lightning/tests/tests_fabric/strategies/test_deepspeed_integration.py

415 lines
16 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 os
from copy import deepcopy
from unittest import mock
from unittest.mock import ANY
import pytest
import torch
import torch.nn as nn
import torch.nn.functional as F
from lightning.fabric import Fabric
from lightning.fabric.plugins import DeepSpeedPrecision
from lightning.fabric.strategies import DeepSpeedStrategy
from torch.utils.data import DataLoader
from tests_fabric.helpers.datasets import RandomDataset, RandomIterableDataset
from tests_fabric.helpers.runif import RunIf
from tests_fabric.test_fabric import BoringModel
@RunIf(min_cuda_gpus=2, standalone=True, deepspeed=True)
def test_deepspeed_multiple_models():
fabric = Fabric(strategy=DeepSpeedStrategy(stage=3, logging_batch_size_per_gpu=1), devices=2, accelerator="gpu")
fabric.launch()
with fabric.init_module():
model = BoringModel()
optimizer = torch.optim.SGD(model.parameters(), lr=0.0001)
model, optimizer = fabric.setup(model, optimizer)
for i in range(2):
optimizer.zero_grad()
x = model(torch.randn(1, 32).to(fabric.device))
loss = x.sum()
if i == 0:
# the weights are not initialized with stage 3 until backward is run once
assert all(w.nelement() == 0 for w in model.state_dict().values())
fabric.backward(loss, model=model)
if i == 0:
# save for later to check that the weights were updated
state_dict = deepcopy(model.state_dict())
optimizer.step()
# check that the model trained, the weights from step 1 do not match the weights from step 2
for mw_b, mw_a in zip(state_dict.values(), model.state_dict().values()):
assert not torch.allclose(mw_b, mw_a)
fabric.seed_everything(42)
model_1 = BoringModel()
optimizer_1 = torch.optim.SGD(model_1.parameters(), lr=0.0001)
fabric.seed_everything(42)
model_2 = BoringModel()
optimizer_2 = torch.optim.SGD(model_2.parameters(), lr=0.0001)
for mw_1, mw_2 in zip(model_1.state_dict().values(), model_2.state_dict().values()):
assert torch.allclose(mw_1, mw_2)
model_1, optimizer_1 = fabric.setup(model_1, optimizer_1)
model_2, optimizer_2 = fabric.setup(model_2, optimizer_2)
# train model_1 first
fabric.seed_everything(42)
data_list = []
for _ in range(2):
optimizer_1.zero_grad()
data = torch.randn(1, 32).to(fabric.device)
data_list.append(data)
x = model_1(data)
loss = x.sum()
fabric.backward(loss, model=model_1)
optimizer_1.step()
# the weights do not match
assert all(w.nelement() > 1 for w in model_1.state_dict().values())
assert all(w.nelement() == 0 for w in model_2.state_dict().values())
# now train model_2 with the same data
for data in data_list:
optimizer_2.zero_grad()
x = model_2(data)
loss = x.sum()
fabric.backward(loss, model=model_2)
optimizer_2.step()
# the weights should match
for mw_1, mw_2 in zip(model_1.state_dict().values(), model_2.state_dict().values()):
assert torch.allclose(mw_1, mw_2)
# Verify collectives works as expected
ranks = fabric.all_gather(torch.tensor([fabric.local_rank]).to(fabric.device))
assert torch.allclose(ranks.cpu(), torch.tensor([[0], [1]]))
assert fabric.broadcast(True)
assert fabric.is_global_zero == (fabric.local_rank == 0)
@RunIf(min_cuda_gpus=1, deepspeed=True)
@pytest.mark.parametrize(
("dataset_cls", "logging_batch_size_per_gpu", "expected_batch_size"),
[
(RandomDataset, None, 1),
(RandomDataset, 10, 10),
(RandomIterableDataset, None, 1),
(RandomIterableDataset, 10, 10),
],
)
def test_deepspeed_auto_batch_size_config_select(dataset_cls, logging_batch_size_per_gpu, expected_batch_size):
"""Test to ensure that the batch size is correctly set as expected for deepspeed logging purposes."""
