2022-12-07 15:04:02 +00:00
|
|
|
import pytest
|
|
|
|
import torch
|
2023-08-24 01:29:25 +00:00
|
|
|
import torch.distributed as dist
|
2022-12-07 15:04:02 +00:00
|
|
|
|
|
|
|
import colossalai
|
2023-10-20 02:35:08 +00:00
|
|
|
from colossalai.testing import DummyDataloader, parameterize, rerun_if_address_is_in_use, spawn
|
2023-09-18 08:31:06 +00:00
|
|
|
from colossalai.utils import set_seed
|
2023-08-24 01:29:25 +00:00
|
|
|
from colossalai.zero import GeminiDDP
|
|
|
|
from colossalai.zero.gemini.chunk import search_chunk_configuration
|
2023-04-04 05:48:16 +00:00
|
|
|
from colossalai.zero.gemini.memory_tracer.runtime_mem_tracer import RuntimeMemTracer
|
2023-10-20 02:35:08 +00:00
|
|
|
from tests.kit.model_zoo import model_zoo, run_fwd_bwd
|
2022-12-07 15:04:02 +00:00
|
|
|
|
|
|
|
# run gemini use the runtime memory tracer
|
|
|
|
|
|
|
|
|
2023-09-19 06:20:26 +00:00
|
|
|
@parameterize("placement_policy", ["auto"])
|
|
|
|
@parameterize("keep_gather", [False])
|
2023-10-20 02:35:08 +00:00
|
|
|
@parameterize("model_name", ["transformers_bert_for_sequence_classification"])
|
2023-09-19 06:20:26 +00:00
|
|
|
@parameterize("use_grad_checkpoint", [False, True])
|
2022-12-07 15:04:02 +00:00
|
|
|
def run_gemini_use_rmt(placement_policy, keep_gather, model_name: str, use_grad_checkpoint: bool = False):
|
|
|
|
set_seed(42)
|
2023-10-20 02:35:08 +00:00
|
|
|
model_builder, data_gen_fn, output_transform_fn, *_ = next(iter(model_zoo.get_sub_registry(model_name).values()))
|
2022-12-07 15:04:02 +00:00
|
|
|
|
2023-10-20 02:35:08 +00:00
|
|
|
model = model_builder().cuda()
|
|
|
|
if use_grad_checkpoint:
|
|
|
|
model.gradient_checkpointing_enable()
|
2022-12-07 15:04:02 +00:00
|
|
|
|
2023-09-19 06:20:26 +00:00
|
|
|
print(f"model_name {model_name}")
|
2022-12-07 15:04:02 +00:00
|
|
|
|
2023-10-20 02:35:08 +00:00
|
|
|
runtime_mem_tracer = RuntimeMemTracer(model)
|
|
|
|
data = data_gen_fn()
|
|
|
|
data = {k: v.cuda() if isinstance(v, torch.Tensor) else v for k, v in data.items()}
|
|
|
|
run_fwd_bwd(runtime_mem_tracer, data, output_transform_fn, optimizer=runtime_mem_tracer)
|
2022-12-07 15:04:02 +00:00
|
|
|
memstats = runtime_mem_tracer.memstats()
|
|
|
|
runtime_tracer_non_model_data = runtime_mem_tracer._memstats._non_model_data_cuda_list
|
2023-09-19 06:20:26 +00:00
|
|
|
print("runtime tracer non model data points: ", len(runtime_tracer_non_model_data))
|
|
|
|
print("runtime tracer: ", runtime_tracer_non_model_data)
|
2022-12-13 09:11:31 +00:00
|
|
|
print([memstats.param_used_step(p) for p in model.parameters()])
|
|
|
|
|
2023-09-19 06:20:26 +00:00
|
|
|
if model_name == "repeated_computed_layers":
|
2022-12-13 09:11:31 +00:00
|
|
|
for idx, p in enumerate(model.parameters()):
|
|
|
|
step_list = memstats.param_used_step(p)
|
|
|
|
if idx < 4:
|
|
|
|
assert len(step_list) == 4
|
2022-12-07 15:04:02 +00:00
|
|
|
|
2023-09-19 06:20:26 +00:00
|
|
|
