github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-11 13:59:08 +00:00
Code Issues Packages Projects Releases Wiki Activity

[test] merge old components to test to model zoo (#4945)

Browse Source 
* [test] add custom models in model zoo

* [test] update legacy test

* [test] update model zoo

* [test] update gemini test

* [test] remove components to test
This commit is contained in:
Hongxin Liu
2023-10-20 10:35:08 +08:00
committed by GitHub
parent 3a41e8304e
commit b8e770c832
49 changed files with 461 additions and 914 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
tests/test_zero/test_gemini/test_grad_clip.py
View File

@@ -7,12 +7,11 @@ from torch.testing import assert_close
import colossalai
from colossalai.legacy.amp import convert_to_apex_amp
from colossalai.nn.optimizer import HybridAdam
from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
from colossalai.testing import DummyDataloader, parameterize, rerun_if_address_is_in_use, spawn
from colossalai.utils import set_seed
from colossalai.zero import GeminiDDP, GeminiOptimizer
from colossalai.zero.gemini.chunk import search_chunk_configuration
from tests.components_to_test import run_fwd_bwd
from tests.components_to_test.registry import non_distributed_component_funcs
from tests.kit.model_zoo import model_zoo, run_fwd_bwd
PLACEMENT_CONFIGS = [
{
@@ -51,12 +50,13 @@ def check_param(model: GeminiDDP, torch_model: torch.nn.Module):
@parameterize("placement_config", PLACEMENT_CONFIGS)
@parameterize("model_name", ["gpt2"])
@parameterize("model_name", ["transformers_gpt_lm"])
@parameterize("master_weights", [True, False])
def exam_grad_clipping(placement_config, model_name: str, master_weights: bool):
set_seed(1912)
get_components_func = non_distributed_component_funcs.get_callable(model_name)
model_builder, train_dataloader, test_dataloader, optimizer_class, criterion = get_components_func()
model_builder, data_gen_fn, output_transform_fn, loss_fn, *_ = next(
iter(model_zoo.get_sub_registry(model_name).values())
)
torch_model = model_builder().cuda()
amp_config = dict(opt_level="O2", keep_batchnorm_fp32=False, loss_scale=32)
@@ -94,21 +94,17 @@ def exam_grad_clipping(placement_config, model_name: str, master_weights: bool):
torch_model.train()
set_seed(dist.get_rank() * 3 + 128)
for i, (data, label) in enumerate(train_dataloader):
train_dataloader = DummyDataloader(data_gen_fn)
for i, data in enumerate(train_dataloader):
if i > 2:
break
data = data.cuda()
label = label.cuda()
data = {k: v.cuda() if isinstance(v, torch.Tensor) else v for k, v in data.items()}
zero_optim.zero_grad()
torch_optim.zero_grad()
torch_loss = run_fwd_bwd(torch_model, data, label, criterion, torch_optim)
loss = run_fwd_bwd(model, data, label, criterion, zero_optim)
# as no master weights leads to error accumulation, we don't check the loss
if master_weights:
assert_close(torch_loss, loss)
run_fwd_bwd(torch_model, data, output_transform_fn, loss_fn, optimizer=torch_optim)
run_fwd_bwd(model, data, output_transform_fn, loss_fn, optimizer=zero_optim)
import apex.amp as apex_amp