tensor2_gathered = strategy.all_gather(tensor2, sync_grads=True)
tensor1_gathered = tensor1_gathered * rank
tensor2_gathered = tensor2_gathered * rank
tensor1_gathered.sum().backward()
tensor2_gathered.sum().backward()
grad1 = torch.zeros_like(tensor1.grad).fill_(torch.arange(world_size).sum().float())
grad2 = torch.zeros_like(tensor2.grad).fill_(torch.arange(world_size).sum().float())
assert torch.allclose(grad1, tensor1.grad)
assert torch.allclose(grad2, tensor2.grad)
@RunIf(skip_windows=True)
@pytest.mark.flaky(reruns=3)
def test_all_gather_ddp_spawn():
spawn_launch(all_gather_ddp_spawn_fn, [torch.device("cpu")] * 3)
@RunIf(min_cuda_gpus=2, skip_windows=True, standalone=True)
def test_all_gather_collection(tmp_path):
class TestModel(BoringModel):
on_train_epoch_end_called = False
def on_train_epoch_end(self):
losses = torch.rand(2, 2).t()
gathered_loss = self.all_gather({
"losses_tensor_int": losses.int(),
"losses_tensor_float": losses,
"losses_tensor_list": [losses, losses],
"losses_np_ndarray": np.array([1, 2, 3]),
"losses_bool": [True, False],
"losses_float": [0.0, 1.0, 2.0],
"losses_int": [0, 1, 2],
})
assert gathered_loss["losses_tensor_int"][0].dtype == torch.int32
assert gathered_loss["losses_tensor_float"][0].dtype == torch.float
assert gathered_loss["losses_np_ndarray"][0].dtype == torch.int64
# torch.bool can't be all_gathered
assert gathered_loss["losses_bool"][0].dtype == torch.uint8
assert gathered_loss["losses_float"][0].dtype == torch.float
assert gathered_loss["losses_int"][0].dtype == torch.int
losses_numel = losses.numel()
assert gathered_loss["losses_tensor_int"].numel() == 2 * losses_numel
assert gathered_loss["losses_tensor_float"].numel() == 2 * losses_numel
assert torch.stack(gathered_loss["losses_tensor_list"]).shape == (2, 2, 2, 2)
assert gathered_loss["losses_np_ndarray"].numel() == 2 * 3
assert torch.stack(gathered_loss["losses_bool"]).shape == (2, 2)
assert torch.stack(gathered_loss["losses_float"]).shape == (3, 2)
assert torch.stack(gathered_loss["losses_int"]).shape == (3, 2)
self.on_train_epoch_end_called = True
model = TestModel()
trainer = Trainer(
default_root_dir=tmp_path,
limit_train_batches=8,
limit_val_batches=0,
max_epochs=1,
log_every_n_steps=1,
accumulate_grad_batches=2,
accelerator="gpu",
devices=2,
strategy="ddp",
enable_progress_bar=False,
enable_model_summary=False,
enable_checkpointing=False,
)
trainer.fit(model)
assert model.on_train_epoch_end_called
@RunIf(min_cuda_gpus=2, skip_windows=True, standalone=True)
def test_all_gather_sync_grads(tmp_path):
class TestModel(BoringModel):
training_step_called = False
def training_step(self, batch, batch_idx):
self.training_step_called = True
tensor = torch.rand(2, 2, requires_grad=True, device=self.device)
gathered_tensor = self.all_gather(tensor, sync_grads=True)
assert gathered_tensor.shape == torch.Size([2, 2, 2])
return gathered_tensor.sum()
model = TestModel()
trainer = Trainer(
default_root_dir=tmp_path,
limit_train_batches=1,
limit_val_batches=0,
max_epochs=1,
accelerator="gpu",
devices=2,
strategy="ddp",
enable_progress_bar=False,
enable_model_summary=False,
enable_checkpointing=False,
)
trainer.fit(model)
assert model.training_step_called