[zero] adapt zero for unsharded paramters (Optimizer part) (#601)

colossalai/utils/checkpointing.py

@@ -6,7 +6,10 @@ import torch.distributed as dist
 from colossalai.communication.collective import scatter_object_list
 from colossalai.context.parallel_mode import ParallelMode
 from colossalai.core import global_context as gpc
-from torch.nn.modules.module import _EXTRA_STATE_KEY_SUFFIX
+try:
+    from torch.nn.modules.module import _EXTRA_STATE_KEY_SUFFIX
+except ImportError:
+    _EXTRA_STATE_KEY_SUFFIX = '_extra_state'
 
 from .common import is_using_pp
 

colossalai/zero/init_ctx/init_context.py

@@ -11,6 +11,7 @@ from colossalai.zero.shard_utils import BaseShardStrategy
 from colossalai.zero.sharded_model._utils import cast_tensor_to_fp16
 from colossalai.zero.sharded_param import ShardedParamV2
 from torch.distributed import ProcessGroup
+from contextlib import AbstractContextManager
 
 
 def _substitute_init_recursively(cls, func):
@@ -88,6 +89,7 @@ class ZeroContextConfig(object):
     """The configuration used to control zero context initialization.
 
     Args:
+        target_device (torch.device): The device where param data are after exiting the context.
         replicated (bool, optional): Whether the param is replicated across data parallel group.
             Some parameters are not replicated, e.g. parameters in MOE experts.
         shard_param (bool, optional): Is param sharded after exiting the context. Defaults to False.
@@ -99,8 +101,13 @@ class ZeroContextConfig(object):
             See torchvision resnet18. Defaults to False.
     """
 
-    def __init__(self, replicated: bool = True, shard_param: bool = False, rm_torch_payload_on_the_fly: bool = False):
+    def __init__(self,
+                 target_device: torch.device,
+                 replicated: bool = True,
+                 shard_param: bool = False,
+                 rm_torch_payload_on_the_fly: bool = False):
         super().__init__()
+        self.target_device = target_device
         self.is_replicated: bool = replicated
         self.shard_param: bool = shard_param
         self.rm_torch_payload_on_the_fly: bool = rm_torch_payload_on_the_fly
@@ -114,7 +121,7 @@ class ZeroInitContext(InsertPostInitMethodToModuleSubClasses):
     3. Shard the param and grad according to flags.
 
     Args:
-        target_device (torch.device): The device where param data after exiting the context.
+        target_device (torch.device): The device where param data are after exiting the context.
         shard_strategy (BaseShardStrategy): Shard strategy instance.
         shard_param (bool, optional): Is param sharded after exiting the context. Defaults to False.
         rm_torch_payload_on_the_fly (bool, optional): If set to `True`, remove tensor payload on `param.data` after module init finished.
@@ -136,17 +143,22 @@ class ZeroInitContext(InsertPostInitMethodToModuleSubClasses):
                  dp_process_group: Optional[ProcessGroup] = None):
 
         super().__init__()
-        self.target_device = target_device
         self.shard_strategy = shard_strategy
         self.initialized_param_list = []
         self.model_numel_tensor = model_numel_tensor
         self.dp_process_group = dp_process_group or gpc.get_group(ParallelMode.DATA)
 
-        self.config = ZeroContextConfig(replicated=True,
+        self.config = ZeroContextConfig(target_device=target_device,
+                                        replicated=True,
                                         shard_param=shard_param,
                                         rm_torch_payload_on_the_fly=rm_torch_payload_on_the_fly)
+
         ZeroContextMgr().current_context = self
 
+    @property
+    def target_device(self):
+        return self.config.target_device
+
     @property
     def is_replicated(self):
         return self.config.is_replicated
@@ -235,8 +247,9 @@ class ZeroContextMgr(metaclass=SingletonMeta):
             self.current_context.config = old_config
 
