Merge branch 'main' into feature/shardformer

tests/test_booster/test_plugin/test_gemini_plugin.py

@@ -98,7 +98,7 @@ def check_gemini_plugin(subset: str, init_method: str = 'none', early_stop: bool
         ]:
             continue
         err = run_fn(init_method, model_fn, data_gen_fn, output_transform_fn)
-
+        torch.cuda.empty_cache()
         if err is None:
             passed_models.append(name)
         else:

tests/test_checkpoint_io/test_low_level_zero_checkpoint_io.py

@@ -14,6 +14,7 @@ from colossalai.testing import (
     rerun_if_address_is_in_use,
     spawn,
 )
+from colossalai.zero import LowLevelZeroOptimizer
 
 
 # stage 1 and 2 process the optimizer/mode the same way
@@ -50,6 +51,17 @@ def check_low_level_zero_checkpointIO(stage: int, shard: bool, offload: bool):
 
         booster.load_model(new_model, model_ckpt_path)
         check_state_dict_equal(model.state_dict(), new_model.state_dict(), False)
+        # check master weight
+        assert isinstance(new_optimizer, LowLevelZeroOptimizer)
+        working_param_id_set = set(id(p) for p in new_model.parameters())
+        for p_id, master_param in new_optimizer._param_store.working_to_master_param.items():
+            assert p_id in working_param_id_set
+            working_param = new_optimizer._param_store.master_to_working_param[id(master_param)]
+            padding = new_optimizer._param_store.get_param_padding_size(working_param)
+            padded_param = torch.nn.functional.pad(working_param.data.view(-1), (0, padding))
+            working_shard = padded_param.chunk(dist.get_world_size())[dist.get_rank()]
+            assert torch.equal(working_shard,
+                               master_param.data.view(-1).to(dtype=padded_param.dtype, device=padded_param.device))
 
         booster.load_optimizer(new_optimizer, optimizer_ckpt_path)
         check_state_dict_equal(optimizer.optim.state_dict(), new_optimizer.optim.state_dict(), False)

tests/test_data_pipeline_tensor_parallel/test_cifar_with_data_pipeline_tensor.py

@@ -1,100 +0,0 @@
-import os
-from pathlib import Path
-
-import pytest
-import torch
-from torchvision import transforms
-from torchvision.datasets import CIFAR10
-
-import colossalai
-from colossalai.amp import AMP_TYPE
-from colossalai.context import ParallelMode
-from colossalai.core import global_context as gpc
-from colossalai.logging import get_dist_logger
-from colossalai.nn import CrossEntropyLoss
-from colossalai.nn.lr_scheduler import CosineAnnealingWarmupLR
-from colossalai.pipeline.pipelinable import PipelinableContext
-from colossalai.testing import rerun_if_address_is_in_use, skip_if_not_enough_gpus, spawn
-from colossalai.trainer import Trainer, hooks
-from colossalai.utils import get_dataloader
-
-BATCH_SIZE = 4
-NUM_EPOCHS = 60
-WARMUP_EPOCHS = 5
-CONFIG = dict(NUM_MICRO_BATCHES=2,
-              parallel=dict(pipeline=2, tensor=dict(size=2, mode='1d')),
-              fp16=dict(mode=AMP_TYPE.NAIVE),
-              gradient_accumulation=2)
-
-
-def run_trainer(rank, world_size, port):
-    colossalai.launch(config=CONFIG, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
-
-    logger = get_dist_logger()
-
-    # get logger
-    logger = get_dist_logger()
-
-    pipelinable = PipelinableContext()
-    try:
-        from titans.model.vit import vit_tiny_patch4_32
-    except ImportError:
-        logger.warning('skip the test_cifar_with_data_pipeline_tensor test because titan is not installed')
-        logger.warning('please install titan from https://github.com/hpcaitech/Titans')
-        return
-    with pipelinable:
-        model = vit_tiny_patch4_32()
-    pipelinable.to_layer_list()
-    pipelinable.policy = "uniform"
-    model = pipelinable.partition(1, gpc.pipeline_parallel_size, gpc.get_local_rank(ParallelMode.PIPELINE))
-
-    # create dataloaders
-    root = Path(os.environ['DATA'])
-    transform_train = transforms.Compose([
-        transforms.RandomCrop(32, padding=4, pad_if_needed=True),
-        transforms.AutoAugment(policy=transforms.AutoAugmentPolicy.CIFAR10),
-        transforms.ToTensor(),
-        transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
-    ])
-    train_dataset = CIFAR10(root=root, train=True, download=True, transform=transform_train)
-    train_dataloader = get_dataloader(dataset=train_dataset, shuffle=True, batch_size=BATCH_SIZE, pin_memory=True)
-
-    # create loss function
-    criterion = CrossEntropyLoss(label_smoothing=0.1)
-
-    # create optimizer
-    optimizer = torch.optim.AdamW(model.parameters(), lr=0.001, weight_decay=0)
-
-    # create lr scheduler
-    lr_scheduler = CosineAnnealingWarmupLR(optimizer=optimizer, total_steps=NUM_EPOCHS, warmup_steps=WARMUP_EPOCHS)
-
-    # initialize
-    engine, train_dataloader, *_ = colossalai.initialize(model=model,
-                                                         optimizer=optimizer,
-                                                         criterion=criterion,
-                                                         train_dataloader=train_dataloader)
-
-    logger = get_dist_logger()
-
-    trainer = Trainer(engine=engine, logger=logger)
-
-    hook_list = [
-        hooks.LRSchedulerHook(lr_scheduler=lr_scheduler, by_epoch=False),
-    ]
-
-    trainer.fit(train_dataloader=train_dataloader,
-                epochs=NUM_EPOCHS,
-                max_steps=2,
-                hooks=hook_list,
-                display_progress=True)
-
-
-@pytest.mark.dist
-@skip_if_not_enough_gpus(min_gpus=8)
-@rerun_if_address_is_in_use()
-def test_hybrid_parallel():
-    spawn(run_trainer, 8)
-
-
-if __name__ == '__main__':
-    test_hybrid_parallel()