fabric = Fabric(
accelerator="cuda",
devices=1,
strategy=DeepSpeedStrategy(logging_batch_size_per_gpu=logging_batch_size_per_gpu, zero_optimization=False),
)
fabric.launch()
assert isinstance(fabric._strategy, DeepSpeedStrategy)
_ = fabric.setup_dataloaders(DataLoader(dataset_cls(32, 64)))
config = fabric._strategy.config
assert config["train_micro_batch_size_per_gpu"] == expected_batch_size
@RunIf(min_cuda_gpus=1, standalone=True, deepspeed=True)
def test_deepspeed_configure_optimizers():
"""Test that the deepspeed strategy with default initialization wraps the optimizer correctly."""
from deepspeed.runtime.zero.stage_1_and_2 import DeepSpeedZeroOptimizer
fabric = Fabric(
strategy=DeepSpeedStrategy(),
accelerator="cuda",
devices=1,
precision="16-mixed",
)
fabric.launch()
model = nn.Linear(3, 3)
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
model, optimizer = fabric.setup(model, optimizer)
assert isinstance(optimizer.optimizer, DeepSpeedZeroOptimizer)
assert isinstance(optimizer.optimizer.optimizer, torch.optim.SGD)
@RunIf(min_cuda_gpus=1, deepspeed=True)
def test_deepspeed_custom_precision_params():
"""Test that if the FP16 parameters are set via the DeepSpeedStrategy, the deepspeed config contains these
changes."""
strategy = DeepSpeedStrategy(
loss_scale=10, initial_scale_power=11, loss_scale_window=12, hysteresis=13, min_loss_scale=14
)
fabric = Fabric(
strategy=strategy,
precision="16-mixed",
accelerator="cuda",
devices=1,
)
fabric.launch()
assert fabric._strategy._config_initialized
assert fabric._strategy.config["fp16"]["loss_scale"] == 10
assert fabric._strategy.config["fp16"]["initial_scale_power"] == 11
assert fabric._strategy.config["fp16"]["loss_scale_window"] == 12
assert fabric._strategy.config["fp16"]["hysteresis"] == 13
assert fabric._strategy.config["fp16"]["min_loss_scale"] == 14
@RunIf(min_cuda_gpus=1, standalone=True, deepspeed=True)
def test_deepspeed_custom_activation_checkpointing_params_forwarded():
"""Test that the activation checkpointing parameters get passed to `deepspeed.checkpointing.configure`
correctly."""
import deepspeed
strategy = DeepSpeedStrategy(
partition_activations=True,
cpu_checkpointing=True,
contiguous_memory_optimization=True,
synchronize_checkpoint_boundary=True,
)
fabric = Fabric(
strategy=strategy,
precision="16-mixed",
accelerator="cuda",
devices=1,
)
fabric.launch()
model = nn.Linear(3, 3)
optimizer = torch.optim.Adam(model.parameters())
with mock.patch("deepspeed.checkpointing.configure", wraps=deepspeed.checkpointing.configure) as configure:
fabric.setup(model, optimizer)
configure.assert_called_with(
mpu_=None,
partition_activations=True,
contiguous_checkpointing=True,
checkpoint_in_cpu=True,
profile=None,
)
@RunIf(min_cuda_gpus=2, standalone=True, deepspeed=True)
def test_deepspeed_multigpu_stage_3():
"""Test to ensure ZeRO Stage 3 works with a parallel model."""
fabric = Fabric(
strategy=DeepSpeedStrategy(stage=3),
accelerator="cuda",
devices=2,
precision="16-mixed",
)
fabric.launch()
def _make_block():
return nn.Sequential(nn.Linear(32, 32, bias=False), nn.ReLU())
with fabric.init_module():
model = nn.Sequential(*(_make_block() for _ in range(5)), nn.Linear(32, 3))
optimizer = torch.optim.Adam(model.parameters(), lr=0.1)
model, optimizer = fabric.setup(model, optimizer)
x = torch.rand(2, 32, device=fabric.device)
y = torch.ones(x.size(0), device=x.device, dtype=torch.long)
x = model(x)
x = x.float() # Ensure output is in float32 for softmax operation
logits = F.softmax(x, dim=1)
loss = F.cross_entropy(logits, y)
fabric.backward(loss)
optimizer.step()
optimizer.zero_grad()
@RunIf(deepspeed=True)
@mock.patch("deepspeed.init_distributed", autospec=True)
@mock.patch("lightning.fabric.accelerators.mps.MPSAccelerator.is_available", return_value=False)
@pytest.mark.parametrize("platform", ["Linux", "Windows"])
def test_deepspeed_env_variables_on_platforms(_, deepspeed_dist_mock, platform):
"""Test to ensure that we set up distributed communication correctly.
