colossalai
54 строки · 1.8 Кб
1from copy import deepcopy
2
3import numpy as np
4import pytest
5import torch
6
7from colossalai.testing import DummyDataloader, clear_cache_before_run
8from colossalai.zero.gemini.memory_tracer.runtime_mem_tracer import RuntimeMemTracer
9from tests.kit.model_zoo import model_zoo, run_fwd_bwd
10
11
12@pytest.mark.skip("this is not used")
13@clear_cache_before_run()
14def test_runtime_mem_tracer():
15test_models = ["gpt2", "bert", "simple_net", "repeated_computed_layers", "nested_model", "albert"]
16
17for model_name in test_models:
18model_builder, data_gen_fn, output_transform_fn, *_ = next(
19iter(model_zoo.get_sub_registry(model_name).values())
20)
21
22model = model_builder().cuda()
23
24model_bk = deepcopy(model)
25runtime_mem_tracer = RuntimeMemTracer(model)
26
27train_dataloader = DummyDataloader(data_gen_fn)
28for i, data in enumerate(train_dataloader):
29if i > 1:
30break
31data = {k: v.cuda() if isinstance(v, torch.Tensor) else v for k, v in data.items()}
32
33run_fwd_bwd(runtime_mem_tracer, data, output_transform_fn, optimizer=runtime_mem_tracer)
34
35for p1, p2 in zip(model_bk.parameters(), model.parameters()):
36torch.allclose(p1.to(torch.half), p2)
37
38non_model_data_list = runtime_mem_tracer._memstats.non_model_data_list("cuda")
39cuda_non_model_data_list = np.array(non_model_data_list) / 1024**2
40print("cuda_non_model_data_list", len(cuda_non_model_data_list))
41print(non_model_data_list)
42
43cnt1 = 0
44for p in runtime_mem_tracer.parameters_in_runtime_order():
45cnt1 += 1
46cnt2 = 0
47for p in model.parameters():
48cnt2 += 1
49assert cnt2 == cnt1, f"visited param number {cnt1} vs real param number {cnt2}"
50del model
51
52
53if __name__ == "__main__":
54test_runtime_mem_tracer()
55