colossalai
63 строки · 1.9 Кб
1import pytest
2import torch
3
4import colossalai
5from colossalai.accelerator import get_accelerator
6from colossalai.testing import rerun_if_address_is_in_use, spawn
7from colossalai.zero.gemini.chunk import init_chunk_manager, search_chunk_configuration
8from tests.kit.model_zoo import model_zoo
9
10
11def exam_search_chunk_size():
12model_builder, data_gen_fn, output_transform_fn, *_ = next(
13iter(model_zoo.get_sub_registry("transformers_gpt_lm").values())
14)
15
16# make sure torch_model and model has the same parameter values
17model = model_builder()
18config_dict, *_ = search_chunk_configuration(
19model, search_range_m=1, search_interval=128, min_chunk_size_m=0, filter_exlarge_params=True
20)
21
22for key in config_dict:
23chunk_size = config_dict[key]["chunk_size"]
24assert chunk_size == 527872
25
26
27def exam_chunk_manager():
28world_size = torch.distributed.get_world_size()
29
30model_builder, data_gen_fn, output_transform_fn, *_ = next(
31iter(model_zoo.get_sub_registry("transformers_gpt_lm").values())
32)
33
34sharded_ddp_model = model_builder()
35chunk_manager = init_chunk_manager(
36sharded_ddp_model,
37get_accelerator().get_current_device(),
38hidden_dim=128,
39search_range_m=1,
40min_chunk_size_m=0,
41filter_exlarge_params=True,
42strict_ddp_flag=True,
43)
44config_dict = chunk_manager.dp_degree_chunk_size_dict
45assert len(config_dict) == 1
46assert config_dict[world_size] == 527872
47
48
49def run_dist(rank, world_size, port):
50colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl")
51exam_search_chunk_size()
52exam_chunk_manager()
53
54
55@pytest.mark.dist
56@pytest.mark.parametrize("world_size", [1, 4])
57@rerun_if_address_is_in_use()
58def test_search(world_size):
59spawn(run_dist, world_size)
60
61
62if __name__ == "__main__":
63test_search(4)
64