When using Windows, ranks environment variables should not be set, and DeepSpeed should handle this.
"""
fabric = Fabric(strategy=DeepSpeedStrategy(stage=3))
strategy = fabric._strategy
assert isinstance(strategy, DeepSpeedStrategy)
with mock.patch("platform.system", return_value=platform) as platform_mock:
strategy._init_deepspeed_distributed()
deepspeed_dist_mock.assert_called()
platform_mock.assert_called()
if platform == "Windows":
# assert no env variables have been set within the DeepSpeedStrategy
assert all(k not in os.environ for k in ("MASTER_PORT", "MASTER_ADDR", "RANK", "WORLD_SIZE", "LOCAL_RANK"))
else:
assert os.environ["MASTER_ADDR"] == str(strategy.cluster_environment.main_address)
assert os.environ["MASTER_PORT"] == str(strategy.cluster_environment.main_port)
assert os.environ["RANK"] == str(strategy.global_rank)
assert os.environ["WORLD_SIZE"] == str(strategy.world_size)
assert os.environ["LOCAL_RANK"] == str(strategy.local_rank)
@RunIf(min_cuda_gpus=2, standalone=True, deepspeed=True, bf16_cuda=True)
def test_deepspeed_with_bfloat16_precision():
"""Test that the DeepSpeed strategy works with bfloat16 precision."""
class Model(nn.Module):
def __init__(self):
super().__init__()
self.layer = nn.Linear(32, 2)
def forward(self, x):
assert x.dtype == torch.bfloat16
return self.layer(x)
fabric = Fabric(accelerator="cuda", devices=2, strategy="deepspeed_stage_3", precision="bf16-mixed")
assert isinstance(fabric._strategy.precision, DeepSpeedPrecision)
assert fabric._strategy.precision.precision == "bf16-mixed"
assert fabric._strategy.config["zero_optimization"]["stage"] == 3
fabric.launch()
with fabric.init_module():
model = Model()
optimizer = torch.optim.Adam(model.parameters(), lr=0.1)
model, optimizer = fabric.setup(model, optimizer)
assert fabric._strategy.config["bf16"]["enabled"]
assert model.layer.weight.dtype == torch.bfloat16
batch = torch.rand(2, 32, device=fabric.device)
assert batch.dtype == torch.float32
loss = model(batch).sum()
fabric.backward(loss)
optimizer.step()
optimizer.zero_grad()
def _assert_saved_model_is_equal(fabric, model, checkpoint_path):
"""Convert the saved checkpoint to a single file with the model weights consolidated to easily verify the full
weights in float32 precision."""
from deepspeed.utils.zero_to_fp32 import convert_zero_checkpoint_to_fp32_state_dict
assert isinstance(fabric.strategy, DeepSpeedStrategy)
# carry out the check only on rank 0
if fabric.is_global_zero:
if fabric.strategy.config["zero_optimization"]["stage"] in (2, 3):
single_ckpt_path = checkpoint_path / "single_model.pt"
# the tag is hardcoded in DeepSpeedStrategy
convert_zero_checkpoint_to_fp32_state_dict(checkpoint_path, single_ckpt_path, tag="checkpoint")
state_dict = torch.load(single_ckpt_path, weights_only=False)
else:
# 'checkpoint' is the tag, hardcoded in DeepSpeedStrategy
single_ckpt_path = checkpoint_path / "checkpoint" / "mp_rank_00_model_states.pt"
state_dict = torch.load(single_ckpt_path, weights_only=False)["module"]
model = model.cpu()
# assert model parameters are identical after loading
for orig_param, saved_model_param in zip(model.parameters(), state_dict.values()):
# perform the equality check in the same precision
saved_model_param = saved_model_param.cpu().to(orig_param.dtype)
assert torch.equal(orig_param, saved_model_param)
fabric.barrier()
@RunIf(min_cuda_gpus=2, standalone=True, deepspeed=True, bf16_cuda=True)
@pytest.mark.parametrize("stage", [1, 2, 3])
def test_deepspeed_save_load_checkpoint_zero_3(stage, tmp_path):
"""Test that DeepSpeed stage 1, 2, and 3 model checkpoints can be saved and loaded successfully."""
