ColossalAI/tests/test_tensor/test_model.py

437 lines
16 KiB
Python

from colossalai.tensor.colo_parameter import ColoParameter
from tests.components_to_test.registry import non_distributed_component_funcs
import colossalai
import pytest
import torch
import torch.multiprocessing as mp
from colossalai.testing import parameterize, rerun_if_address_is_in_use
from colossalai.utils.cuda import get_current_device
from colossalai.utils import free_port
from colossalai.utils import ColoInitContext
from colossalai.tensor import named_params_with_colotensor, TensorSpec, ComputePattern, \
ParallelAction, ColoTensor, ColoOptimizer, dist_spec, DistSpecManager
from colossalai.context import ParallelMode
from colossalai.core import global_context as gpc
from functools import partial
import random
import os
import numpy as np
# Hack huggingface Bert ModelOutput
# Make it available to our ColoTensor
from transformers.file_utils import ModelOutput
from dataclasses import fields
def _post_init_colotensor(self):
class_fields = fields(self)
# Safety and consistency checks
if len(class_fields) == 0:
raise ValueError(f"{self.__class__.__name__} has no fields.")
if not all(field.default is None for field in class_fields[1:]):
raise ValueError(f"{self.__class__.__name__} should not have more than one required field.")
first_field = getattr(self, class_fields[0].name)
other_fields_are_none = all(getattr(self, field.name) is None for field in class_fields[1:])
def is_tensor_with_colo(x):
"""
Tests if `x` is a `ColoTensor` or `torch.Tensor`.
"""
if isinstance(x, torch.Tensor):
return True
return isinstance(x, ColoTensor)
if other_fields_are_none and not is_tensor_with_colo(first_field):
if isinstance(first_field, dict):
iterator = first_field.items()
first_field_iterator = True
else:
try:
iterator = iter(first_field)
first_field_iterator = True
except TypeError:
first_field_iterator = False
# if we provided an iterator as first field and the iterator is a (key, value) iterator
# set the associated fields
if first_field_iterator:
for element in iterator:
if (not isinstance(element, (list, tuple)) or not len(element) == 2 or not isinstance(element[0], str)):
break
setattr(self, element[0], element[1])
if element[1] is not None:
self[element[0]] = element[1]
elif first_field is not None:
self[class_fields[0].name] = first_field
else:
for field in class_fields:
v = getattr(self, field.name)
if v is not None:
self[field.name] = v
ModelOutput.__post_init__ = _post_init_colotensor
# complete the hack
def set_seed(seed):
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
def init_1d_row_linear(weight):
spec = TensorSpec(
dist_spec.shard(gpc.get_group(ParallelMode.PARALLEL_1D), [-1], [gpc.get_world_size(ParallelMode.PARALLEL_1D)]),
[ParallelAction(priority=1, compute_pattern=ComputePattern.TP1D, parallel_mode=ParallelMode.PARALLEL_1D)])
with DistSpecManager.no_grad():
weight.set_spec(spec)
def init_1d_col_linear(weight, gather_out=True):
spec = TensorSpec(
dist_spec.shard(gpc.get_group(ParallelMode.PARALLEL_1D), [0], [gpc.get_world_size(ParallelMode.PARALLEL_1D)]), [
ParallelAction(priority=1,
compute_pattern=ComputePattern.TP1D,
parallel_mode=ParallelMode.PARALLEL_1D,
gather_out=gather_out)
])
with DistSpecManager.no_grad():
weight.set_spec(spec)
def init_1d_row_embedding(weight):
spec = TensorSpec(
dist_spec.shard(gpc.get_group(ParallelMode.PARALLEL_1D), [0], [gpc.get_world_size(ParallelMode.PARALLEL_1D)]),
[ParallelAction(priority=1, compute_pattern=ComputePattern.TP1D, parallel_mode=ParallelMode.PARALLEL_1D)])
with DistSpecManager.no_grad():
weight.set_spec(spec)
def init_1d_col_embedding(weight):
spec = TensorSpec(
dist_spec.shard(gpc.get_group(ParallelMode.PARALLEL_1D), [-1], [gpc.get_world_size(ParallelMode.PARALLEL_1D)]),