 
-def no_shard_zero_context(is_replicated: bool = True):
+def no_shard_zero_context(is_replicated: bool = True) -> AbstractContextManager:
-    return ZeroContextMgr().hijack_context_config(replicated=is_replicated,
+    return ZeroContextMgr().hijack_context_config(target_device=torch.device('cuda', torch.cuda.current_device()),
+                                                  replicated=is_replicated,
                                                   shard_param=False,
                                                   rm_torch_payload_on_the_fly=False)
 

colossalai/zero/sharded_optim/sharded_optim_v2.py

@@ -12,13 +12,12 @@ from colossalai.logging import get_dist_logger
 from colossalai.nn.optimizer import ColossalaiOptimizer
 from colossalai.utils.memory_tracer.model_data_memtracer import \
     GLOBAL_MODEL_DATA_TRACER
-from colossalai.zero.shard_utils.tensor_utils import (colo_model_tensor_clone, colo_tensor_mem_usage)
+from colossalai.zero.shard_utils.tensor_utils import (colo_model_data_tensor_move_inline, colo_model_tensor_clone,
+                                                      colo_tensor_mem_usage)
 from colossalai.zero.sharded_model import ShardedModelV2
 from colossalai.zero.sharded_model._utils import cast_tensor_to_fp32
 from colossalai.zero.sharded_optim._utils import has_inf_or_nan
 from colossalai.zero.sharded_param.tensorful_state import (StatefulTensor, TensorState)
-from colossalai.zero.shard_utils.tensor_utils import colo_model_data_tensor_move_inline
-
 from torch import Tensor
 from torch.distributed import ProcessGroup
 from torch.nn.parameter import Parameter
@@ -69,6 +68,9 @@ class ShardedOptimizerV2(ColossalaiOptimizer):
         backoff_factor (float, optional): backoff_factor used by DynamicGradScaler. Defaults to 0.5.
         growth_interval (float, optional): growth_interval used by DynamicGradScaler. Defaults to 1000.
         hysteresis (float, optional): hysteresis used by DynamicGradScaler. Defaults to 2.
+        keep_unsharded (bool, optional): if True, optimizer won't shard unsharded parameters.
+            In Zero-2, set keep_unsharded to False.
+            In Zero-3, set keep_unsharded to True.
         max_scale (int, optional): max_scale used by DynamicGradScaler. Defaults to 2**32.
         dp_process_group (Optional[ProcessGroup], optional): data paralle process group. Defaults to None.
         mp_process_group (Optional[ProcessGroup], optional): model paralle process group. Defaults to None.
@@ -89,6 +91,7 @@ class ShardedOptimizerV2(ColossalaiOptimizer):
                  growth_interval: float = 1000,
                  hysteresis: float = 2,
                  max_scale: int = 2**32,
+                 keep_unsharded: bool = False,
                  dp_process_group: Optional[ProcessGroup] = None,
                  mp_process_group: Optional[ProcessGroup] = None) -> None:
         assert isinstance(sharded_model, ShardedModelV2), 'model must be wrapped with ShardedModel'
@@ -122,24 +125,12 @@ class ShardedOptimizerV2(ColossalaiOptimizer):
         self._found_overflow: Tensor = torch.FloatTensor([0]).to(torch.cuda.current_device())
         self._logger = get_dist_logger("ShardedOptimizerV2")
 
-        # Store fp32 param shards
+        assert not (keep_unsharded and self._should_move_fp32_shards_h2d), \
-        self.master_params: Dict[Parameter, StatefulTensor] = {}
+            "Keeping unsharded parameters can't be used with hybrid OS placement right now."
+        self.keep_unshard = keep_unsharded
 
