ColossalAI/tests/test_shardformer/test_model/test_shard_opt.py

100 lines
4.0 KiB
Python
Raw Normal View History

import copy
import os
import pytest
import torch
import colossalai
from colossalai.logging import disable_existing_loggers
from colossalai.tensor.d_tensor.api import is_customized_distributed_tensor, is_distributed_tensor
from colossalai.testing import (
assert_hf_output_close,
clear_cache_before_run,
parameterize,
rerun_if_address_is_in_use,
spawn,
)
from tests.kit.model_zoo import model_zoo
from tests.test_shardformer.test_model._utils import build_model, check_state_dict, run_forward
os.environ['TRANSFORMERS_NO_ADVISORY_WARNINGS'] = 'true'
def check_forward_backward(org_model, sharded_model, data_gen_fn, output_transform_fn, loss_fn):
org_output, org_loss, shard_output, shard_loss = run_forward(org_model, sharded_model, data_gen_fn,
output_transform_fn, loss_fn)
assert_hf_output_close(org_output, shard_output, ignore_keys=['past_key_values'], rtol=1e-4)
# run backward
org_loss.backward()
shard_loss.backward()
assert torch.allclose(org_loss, shard_loss,
atol=1e-5), f"shard model loss is not equal to orgin model loss\n{org_loss}\n{shard_loss}"
# unwrap model
if hasattr(org_model, 'model'):
opt_model = org_model.model
shard_opt_model = sharded_model.model
else:
opt_model = org_model
shard_opt_model = sharded_model
# check attention grad
org_grad = opt_model.decoder.layers[0].self_attn.q_proj.weight.grad
shard_grad = shard_opt_model.decoder.layers[0].self_attn.q_proj.weight.grad
shard_weight = shard_opt_model.decoder.layers[0].self_attn.q_proj.weight
if is_distributed_tensor(shard_weight) or is_customized_distributed_tensor(shard_weight):
shard_grad_list = [torch.zeros([*shard_grad.shape]).to('cuda') for _ in range(4)]
torch.distributed.all_gather(shard_grad_list, shard_grad)
all_shard_grad = torch.cat(shard_grad_list, dim=0)
else:
all_shard_grad = shard_grad
assert torch.allclose(org_grad, all_shard_grad,
atol=1e-5), f"shard model grad is not equal to orgin model grad\n{org_grad}\n{all_shard_grad}"
# check embedding grad
org_grad = opt_model.decoder.embed_tokens.weight.grad
shard_grad = shard_opt_model.decoder.embed_tokens.weight.grad
shard_weight = shard_opt_model.decoder.embed_tokens.weight
if is_distributed_tensor(shard_weight) or is_customized_distributed_tensor(shard_weight):
shard_grad_list = [torch.zeros([*shard_grad.shape]).to('cuda') for _ in range(4)]
torch.distributed.all_gather(shard_grad_list, shard_grad)
all_shard_grad = torch.cat(shard_grad_list, dim=0)
else:
all_shard_grad = shard_grad
assert torch.allclose(org_grad, all_shard_grad,
atol=1e-5), f"shard model grad is not equal to orgin model grad\n{org_grad}\n{all_shard_grad}"
@parameterize('enable_fused_normalization', [True, False])
@parameterize('enable_tensor_parallelism', [True, False])
@parameterize('use_lazy_init', [False, True])
def run_t5_test(enable_fused_normalization, enable_tensor_parallelism, use_lazy_init):
sub_model_zoo = model_zoo.get_sub_registry('transformers_opt')
for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items():
org_model, sharded_model = build_model(model_fn, enable_fused_normalization, enable_tensor_parallelism,
use_lazy_init)
check_state_dict(org_model, sharded_model, name=name)
check_forward_backward(org_model, sharded_model, data_gen_fn, output_transform_fn, loss_fn)
torch.cuda.empty_cache()
def check_OPTModel(rank, world_size, port):
disable_existing_loggers()
colossalai.launch(config={}, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
run_t5_test()
@pytest.mark.dist
@rerun_if_address_is_in_use()
@clear_cache_before_run()
def test_OPTModel():
spawn(check_OPTModel, 4)
if __name__ == '__main__':
test_OPTModel()