from deepspeed import DeepSpeedEngine
fabric = Fabric(accelerator="cuda", devices=2, strategy=DeepSpeedStrategy(stage=stage), precision="bf16-mixed")
fabric.launch()
checkpoint_path = fabric.broadcast(tmp_path / "deepspeed-checkpoint")
with fabric.init_module():
model = BoringModel()
optimizer = torch.optim.SGD(model.parameters(), lr=0.0001)
model, optimizer = fabric.setup(model, optimizer)
assert isinstance(model._forward_module, DeepSpeedEngine)
# TODO(fabric): The dtype on the model is not correct, should be torch.bfloat16
assert model.dtype == torch.float32
assert next(model.parameters()).dtype == torch.bfloat16
# dummy training step
output = model(torch.randn(1, 32).to(fabric.device))
loss = output.sum()
fabric.backward(loss)
optimizer.step()
optimizer.zero_grad()
state = {"model": model, "optimizer": optimizer, "steps": 1}
fabric.save(checkpoint_path, state)
# re-init all objects and resume
fabric = Fabric(accelerator="cuda", devices=2, strategy=DeepSpeedStrategy(stage=stage), precision="bf16")
fabric.launch()
with fabric.init_module():
model = BoringModel()
optimizer = torch.optim.SGD(model.parameters(), lr=0.0001)
model, optimizer = fabric.setup(model, optimizer)
state = {"model": model, "optimizer": optimizer, "steps": 0}
metadata = fabric.load(checkpoint_path, state)
# check user data in state reloaded
assert state["steps"] == 1
# the remainder of the deepspeed checkpoint contains metadata
assert "ds_version" in metadata
_assert_saved_model_is_equal(fabric, model, checkpoint_path)
@RunIf(min_cuda_gpus=2, standalone=True, deepspeed=True, bf16_cuda=True)
@pytest.mark.parametrize("empty_init", [None, True])
def test_deepspeed_init_module_with_stage_3(empty_init):
"""Tests how `.init_module()` behaves with ZeRO stage 3."""
strategy = DeepSpeedStrategy(stage=3)
fabric = Fabric(accelerator="cuda", devices=2, strategy=strategy, precision="bf16-true")
fabric.launch()
with mock.patch("deepspeed.zero.Init") as zero_init_mock, fabric.init_module(empty_init=empty_init):
BoringModel()
fabric.barrier()
zero_init_mock.assert_called_once_with(enabled=True, remote_device=None, config_dict_or_path=ANY)
@RunIf(min_cuda_gpus=2, standalone=True, deepspeed=True, bf16_cuda=True)
@pytest.mark.parametrize("stage", [1, 2])
@pytest.mark.parametrize("empty_init", [None, False, True])
def test_deepspeed_init_module_with_stages_1_2(stage, empty_init):
"""Tests how `.init_module()` behaves with ZeRO stages 1 and 2."""
strategy = DeepSpeedStrategy(stage=stage)
fabric = Fabric(accelerator="cuda", devices=2, strategy=strategy, precision="bf16-true")
fabric.launch()
with mock.patch("deepspeed.zero.Init") as zero_init_mock, mock.patch(
"torch.Tensor.uniform_"
) as init_mock, fabric.init_module(empty_init=empty_init):
model = BoringModel()
zero_init_mock.assert_called_with(enabled=False, remote_device=None, config_dict_or_path=ANY)
assert init_mock.call_count == int(not empty_init)
assert model.layer.weight.dtype == torch.bfloat16