-        for group in self.optim.param_groups:
+        # Store fp32 param shards
-            for p in group['params']:
+        self._register_master_weight()
-                assert hasattr(p, 'colo_attr'), 'The parameter must be wrapped with ShardedParam'
-                is_param_sharded = p.colo_attr.sharded_data_tensor.is_sharded
-                if not is_param_sharded:
-                    # TODO (ver217): we may not use shard / gather here
-                    # Param is no sharded, which means we use ZeRO-2 here
-                    # As we only store param shard, we shard it here
-                    self.shard_strategy.shard([p.colo_attr.sharded_data_tensor], self.dp_process_group)
-                self.master_params[p] = StatefulTensor(
-                    cast_tensor_to_fp32(p.colo_attr.sharded_data_tensor.payload).to(self.device))
-                if not is_param_sharded:
-                    # In this branch, there's no need to shard param
-                    # So we gather here
-                    self.shard_strategy.gather([p.colo_attr.sharded_data_tensor], self.dp_process_group)
 
         self._logger.debug(f"After init ShardedOptimizerV2 consumes {self.get_memory_usage()[0]/1e6} MB CUDA Memory!",
                            ranks=[0])
@@ -283,6 +274,23 @@ class ShardedOptimizerV2(ColossalaiOptimizer):
     def sync_grad(self):
         pass
 
+    def _register_master_weight(self):
+        self.master_params: Dict[Parameter, StatefulTensor] = {}
+        for group in self.optim.param_groups:
+            for p in group['params']:
+                assert hasattr(p, 'colo_attr'), 'The parameter must be wrapped with ShardedParam'
+                is_param_sharded = p.colo_attr.sharded_data_tensor.is_sharded
+                if not is_param_sharded and not self.keep_unshard:
+                    # Please use keep_unsharded to control whether shard unsharded paramters
+                    # As we only store param shard, we shard it here
+                    self.shard_strategy.shard([p.colo_attr.sharded_data_tensor], self.dp_process_group)
+                self.master_params[p] = StatefulTensor(
+                    cast_tensor_to_fp32(p.colo_attr.sharded_data_tensor.payload).to(self.device))
+                if not is_param_sharded and not self.keep_unshard:
+                    # In this branch, there's no need to shard param
+                    # So we gather here
+                    self.shard_strategy.gather([p.colo_attr.sharded_data_tensor], self.dp_process_group)
+
     def _maybe_move_fp32_shards(self):
         if self._should_move_fp32_shards_h2d:
             self._should_move_fp32_shards_h2d = False
@@ -328,7 +336,7 @@ class ShardedOptimizerV2(ColossalaiOptimizer):
         for group in self.optim.param_groups:
             for p in group['params']:
                 is_param_sharded = p.colo_attr.sharded_data_tensor.is_sharded
-                if not is_param_sharded:
+                if not is_param_sharded and not self.keep_unshard:
                     # We use ZeRO-2 here
                     # The `p.colo_attr.sharded_data_tensor` saves full fp16 param
                     # But we only have updated fp32 param shard here
@@ -342,7 +350,7 @@ class ShardedOptimizerV2(ColossalaiOptimizer):
                 p.colo_attr.sharded_data_tensor.reset_payload(
                     colo_model_tensor_clone(p.half(), torch.cuda.current_device()))
 
-                if not is_param_sharded:
+                if not is_param_sharded and not self.keep_unshard:
                     # We gather full fp16 param here
                     self.shard_strategy.gather([p.colo_attr.sharded_data_tensor], self.dp_process_group)
                 p.data = p.colo_attr.sharded_data_tensor.payload

tests/components_to_test/no_leaf_module.py

@@ -42,4 +42,5 @@ def get_training_components():
     testloader = DummyDataLoader()
 
     criterion = torch.nn.CrossEntropyLoss()
-    return model_builder, trainloader, testloader, torch.optim.Adam, criterion
+    from colossalai.nn.optimizer import HybridAdam
+    return model_builder, trainloader, testloader, HybridAdam, criterion

tests/test_moe/test_moe_zero_init.py

@@ -76,8 +76,11 @@ def run_moe_zero_init(init_device_type, shard_strategy_class):
             else:
                 assert param.is_replicated
 
-            assert param.colo_attr.sharded_data_tensor.payload.device.type == init_device.type, \
+            if param.colo_attr.param_is_sharded:
-                f'{param.colo_attr.sharded_data_tensor.payload.device.type} vs. {init_device.type}'
+                assert param.colo_attr.sharded_data_tensor.payload.device.type == init_device.type, \
+                    f'{param.colo_attr.sharded_data_tensor.payload.device.type} vs. {init_device.type}'
+            else:
+                assert param.colo_attr.sharded_data_tensor.payload.device.type == 'cuda'
 
 
 def _run_dist(rank, world_size, port):

tests/test_moe/test_moe_zero_model.py

@@ -67,7 +67,7 @@ def run_dist(rank, world_size, port):
 