[ParallelAction(priority=1, compute_pattern=ComputePattern.TP1D, parallel_mode=ParallelMode.PARALLEL_1D)])
with DistSpecManager.no_grad():
weight.set_spec(spec)
def run_1d_hybrid_tp(model_name):
# A simple net with two stacked nn.Linear
get_components_func = non_distributed_component_funcs.get_callable(model_name)
model_builder, train_dataloader, test_dataloader, optimizer_class, criterion = get_components_func()
rank = gpc.get_local_rank(ParallelMode.PARALLEL_1D)
set_seed(1)
with ColoInitContext(device=get_current_device()):
model = model_builder(checkpoint=True)
if rank == 0:
model_torch = model_builder(checkpoint=True)
model_torch = model_torch.cuda()
colo_optimizer_torch = ColoOptimizer(dict(model_torch.named_parameters()), torch.optim.SGD, lr=0.1)
# Make two models have the same init params
for p1, p2 in zip(model.parameters(), model_torch.parameters()):
p2.data.copy_(p1.data)
if 'bert' == model_name:
for name, p in model.colo_named_parameters():
if not isinstance(p, ColoTensor):
continue
# print(name)
# num_class = type_vocab_size = 2 | (8, 2)
if 'classifier' in name and 'weight' in name:
init_1d_row_linear(p)
# num_class = vocab_size = 30524 | (30524, 8)
if 'word_embeddings' in name and 'weight' in name:
init_1d_row_embedding(p)
# num_class = seq_len = 512 | (512, 8)
if 'position_embeddings' in name and 'weight' in name:
init_1d_row_embedding(p)
# num_class = type_vocab_size = 2 | (2, 8)
if 'token_type_embeddings' in name and 'weight' in name:
init_1d_col_embedding(p)
elif "simple_net" == model_name:
# A naive way to set spec for all weights in Linear
for name, p in model.colo_named_parameters():
if not isinstance(p, ColoTensor):
continue
if 'embed' in name and 'weight' in name:
init_1d_col_embedding(p)
if 'proj1' in name and ('weight' in name or 'bias' in name):
init_1d_col_linear(p)
if 'proj2' in name and 'weight' in name:
init_1d_row_linear(p)
if 'classifier' in name and ('weight' in name or 'bias' in name):
init_1d_col_linear(p, gather_out=False)
model = model.cuda()
colo_optimizer = ColoOptimizer(dict(model.named_parameters()), torch.optim.SGD, lr=0.1)
for i, (data, label) in enumerate(train_dataloader):
model.eval()
colo_optimizer.zero_grad()
if rank == 0:
model_torch.eval()
colo_optimizer_torch.zero_grad()
data = data.to(get_current_device())
label = label.to(get_current_device())
torch.distributed.broadcast(data, 0, group=gpc.get_group(ParallelMode.PARALLEL_1D))
torch.distributed.broadcast(label, 0, group=gpc.get_group(ParallelMode.PARALLEL_1D))
# Bcast rank0 data to all processes
if criterion:
output = model(data)
loss = criterion(output, label)
else:
output = model(data, label)
loss = output
# For reference
if rank == 0:
if criterion:
output_torch = model_torch(data)
loss_torch = criterion(output_torch, label)
else:
output_torch = model_torch(data, label)
loss_torch = output_torch
if rank == 0:
# print(loss.torch_tensor().item())
# print('loss torch', loss_torch.item())
with torch.no_grad():
assert torch.allclose(loss.torch_tensor(), loss_torch, rtol=1e-2)
loss.backward()
colo_optimizer.step()
if rank == 0:
loss_torch.backward()
colo_optimizer_torch.step()
with torch.no_grad():
# check param
for p1, p2 in zip(model.parameters(), model_torch.parameters()):
if p1.size() == p2.size():
assert torch.allclose(p1, p2)
else:
# TODO(jzy) Only check 1D spec. Need to be replaced by new DistSpec.
if p1.size(-1) < p2.size(-1): # col
world_size = p2.size(-1) // p1.size(-1)
split_p2 = torch.chunk(p2, world_size, dim=-1)[0]
elif p1.size(0) < p2.size(0): # row
world_size = p2.size(0) // p1.size(0)
split_p2 = torch.chunk(p2, world_size, dim=0)[0]
assert torch.allclose(p1, split_p2)
if i > 5:
break
# Test the overrided parameters() and named_parameters() member functions
def test_model_parameters():
