xref: /aosp_15_r20/external/pytorch/test/distributed/fsdp/test_distributed_checkpoint.py (revision da0073e96a02ea20f0ac840b70461e3646d07c45) 
1# Owner(s): ["oncall: distributed"]
2
3import sys
4
5import torch
6from torch import distributed as dist
7from torch.distributed.checkpoint import (
8    FileSystemReader,
9    FileSystemWriter,
10    load_state_dict,
11    save_state_dict,
12)
13from torch.distributed.fsdp import FullyShardedDataParallel as FSDP, StateDictType
14from torch.distributed.fsdp.fully_sharded_data_parallel import FullyShardedDataParallel
15from torch.distributed.fsdp.wrap import enable_wrap, wrap
16from torch.testing._internal.common_distributed import skip_if_lt_x_gpu
17from torch.testing._internal.common_fsdp import FSDPTest, SkipModel
18from torch.testing._internal.common_utils import (
19    instantiate_parametrized_tests,
20    parametrize,
21    run_tests,
22    TEST_WITH_DEV_DBG_ASAN,
23)
24from torch.testing._internal.distributed.checkpoint_utils import with_temp_dir
25
26
27if not dist.is_available():
28    print("Distributed not available, skipping tests", file=sys.stderr)
29    sys.exit(0)
30
31if TEST_WITH_DEV_DBG_ASAN:
32    print(
33        "Skip dev-asan as torch + multiprocessing spawn have known issues",
34        file=sys.stderr,
35    )
36    sys.exit(0)
37
38
39_DISTRIBUTED_STATE_DICT_IMPLS = {
40    StateDictType.LOCAL_STATE_DICT,
41    StateDictType.SHARDED_STATE_DICT,
42}
43
44
45class TestDistributedCheckpoint(FSDPTest):
46    @property
47    def world_size(self):
48        return 2
49
50    @skip_if_lt_x_gpu(2)
51    @with_temp_dir
52    @parametrize("state_dict_type", _DISTRIBUTED_STATE_DICT_IMPLS)
53    def test_distributed_checkpoint(self, state_dict_type) -> None:
54        with enable_wrap(wrapper_cls=FSDP):
55            torch.manual_seed(100)
56            model = wrap(SkipModel(double_nest=True))
57            torch.manual_seed(200)
58            new_model = wrap(SkipModel(double_nest=True))
59
60        with FullyShardedDataParallel.summon_full_params(
61            model
62        ), FullyShardedDataParallel.summon_full_params(new_model):
63            params = list(model.parameters())
64            new_params = list(new_model.parameters())
65            self.assertNotEqual(params, new_params)
66
67        writer = FileSystemWriter(self.temp_dir)
68        reader = FileSystemReader(self.temp_dir)
69        with FSDP.state_dict_type(model, state_dict_type), FSDP.state_dict_type(
70            new_model, state_dict_type
71        ):
72            state_dict = model.state_dict()
73
74        save_state_dict(state_dict, writer)
75
76        with FSDP.state_dict_type(model, state_dict_type), FSDP.state_dict_type(
77            new_model, state_dict_type
78        ):
79            state_dict = new_model.state_dict()
80            load_state_dict(state_dict, reader)
81            new_model.load_state_dict(state_dict)
82
83        with FullyShardedDataParallel.summon_full_params(
84            model
85        ), FullyShardedDataParallel.summon_full_params(new_model):
86            params = list(model.parameters())
87            new_params = list(new_model.parameters())
88            self.assertEqual(params, new_params)
89
90        # TODO: add resharding test case.
91
92
93instantiate_parametrized_tests(TestDistributedCheckpoint)
94
95if __name__ == "__main__":
96    run_tests()
97