if model_name == "repeated_computed_layers":
|
2022-12-13 08:34:10 +00:00
|
|
|
for idx, p in enumerate(model.parameters()):
|
2022-12-13 16:47:06 +00:00
|
|
|
step_list = memstats.param_used_step(p)
|
2022-12-13 08:34:10 +00:00
|
|
|
if idx < 4:
|
|
|
|
assert len(step_list) == 4
|
|
|
|
|
2022-12-13 06:14:55 +00:00
|
|
|
world_size = torch.distributed.get_world_size()
|
2023-06-25 05:34:15 +00:00
|
|
|
config_dict, *_ = search_chunk_configuration(model, search_range_m=1, search_interval=100)
|
2023-09-19 06:20:26 +00:00
|
|
|
config_dict[world_size]["chunk_size"] = 5000
|
|
|
|
config_dict[world_size]["keep_gathered"] = keep_gather
|
|
|
|
model = GeminiDDP(
|
|
|
|
model, chunk_config_dict=config_dict, placement_policy=placement_policy, pin_memory=True, memstats=memstats
|
|
|
|
)
|
2022-12-07 15:04:02 +00:00
|
|
|
|
2023-08-24 01:29:25 +00:00
|
|
|
set_seed(dist.get_rank())
|
2023-10-20 02:35:08 +00:00
|
|
|
train_dataloader = DummyDataloader(data_gen_fn)
|
|
|
|
for i, data in enumerate(train_dataloader):
|
2022-12-07 15:04:02 +00:00
|
|
|
# you can only test a single fwd + bwd.
|
|
|
|
# after bwd param is grad for Gemini, due to the chunk reuse optimization.
|
2022-12-12 10:06:16 +00:00
|
|
|
# print(f'iteration {i}')
|
|
|
|
if i > 4:
|
2022-12-07 15:04:02 +00:00
|
|
|
break
|
2023-10-20 02:35:08 +00:00
|
|
|
data = {k: v.cuda() if isinstance(v, torch.Tensor) else v for k, v in data.items()}
|
2022-12-07 15:04:02 +00:00
|
|
|
|
|
|
|
set_seed(42)
|
2023-10-20 02:35:08 +00:00
|
|
|
run_fwd_bwd(model, data, output_transform_fn, optimizer=model)
|
2022-12-07 15:04:02 +00:00
|
|
|
|
2023-09-19 06:20:26 +00:00
|
|
|
gemini_non_model_data = model.gemini_manager._mem_stats_collector._memstats.non_model_data_list("cuda")
|
2022-12-07 15:04:02 +00:00
|
|
|
|
|
|
|
# print('gemini non model data:', gemini_non_model_data)
|
|
|
|
|
2023-09-19 06:20:26 +00:00
|
|
|
assert len(gemini_non_model_data) == len(
|
|
|
|
runtime_tracer_non_model_data
|
|
|
|
), f"model_name {model_name} {len(gemini_non_model_data)} vs {len(runtime_tracer_non_model_data)}"
|
2022-12-07 15:04:02 +00:00
|
|
|
|
|
|
|
|
|
|
|
def run_dist(rank, world_size, port):
|
|
|
|
config = {}
|
2023-09-19 06:20:26 +00:00
|
|
|
colossalai.launch(config=config, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl")
|
2022-12-07 15:04:02 +00:00
|
|
|
run_gemini_use_rmt()
|
|
|
|
|
|
|
|
|
2023-08-24 01:29:25 +00:00
|
|
|
@pytest.mark.skip("this is not used")
|
2022-12-07 15:04:02 +00:00
|
|
|
@pytest.mark.dist
|
2023-09-19 06:20:26 +00:00
|
|
|
@pytest.mark.parametrize("world_size", [1, 4])
|
2022-12-07 15:04:02 +00:00
|
|
|
@rerun_if_address_is_in_use()
|
|
|
|
def test_gemini_use_rmt(world_size):
|
2023-04-06 06:51:35 +00:00
|
|
|
spawn(run_dist, world_size)
|
2022-12-07 15:04:02 +00:00
|
|
|
|
|
|
|
|
2023-09-19 06:20:26 +00:00
|
|
|
if __name__ == "__main__":
|
2022-12-07 15:04:02 +00:00
|
|
|
test_gemini_use_rmt(1)
|