from copy import copy import pytest from functools import partial import torch import torch.multiprocessing as mp from colossalai.tensor import TensorSpec, ComputePattern, ParallelAction from colossalai.nn import init_colo_module, check_colo_module from _utils import tensor_equal, tensor_shard_equal, set_seed import colossalai from colossalai.utils.cuda import get_current_device from colossalai.utils.model.colo_init_context import ColoInitContext from colossalai.context.parallel_mode import ParallelMode from colossalai.tensor import distspec from colossalai.testing import rerun_if_address_is_in_use from colossalai.utils import free_port from colossalai.core import global_context as gpc from tests.components_to_test.registry import non_distributed_component_funcs def run_model_with_spec(mode, model_name): 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=False) if rank == 0: model_seq = model_builder(checkpoint=False) model_seq = model_seq.cuda() # Make two models have the same init params for p1, p2 in zip(model.parameters(), model_seq.parameters()): p2.data.copy_(p1.data) parallel_action = ParallelAction(ComputePattern.TP1D) # Not all layers in Bert can be mod by 4. # e.g. row shard for all layers is invalid because the first dim of some layer is the classification type size 2. if 'bert' == model_name: if 'col' == mode: init_colo_module(model.bert.embeddings, parallel_action, recursive=True, mode=mode) init_colo_module(model.bert.encoder, parallel_action, recursive=True, mode=mode) init_colo_module(model.classifier, parallel_action, recursive=True, mode='row') elif 'row' == mode: init_colo_module(model.bert.embeddings, parallel_action, recursive=True, mode='col') init_colo_module(model.bert.encoder, parallel_action, recursive=True, mode=mode) init_colo_module(model.classifier, parallel_action, recursive=True, mode=mode) elif 'simple_net' == model_name: init_colo_module(model, parallel_action, recursive=True, mode=mode) 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)) if criterion: output = model(data) loss = criterion(output, label) else: output = model(data, label) loss = output # For reference if rank == 0: if criterion: output_seq = model_seq(data) loss_seq = criterion(output_seq, label) else: output_seq = model_seq(data, label) loss_seq = output_seq if rank == 0: with torch.no_grad(): assert torch.allclose(loss, loss_seq, rtol=1e-2) loss.backward() if rank == 0: loss_seq.backward() with torch.no_grad(): # check param for p1, p2 in zip(model.parameters(), model_seq.parameters()): if p1.size() == p2.size(): assert torch.allclose(p1, p2) else: 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 > 3: break def run_linear_with_spec(mode): with ColoInitContext(device=get_current_device()): model = torch.nn.Linear(4, 8) model_handy = copy(model) parallel_action = ParallelAction(ComputePattern.TP1D) init_colo_module(model, parallel_action, recursive=True, mode=mode) x = torch.rand(2, 4).cuda() out = model(x) colo_out = model_handy(x) assert tensor_equal(out, colo_out) grad = torch.rand_like(out) out.backward(grad) colo_out.backward(grad) assert tensor_shard_equal(model.weight.grad, model_handy.weight.grad) assert tensor_shard_equal(model.bias.grad, model_handy.bias.grad) def run_check_shared_param(): from transformers import BertForMaskedLM, BertConfig hidden_dim = 8 num_head = 4 sequence_length = 12 num_layer = 2 vocab_size = 24 config = BertConfig(vocab_size=vocab_size, hidden_size=hidden_dim, intermediate_size=hidden_dim * 4, num_attention_heads=num_head, max_position_embeddings=sequence_length, num_hidden_layers=num_layer, hidden_dropout_prob=0., attention_probs_dropout_prob=0.) with ColoInitContext(lazy_memory_allocate=False, device=get_current_device()): model = BertForMaskedLM(config) model = model.cuda() parallel_action = ParallelAction(ComputePattern.TP1D) # model.cls.predictions.decoder and model.cls.predictions share the bias, so they should have the same spec assert len(model.cls.predictions.decoder.bias.shared_param_modules) == 2 # They are all Linear, so both row is allowed. This should pass check. init_colo_module(model, parallel_action, recursive=True, mode='row') # This should be detected by check because you can not set weight as row while set bias as col. col_spec = TensorSpec( distspec.shard(gpc.get_group(ParallelMode.PARALLEL_1D), [0], [gpc.get_world_size(ParallelMode.PARALLEL_1D)]), ParallelAction(ComputePattern.TP1D)) model.cls.predictions.bias.set_spec(col_spec) try: check_colo_module(model.cls.predictions.decoder, recursive=False) except Exception as e: assert 'incorrectly sharded' in str(e) def run_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_linear_with_spec('col') run_linear_with_spec('row') def run_dist_model(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 model_name in ['simple_net', 'bert']: run_model_with_spec('col', model_name) run_model_with_spec('row', model_name) def run_dist_check(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_check_shared_param() @pytest.mark.dist @pytest.mark.parametrize('world_size', [1, 4]) @rerun_if_address_is_in_use() def test_module_linear_1d(world_size): run_func = partial(run_dist, world_size=world_size, port=free_port()) mp.spawn(run_func, nprocs=world_size) @pytest.mark.dist @pytest.mark.parametrize('world_size', [1, 4]) @rerun_if_address_is_in_use() def test_module_model(world_size): run_func = partial(run_dist_model, world_size=world_size, port=free_port()) mp.spawn(run_func, nprocs=world_size) @pytest.mark.dist @pytest.mark.parametrize('world_size', [1, 2]) @rerun_if_address_is_in_use() def test_module_check(world_size): run_func = partial(run_dist_check, world_size=world_size, port=free_port()) mp.spawn(run_func, nprocs=world_size) if __name__ == '__main__': test_module_check(2)