def test_tpu(self):
distributed_args = f"""
{self.test_dir}/xla_spawn.py
--num_cores 8
{self.test_file_path}
""".split()
cmd = [sys.executable] + distributed_args
execute_subprocess_async(cmd, env=os.environ.copy())
def test_ddp_kwargs(self):
distributed_args = f"""
-m torch.distributed.launch
--nproc_per_node={torch.cuda.device_count()}
--use_env
{inspect.getfile(self.__class__)}
""".split()
cmd = [sys.executable] + distributed_args
execute_subprocess_async(cmd, env=os.environ.copy())
def test_multi_gpu(self):
print(f"Found {torch.cuda.device_count()} devices.")
distributed_args = f"""
-m torch.distributed.launch
--nproc_per_node={torch.cuda.device_count()}
--use_env
{self.test_file_path}
""".split()
cmd = [sys.executable] + distributed_args
execute_subprocess_async(cmd, env=os.environ.copy())
def test_command(args):
script_name = os.path.sep.join(
__file__.split(os.path.sep)[:-2] + ["test_utils", "test_script.py"])
test_args = f"""
--config_file={args.config_file} {script_name}
""".split()
cmd = ["accelerate-launch"] + test_args
result = execute_subprocess_async(cmd, env=os.environ.copy())
if result.returncode == 0:
print(
"Test is a success! You are ready for your distributed training!")