SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

[deepspeed] Enable multiple test runs on single box, defer to DS_TEST_PORT if set (#14331)

Browse Source 
* defer to DS_TEST_PORT if set

* style

Co-authored-by: Stas Bekman <stas@stason.org>
This commit is contained in:
Jeff Rasley
2021-11-08 12:40:29 -08:00
committed by GitHub
parent dfb00bf644
commit d0e96c6de6
2 changed files with 14 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
tests/deepspeed/test_deepspeed.py
View File

@@ -51,6 +51,9 @@ with ExtendSysPath(tests_dir):
set_seed(42) set_seed(42)
# default torch.distributed port
DEFAULT_MASTER_PORT = "10999"
T5_SMALL = "t5-small" T5_SMALL = "t5-small"
T5_TINY = "patrickvonplaten/t5-tiny-random" T5_TINY = "patrickvonplaten/t5-tiny-random"
GPT2_TINY = "sshleifer/tiny-gpt2" GPT2_TINY = "sshleifer/tiny-gpt2"
@@ -89,7 +92,8 @@ def get_launcher(distributed=False):
# 2. for now testing with just 2 gpus max (since some quality tests may give different # 2. for now testing with just 2 gpus max (since some quality tests may give different
# results with mode gpus because we use very little data) # results with mode gpus because we use very little data)
num_gpus = min(2, get_gpu_count()) if distributed else 1 num_gpus = min(2, get_gpu_count()) if distributed else 1
return f"deepspeed --num_nodes 1 --num_gpus {num_gpus}".split() master_port = os.environ.get("DS_TEST_PORT", DEFAULT_MASTER_PORT)
return f"deepspeed --num_nodes 1 --num_gpus {num_gpus} --master_port {master_port}".split()
ZERO2 = "zero2" ZERO2 = "zero2"
@@ -107,8 +111,9 @@ class CoreIntegrationDeepSpeed(TestCasePlus, TrainerIntegrationCommon):
def setUp(self): def setUp(self):
super().setUp() super().setUp()
master_port = os.environ.get("DS_TEST_PORT", DEFAULT_MASTER_PORT)
self.dist_env_1_gpu = dict( self.dist_env_1_gpu = dict(
MASTER_ADDR="localhost", MASTER_PORT="10999", RANK="0", LOCAL_RANK="0", WORLD_SIZE="1" MASTER_ADDR="localhost", MASTER_PORT=master_port, RANK="0", LOCAL_RANK="0", WORLD_SIZE="1"
) )
def test_init_zero3(self): def test_init_zero3(self):
@@ -176,8 +181,9 @@ class TrainerIntegrationDeepSpeed(TestCasePlus, TrainerIntegrationCommon):
self.n_epochs = args.num_train_epochs self.n_epochs = args.num_train_epochs
self.batch_size = args.train_batch_size self.batch_size = args.train_batch_size
master_port = os.environ.get("DS_TEST_PORT", DEFAULT_MASTER_PORT)
self.dist_env_1_gpu = dict( self.dist_env_1_gpu = dict(
MASTER_ADDR="localhost", MASTER_PORT="10999", RANK="0", LOCAL_RANK="0", WORLD_SIZE="1" MASTER_ADDR="localhost", MASTER_PORT=master_port, RANK="0", LOCAL_RANK="0", WORLD_SIZE="1"
) )
self.ds_config_file = dict( self.ds_config_file = dict(

6
tests/deepspeed/test_model_zoo.py
View File

@@ -41,6 +41,9 @@ with ExtendSysPath(tests_dir):
set_seed(42) set_seed(42)
# default torch.distributed port
DEFAULT_MASTER_PORT = "10999"
# translation # translation
FSMT_TINY = "stas/tiny-wmt19-en-de" FSMT_TINY = "stas/tiny-wmt19-en-de"
BART_TINY = "sshleifer/bart-tiny-random" BART_TINY = "sshleifer/bart-tiny-random"
@@ -89,7 +92,8 @@ def get_launcher(distributed=False):
# 2. for now testing with just 2 gpus max (since some quality tests may give different # 2. for now testing with just 2 gpus max (since some quality tests may give different
# results with mode gpus because we use very little data) # results with mode gpus because we use very little data)
num_gpus = min(2, get_gpu_count()) if distributed else 1 num_gpus = min(2, get_gpu_count()) if distributed else 1
return f"deepspeed --num_nodes 1 --num_gpus {num_gpus}".split() master_port = os.environ.get("DS_TEST_PORT", DEFAULT_MASTER_PORT)
return f"deepspeed --num_nodes 1 --num_gpus {num_gpus} --master_port {master_port}".split()
def make_task_cmds(): def make_task_cmds():