1313# limitations under the License.
1414import os
1515from unittest import mock
16- from unittest .mock import patch
1716
1817import pytest
1918import torch
@@ -59,34 +58,29 @@ def test_multi_gpu_model_ddp_fit_test(tmpdir):
5958
6059
6160@RunIf (skip_windows = True )
62- @pytest . mark . skipif ( torch . cuda . is_available (), reason = "test doesn't requires GPU machine" )
61+ @mock . patch . dict ( os . environ , { "CUDA_VISIBLE_DEVICES" : "0,1" , "WORLD_SIZE" : "2" }, clear = True )
6362@mock .patch ("lightning_lite.utilities.device_parser.is_cuda_available" , return_value = True )
64- def test_torch_distributed_backend_env_variables (tmpdir ):
63+ @mock .patch ("lightning_lite.utilities.device_parser.num_cuda_devices" , return_value = 2 )
64+ def test_torch_distributed_backend_invalid (_ , __ , tmpdir ):
6565 """This test set `undefined` as torch backend and should raise an `Backend.UNDEFINED` ValueError."""
66- _environ = {"PL_TORCH_DISTRIBUTED_BACKEND" : "undefined" , "CUDA_VISIBLE_DEVICES" : "0,1" , "WORLD_SIZE" : "2" }
67- with patch .dict (os .environ , _environ ), patch (
68- "lightning_lite.utilities.device_parser.num_cuda_devices" , return_value = 2
69- ):
70- with pytest .deprecated_call (match = "Environment variable `PL_TORCH_DISTRIBUTED_BACKEND` was deprecated in v1.6" ):
71- with pytest .raises (ValueError , match = "Invalid backend: 'undefined'" ):
72- model = BoringModel ()
73- trainer = Trainer (
74- default_root_dir = tmpdir ,
75- fast_dev_run = True ,
76- strategy = "ddp" ,
77- accelerator = "gpu" ,
78- devices = 2 ,
79- logger = False ,
80- )
81- trainer .fit (model )
66+ model = BoringModel ()
67+ trainer = Trainer (
68+ default_root_dir = tmpdir ,
69+ fast_dev_run = True ,
70+ strategy = DDPStrategy (process_group_backend = "undefined" ),
71+ accelerator = "cuda" ,
72+ devices = 2 ,
73+ logger = False ,
74+ )
75+ with pytest .raises (ValueError , match = "Invalid backend: 'undefined'" ):
76+ trainer .fit (model )
8277
8378
8479@RunIf (skip_windows = True )
8580@mock .patch ("torch.cuda.set_device" )
8681@mock .patch ("lightning_lite.utilities.device_parser.is_cuda_available" , return_value = True )
8782@mock .patch ("lightning_lite.utilities.device_parser.num_cuda_devices" , return_value = 1 )
8883@mock .patch ("pytorch_lightning.accelerators.gpu.CUDAAccelerator.is_available" , return_value = True )
89- @mock .patch .dict (os .environ , {"PL_TORCH_DISTRIBUTED_BACKEND" : "gloo" }, clear = True )
9084def test_ddp_torch_dist_is_available_in_setup (
9185 mock_gpu_is_available , mock_device_count , mock_cuda_available , mock_set_device , tmpdir
9286):
@@ -98,10 +92,15 @@ def setup(self, stage: str) -> None:
9892 raise SystemExit ()
9993
10094 model = TestModel ()
101- trainer = Trainer (default_root_dir = tmpdir , fast_dev_run = True , strategy = "ddp" , accelerator = "gpu" , devices = 1 )
102- with pytest .deprecated_call (match = "Environment variable `PL_TORCH_DISTRIBUTED_BACKEND` was deprecated in v1.6" ):
103- with pytest .raises (SystemExit ):
104- trainer .fit (model )
95+ trainer = Trainer (
96+ default_root_dir = tmpdir ,
97+ fast_dev_run = True ,
98+ strategy = DDPStrategy (process_group_backend = "gloo" ),
99+ accelerator = "gpu" ,
100+ devices = 1 ,
101+ )
102+ with pytest .raises (SystemExit ):
103+ trainer .fit (model )
105104
106105
107106@RunIf (min_cuda_gpus = 2 , min_torch = "1.8.1" , standalone = True )
@@ -143,17 +142,15 @@ def on_train_start(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule")
143142
144143
145144@pytest .mark .parametrize (
146- ["process_group_backend" , "env_var" , " device_str""expected_process_group_backend" ],
145+ ["process_group_backend" , "device_str" , "expected_process_group_backend" ],
147146 [
148- pytest .param ("foo" , None , "cpu" , "foo" ),
149- pytest .param ("foo" , "BAR" , "cpu" , "foo" ),
150- pytest .param ("foo" , "BAR" , "cuda:0" , "foo" ),
151- pytest .param (None , "BAR" , "cuda:0" , "BAR" ),
152- pytest .param (None , None , "cuda:0" , "nccl" ),
153- pytest .param (None , None , "cpu" , "gloo" ),
147+ pytest .param ("foo" , "cpu" , "foo" ),
148+ pytest .param ("foo" , "cuda:0" , "foo" ),
149+ pytest .param (None , "cuda:0" , "nccl" ),
150+ pytest .param (None , "cpu" , "gloo" ),
154151 ],
155152)
156- def test_ddp_process_group_backend (process_group_backend , env_var , device_str , expected_process_group_backend ):
153+ def test_ddp_process_group_backend (process_group_backend , device_str , expected_process_group_backend ):
157154 """Test settings for process group backend."""
158155
159156 class MockDDPStrategy (DDPStrategy ):
@@ -166,14 +163,7 @@ def root_device(self):
166163 return self ._root_device
167164
168165 strategy = MockDDPStrategy (process_group_backend = process_group_backend , root_device = torch .device (device_str ))
169- if not process_group_backend and env_var :
170- with mock .patch .dict (os .environ , {"PL_TORCH_DISTRIBUTED_BACKEND" : env_var }):
171- with pytest .deprecated_call (
172- match = "Environment variable `PL_TORCH_DISTRIBUTED_BACKEND` was deprecated in v1.6"
173- ):
174- assert strategy ._get_process_group_backend () == expected_process_group_backend
175- else :
176- assert strategy ._get_process_group_backend () == expected_process_group_backend
166+ assert strategy ._get_process_group_backend () == expected_process_group_backend
177167
178168
179169@pytest .mark .parametrize (
0 commit comments