tests/test_legacy/test_trainer/test_trainer_with_non_pipe_schedule.py

@@ -3,9 +3,9 @@ import torch
 
 import colossalai
 from colossalai.amp.amp_type import AMP_TYPE
+from colossalai.legacy.trainer import Trainer
 from colossalai.logging import get_dist_logger
 from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
-from colossalai.trainer import Trainer
 from colossalai.utils import MultiTimer
 from tests.components_to_test.registry import non_distributed_component_funcs
 

tests/test_legacy/test_trainer/test_trainer_with_pipe_schedule.py

@@ -12,9 +12,9 @@ from torchvision.models import resnet18
 import colossalai
 from colossalai.context.parallel_mode import ParallelMode
 from colossalai.core import global_context as gpc
+from colossalai.legacy.trainer import Trainer
 from colossalai.logging import get_dist_logger
 from colossalai.testing import rerun_if_address_is_in_use, spawn
-from colossalai.trainer import Trainer
 from colossalai.utils import MultiTimer, get_dataloader
 
 BATCH_SIZE = 4

tests/test_moe/test_grad_handler.py

@@ -5,7 +5,7 @@ import torch.nn as nn
 
 import colossalai
 from colossalai.context.moe_context import MOE_CONTEXT
-from colossalai.engine.gradient_handler import MoeGradientHandler
+from colossalai.legacy.engine.gradient_handler import MoeGradientHandler
 from colossalai.nn.layer.moe import Experts, MoeLayer, Top1Router, UniformNoiseGenerator
 from colossalai.testing import assert_equal_in_group, rerun_if_address_is_in_use, spawn
 from colossalai.utils import get_current_device

tests/test_moe/test_moe_zero_model.py

@@ -3,7 +3,7 @@ import torch
 
 import colossalai
 from colossalai.context import MOE_CONTEXT
-from colossalai.engine.gradient_handler import MoeGradientHandler
+from colossalai.legacy.engine.gradient_handler import MoeGradientHandler
 from colossalai.nn import MoeLoss
 from colossalai.testing import assert_equal_in_group, parameterize, rerun_if_address_is_in_use, spawn
 from colossalai.zero.legacy.init_ctx import ZeroInitContext

tests/test_moe/test_moe_zero_optim.py

@@ -4,7 +4,7 @@ import torch
 import colossalai
 from colossalai.amp import convert_to_apex_amp
 from colossalai.context import MOE_CONTEXT
-from colossalai.engine.gradient_handler import MoeGradientHandler
+from colossalai.legacy.engine.gradient_handler import MoeGradientHandler
 from colossalai.nn import MoeLoss
 from colossalai.nn.optimizer import CPUAdam
 from colossalai.testing import assert_equal_in_group, parameterize, rerun_if_address_is_in_use, spawn

tests/test_pipeline/test_cuda_rpc_performance.py

@@ -1,25 +1,16 @@
 import os
-from typing import Callable, List, Optional, Type, Union
 import time
 
 import pytest
 import torch
 import torch.nn as nn
+from rpc_test_utils import parse_args, rpc_run
 from titans.dataloader.cifar10 import build_cifar
 from torchvision.models import resnet50
-from torchvision.models.resnet import BasicBlock, Bottleneck, conv1x1
 from tqdm import tqdm
 
-from rpc_test_utils import rpc_run, parse_args
-import colossalai
-import colossalai.nn as col_nn
-from colossalai.logging import disable_existing_loggers, get_dist_logger
-from colossalai.trainer import Trainer, hooks
-from colossalai.utils import MultiTimer, get_dataloader
-from colossalai.context import ParallelMode
-from colossalai.pipeline.pipelinable import PipelinableContext, PipelinableModel
-from colossalai.pipeline.rpc import OneFOneBPipelineEngine, ChimeraPipelineEngine
-from colossalai.pipeline.pipeline_process_group import ppg
+from colossalai.pipeline.pipelinable import PipelinableContext
+from colossalai.pipeline.rpc import OneFOneBPipelineEngine
 
 
 def flatten(x):