 
 @pytest.mark.dist
-@pytest.mark.parametrize("world_size", [1, 2])
+@pytest.mark.parametrize("world_size", [2])
 @rerun_on_exception(exception_type=mp.ProcessRaisedException, pattern=".*Address already in use.*")
 def test_moe_zero_model(world_size):
     run_func = partial(run_dist, world_size=world_size, port=free_port())

tests/test_moe/test_moe_zero_optim.py

@@ -0,0 +1,134 @@
+from functools import partial
+
+import colossalai
+from colossalai.utils.cuda import get_current_device
+import pytest
+import torch
+import torch.multiprocessing as mp
+from colossalai.amp import convert_to_apex_amp
+from colossalai.nn.optimizer import CPUAdam
+from colossalai.testing import parameterize, rerun_on_exception
+from colossalai.utils import free_port
+from colossalai.zero.init_ctx import ZeroInitContext
+from colossalai.zero.shard_utils import (BucketTensorShardStrategy, TensorShardStrategy)
+from colossalai.zero.sharded_model import ShardedModelV2
+from colossalai.zero.sharded_model.utils import col_model_deepcopy
+from colossalai.zero.sharded_optim import ShardedOptimizerV2
+from colossalai.zero.sharded_optim._utils import has_inf_or_nan
+from colossalai.utils import get_current_device
+from tests.components_to_test.registry import non_distributed_component_funcs
+from colossalai.engine.gradient_handler import MoeGradientHandler
+from colossalai.context import MOE_CONTEXT
+from colossalai.testing import assert_equal_in_group
+
+from tests.test_zero_data_parallel.common import CONFIG, check_sharded_model_params
+from tests.test_moe.test_moe_zero_init import MoeModel
+
+
+def _run_step(model, optimizer, data, label, criterion, grad_handler):
+    model.train()
+    optimizer.zero_grad()
+
+    if criterion:
+        y = model(data)
+        loss = criterion(y, label)
+    else:
+        loss = model(data, label)
+
+    loss = loss.float()
+    if isinstance(model, ShardedModelV2):
+        optimizer.backward(loss)
+    else:
+        loss.backward()
+
+    if grad_handler is not None:
+        grad_handler.handle_gradient()
+
+    optimizer.step()
+
+
+@parameterize("cpu_offload", [True, False])
+@parameterize("use_cpuadam", [True, False])
+@parameterize("shard_strategy_class", [TensorShardStrategy, BucketTensorShardStrategy])
+def _run_test_sharded_optim_v2(cpu_offload, shard_strategy_class, use_cpuadam, gpu_margin_mem_ratio=0.0):
+    MOE_CONTEXT.reset_loss()
+    shard_strategy = shard_strategy_class()
+    if use_cpuadam and cpu_offload is False:
+        return
+
+    get_components_func = non_distributed_component_funcs.get_callable('no_leaf_module')
+    _, train_dataloader, _, optimizer_class, criterion = get_components_func()
+
+    with ZeroInitContext(
+            target_device=torch.device('cpu') if cpu_offload else torch.device(f'cuda:{get_current_device()}'),
+            shard_strategy=shard_strategy,
+            shard_param=True,
+            rm_torch_payload_on_the_fly=False):
+        zero_model = MoeModel()
+
+    zero_model = ShardedModelV2(
+        zero_model,
+        shard_strategy,
+        offload_config=dict(device='cpu') if cpu_offload else None,
+        use_memory_tracer=gpu_margin_mem_ratio > 0.0,
+        reuse_fp16_shard=use_cpuadam,
+    )
+
+    # check whether parameters are identical in ddp
+    for name, p in zero_model.named_parameters():
+        if not p.colo_attr.param_is_sharded and p.is_replicated:
+            assert_equal_in_group(p.data.to(get_current_device()))
+
+    model = MoeModel().half()
+    col_model_deepcopy(zero_model, model)
+    model = model.cuda().float()
+
+    if use_cpuadam:
+        optimizer_class = CPUAdam
+    optim = optimizer_class(model.parameters(), lr=1e-3)
+    sharded_optim = optimizer_class(zero_model.parameters(), lr=1e-3)
+    sharded_optim = ShardedOptimizerV2(zero_model,
+                                       sharded_optim,
+                                       cpu_offload=cpu_offload,
+                                       initial_scale=2**5,
+                                       gpu_margin_mem_ratio=gpu_margin_mem_ratio,
+                                       keep_unsharded=True)
+
+    amp_config = dict(opt_level='O2', keep_batchnorm_fp32=False)
+    apex_model, apex_optimizer = convert_to_apex_amp(model, optim, amp_config)
+    apex_grad_handler = MoeGradientHandler(model)
+
+    # Since MOE is not compatible with apex_amp now, we need to convert gate weight to fp32
+    for (n, p), zp in zip(apex_model.named_parameters(), zero_model.parameters()):
+        if 'gate' in n:
+            p.data = p.float()
+            p.data.copy_(zp.data)
+
+    for i, (data, label) in enumerate(train_dataloader):
+        if i > 5:
+            break
+        data, label = data.cuda(), label.cuda()
+        _run_step(apex_model, apex_optimizer, data, label, criterion, apex_grad_handler)
+        _run_step(zero_model, sharded_optim, data, label, criterion, None)
+        check_sharded_model_params(model, zero_model, loose=True, reuse_fp16_shard=use_cpuadam)
+        for param in model.parameters():
+            assert not has_inf_or_nan(param)
+
+
+def _run_dist(rank, world_size, port):
+    colossalai.launch(config=CONFIG, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
+    MOE_CONTEXT.setup(seed=42)
+    _run_test_sharded_optim_v2()
+
+
+# use_cpuadam = True can be used with cpu_offload = False
+@pytest.mark.dist
+@pytest.mark.parametrize("world_size", [2])
+@rerun_on_exception(exception_type=mp.ProcessRaisedException, pattern=".*Address already in use.*")
+def test_moe_zero_optim(world_size):
+    run_func = partial(_run_dist, world_size=world_size, port=free_port())
+    mp.spawn(run_func, nprocs=world_size)
+
+
+if __name__ == '__main__':
+    test_moe_zero_optim(world_size=2)