# build a module with 2 Linear, 4 parameters in total.
class Net(torch.nn.Module):
def __init__(self):
super().__init__()
self.fcs = torch.nn.Sequential(torch.nn.Linear(2, 3), torch.nn.Linear(3, 2))
self.extra_param = torch.nn.Parameter(torch.randn(2))
with ColoInitContext(device=get_current_device()):
model = Net()
param_cnt = 0
for name, p in model.named_parameters():
param_cnt += 1
assert param_cnt == 5
for name, colo_p in model.colo_named_parameters():
assert colo_p.is_model_data()
param_cnt = 0
for name, p in model.named_parameters(recurse=False):
param_cnt += 1
assert param_cnt == 1
param_cnt = 0
for p in model.fcs[0].parameters(recurse=False):
param_cnt += 1
assert param_cnt == 2
def test_colo_optimizer():
get_components_func = non_distributed_component_funcs.get_callable('simple_net')
model_builder, train_dataloader, test_dataloader, optimizer_class, criterion = get_components_func()
set_seed(1)
with ColoInitContext(lazy_memory_allocate=False, device=get_current_device()):
model = model_builder(checkpoint=True)
colo_optimizer = ColoOptimizer(dict(model.named_parameters()), torch.optim.SGD, lr=0.1)
for i, (data, label) in enumerate(train_dataloader):
colo_optimizer.zero_grad()
data = data.to(get_current_device())
label = label.to(get_current_device())
# Bcast rank0 data to all processes
if criterion:
output = model(data)
loss = criterion(output, label)
else:
output = model(data, label)
loss = output
loss.backward()
colo_optimizer.step()
if i > 5:
break
def run_1d_row_tp(model_name: str):
# A simple net with two stacked nn.Linear
get_components_func = non_distributed_component_funcs.get_callable(model_name)
model_builder, train_dataloader, test_dataloader, optimizer_class, criterion = get_components_func()
rank = gpc.get_local_rank(ParallelMode.PARALLEL_1D)
set_seed(1)
with ColoInitContext(device=get_current_device()):
model = model_builder(checkpoint=True)
set_seed(1)
if rank == 0:
model_torch = model_builder(checkpoint=True)
model_torch = model_torch.cuda()
# A naive way to set spec for all weights in Linear
for name, p in model.colo_named_parameters():
if not isinstance(p, ColoTensor):
continue
if 'weight' in name and 'LayerNorm' not in name and 'ln' not in name and 'embed' not in name:
init_1d_row_linear(p)
if 'embed' in name and 'weight' in name:
init_1d_row_embedding(p)
model = model.cuda()
for i, (data, label) in enumerate(train_dataloader):
data = data.to(get_current_device())
label = label.to(get_current_device())
torch.distributed.broadcast(data, 0, group=gpc.get_group(ParallelMode.PARALLEL_1D))
torch.distributed.broadcast(label, 0, group=gpc.get_group(ParallelMode.PARALLEL_1D))
# Bcast rank0 data to all processes
if criterion:
output = model(data)
loss = criterion(output, label)
else:
output = model(data, label)
loss = output
# For reference
if rank == 0:
if criterion:
output_torch = model_torch(data)
loss_torch = criterion(output_torch, label)
else:
output_torch = model_torch(data, label)
loss_torch = output_torch
if rank == 0:
# print(loss.torch_tensor().item())
# print('loss torch', loss_torch.item())
assert torch.allclose(loss.torch_tensor(), loss_torch, rtol=1e-2)
loss.backward()
if rank == 0:
loss_torch.backward()
if i > 5:
break
def _run_pretrain_load():
from _utils import check_equal
from transformers import BertForMaskedLM
set_seed(1)
model_pretrained = BertForMaskedLM.from_pretrained('bert-base-uncased')
with ColoInitContext(lazy_memory_allocate=False, device=get_current_device()):
model = BertForMaskedLM.from_pretrained('bert-base-uncased')
model_pretrained = model_pretrained.cuda()
model = model.cuda()
dict_pretrained = {}
dict_col = {}
c_ref = 0
for name, param in model_pretrained.named_parameters():
dict_pretrained[name] = param
c_ref += 1
c1 = 0
c2 = 0
for name, param in model.colo_named_parameters():
if isinstance(param, ColoParameter):
c1 += 1
else:
c2 += 1
dict_col[name] = param
assert c_ref == c1
assert c2 == 0
if model_pretrained.cls.predictions.decoder.bias is model_pretrained.cls.predictions.bias:
assert model.cls.predictions.decoder.bias is model.cls.predictions.bias
for name, param in dict_pretrained.items():
check_equal(param, dict_col[name])
def run_model_dist(rank, world_size, port):
config = dict(parallel=dict(tensor=dict(mode="1d", size=world_size),))
colossalai.launch(config=config, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
for name in ['simple_net']:
run_1d_row_tp(name)
for name in ['bert', 'simple_net']:
run_1d_hybrid_tp(name)
@pytest.mark.dist
@pytest.mark.parametrize('world_size', [1, 4])
# @parameterize('world_size', [1, 4])
@rerun_if_address_is_in_use()
def test_model(world_size):
run_func = partial(run_model_dist, world_size=world_size, port=free_port())
mp.spawn(run_func, nprocs=world_size)
def run_pretrain_load_dist(rank, world_size, port):
config = dict(parallel=dict(tensor=dict(mode="1d", size=world_size),))
colossalai.launch(config=config, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
_run_pretrain_load()
# The test case has to download huggingface pretrained models from the internet
# So we manually trigger the test.
@pytest.mark.dist
@pytest.mark.parametrize('world_size', [1, 4])
@rerun_if_address_is_in_use()
def _test_pretrain_load(world_size):
run_func = partial(run_pretrain_load_dist, world_size=world_size, port=free_port())
mp.spawn(run_func, nprocs=world_size)
if __name__ == '__main__':
# test_model_parameters()
# test_colo_optimizer()
test_model(4)
# _test_pretrain_load(4)