tests/test_zero_data_parallel/common.py

@@ -124,16 +124,18 @@ def check_params_padding(model, zero_model, loose=False):
 
 def check_sharded_model_params(model, zero_model, loose=False, reuse_fp16_shard=False):
     rank = dist.get_rank()
-    for p, zero_p in zip(model.parameters(), zero_model.parameters()):
+    for (name, p), (zero_name, zero_p) in zip(model.named_parameters(), zero_model.named_parameters()):
-        if reuse_fp16_shard:
+        if zero_p.colo_attr.param_is_sharded:
-            zero_p = zero_p.data.to(p.device).float()
+            if reuse_fp16_shard:
-        else:
+                zero_p = zero_p.data.to(p.device).float()
-            zero_p = zero_p.colo_attr.sharded_data_tensor.payload.to(p.device).float()
+            else:
-        chunks = torch.flatten(p).chunk(dist.get_world_size())
+                zero_p = zero_p.colo_attr.sharded_data_tensor.payload.to(p.device).float()
-        if rank >= len(chunks):
+            chunks = torch.flatten(p).chunk(dist.get_world_size())
-            continue
+            if rank >= len(chunks):
-        p = chunks[rank].float()
+                continue
-        if zero_p.size(0) > p.size(0):
+            p = chunks[rank].float()
-            zero_p = zero_p[:p.size(0)]
+            if zero_p.size(0) > p.size(0):
+                zero_p = zero_p[:p.size(0)]
+
         assert p.dtype == zero_p.dtype
         assert allclose(p, zero_p, loose=loose), f'{p} vs {zero_p}'