[zero] adapt zero for unsharded parameters (#561)

* support existing sharded and unsharded parameters in zero

* add unitest for moe-zero model init

* polish moe gradient handler

colossalai/engine/gradient_handler/_moe_gradient_handler.py

@@ -16,6 +16,9 @@ class MoeGradientHandler(BaseGradientHandler):
     the same type to improve the efficiency of communication.
     """
 
+    def __init__(self, model, optimizer=None):
+        super().__init__(model, optimizer)
+
     def handle_gradient(self):
         """A method running an all-reduce operation in a data parallel group.
         Then running an all-reduce operation for all parameters in experts
@@ -24,13 +27,15 @@ class MoeGradientHandler(BaseGradientHandler):
         global_data = gpc.data_parallel_size
 
         if global_data > 1:
-            param_dict = get_moe_epsize_param_dict(self._model)
+            epsize_param_dict = get_moe_epsize_param_dict(self._model)
 
+            # epsize is 1, indicating the params are replicated among processes in data parallelism
+            # use the ParallelMode.DATA to get data parallel group
             # reduce gradients for all parameters in data parallelism
-            if 1 in param_dict:
-                bucket_allreduce(param_list=param_dict[1], group=gpc.get_group(ParallelMode.DATA))
+            if 1 in epsize_param_dict:
+                bucket_allreduce(param_list=epsize_param_dict[1], group=gpc.get_group(ParallelMode.DATA))
 
-            for ep_size in param_dict:
+            for ep_size in epsize_param_dict:
                 if ep_size != 1 and ep_size != MOE_CONTEXT.world_size:
-                    bucket_allreduce(param_list=param_dict[ep_size],
+                    bucket_allreduce(param_list=epsize_param_dict[ep_size],
                                      group=MOE_CONTEXT.parallel_info_dict[ep_size].dp_group)

colossalai/engine/ophooks/__init__.py

@@ -1,4 +1,4 @@
-from typing import List
+from typing import List, Callable, Optional
 
 import torch
 
@@ -61,18 +61,25 @@ class PostBackwardFunction(torch.autograd.Function):
         return (None, None) + args
 
 
-def register_ophooks_recursively(module: torch.nn.Module, ophook_list: List[BaseOpHook] = None, name: str = ""):
+def register_ophooks_recursively(module: torch.nn.Module,
+                                 ophook_list: List[BaseOpHook] = None,
+                                 name: str = "",
+                                 filter_fn: Optional[Callable] = None):
     r"""Recursilvely register pre/post hooks for all submodules in the module in FWD and BWD."""
     assert isinstance(module, torch.nn.Module)
 
     # Add hooks for submodules
     for child_name, child in module.named_children():
-        register_ophooks_recursively(child, ophook_list, name + child_name)
+        register_ophooks_recursively(child, ophook_list, name + child_name, filter_fn)
 
     # Early return on modules with no parameters.
     if len(list(module.parameters(recurse=False))) == 0:
         return
 
+    # return from flitered module
+    if filter_fn is not None and filter_fn(module):
+        return
+
     if ophook_list is not None:
         for hook in ophook_list:
             assert (isinstance(hook, BaseOpHook))

colossalai/nn/layer/moe/experts.py

@@ -35,7 +35,7 @@ class Experts(MoeExperts):
         expert_args: Args used to initialize experts, the args could be found in corresponding expert class
     """
 
-    @no_shard_zero_decrator
+    @no_shard_zero_decrator(is_replicated=False)
     def __init__(self, expert_cls: Type[nn.Module], num_experts: int, **expert_args):
         super().__init__("all_to_all", num_experts)
 

colossalai/nn/layer/moe/layers.py

@@ -9,7 +9,7 @@ from colossalai.context.moe_context import MOE_CONTEXT
 from colossalai.utils import get_current_device
 from ._operation import COL_MOE_KERNEL_FLAG, AllToAll, AllGather, ReduceScatter, MoeDispatch, MoeCombine, moe_cumsum
 from .experts import MoeExperts, Experts
-from .utils import ForceFP32Parameter, UniformNoiseGenerator, NormalNoiseGenerator
+from .utils import ForceFP32Parameter, UniformNoiseGenerator, NormalNoiseGenerator, autocast_softmax
 from colossalai.zero.init_ctx import no_shard_zero_context, no_shard_zero_decrator
 from typing import Callable, Optional, Type
 from torch.distributed import ProcessGroup
@@ -66,7 +66,7 @@ class Top1Router(nn.Module):
         if self.noisy_func is not None and self.training:
             inputs = self.noisy_func(inputs)
 
-        logits = F.softmax(inputs, dim=-1)
+        logits = autocast_softmax(inputs, dim=-1)
         num_experts = logits.size(-1)
         capacity = self.get_capacity(logits.shape)
 
@@ -152,7 +152,7 @@ class Top2Router(nn.Module):
         if self.noisy_func is not None and self.training:
             inputs = self.noisy_func(inputs)
 
-        logits = F.softmax(inputs, dim=-1)    # logits: [s, e]
+        logits = autocast_softmax(inputs, dim=-1)    # logits: [s, e]
         num_experts = logits.size(-1)
         capacity = self.get_capacity(logits.shape)
 
@@ -241,7 +241,7 @@ class MoeLayer(nn.Module):
         experts (:class:`torch.nn.Module`): Instance of experts generated by Expert.
     """
 
-    @no_shard_zero_decrator
+    @no_shard_zero_decrator(is_replicated=True)
     def __init__(self, dim_model: int, num_experts: int, router: nn.Module, experts: MoeExperts):
         super().__init__()
         self.d_model = dim_model

colossalai/nn/layer/moe/utils.py

@@ -1,4 +1,5 @@
 import torch
+import torch.nn.functional as F
 from colossalai.utils import get_current_device
 from colossalai.context.moe_context import MOE_CONTEXT
 from .experts import FFNExperts, TPExperts
@@ -51,6 +52,12 @@ class UniformNoiseGenerator:
         return inputs * noisy
 
 
+def autocast_softmax(logit: torch.Tensor, dim: int):
+    if logit.dtype != torch.float32:
+        logit = logit.float()
+    return F.softmax(logit, dim=dim)
+
+
 def build_ffn_experts(num_experts: int, d_model: int, d_ff: int, activation=None, drop_rate: float = 0):
     mep_size = MOE_CONTEXT.max_ep_size
     if num_experts % mep_size == 0 or mep_size % num_experts == 0:

colossalai/utils/memory_tracer/model_data_memtracer.py

@@ -39,7 +39,7 @@ def colo_model_mem_usage(model: torch.nn.Module) -> Tuple[int, int]:
         if t.device.type == 'cpu':
             _cpu_mem_usage += t.numel() * t.element_size()
         elif t.device.type == 'cuda':
-            _cuda_mem_usages += t.numel() * t.element_size()
+            _cuda_mem_usage += t.numel() * t.element_size()
         return _cuda_mem_usage, _cpu_mem_usage
 
     cuda_mem_usage = 0

colossalai/zero/init_ctx/init_context.py

@@ -88,6 +88,8 @@ class ZeroContextConfig(object):
     """The configuration used to control zero context initialization.
 
     Args:
+        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.
         rm_torch_payload_on_the_fly (bool, optional): If set to `True`, remove tensor payload on `param.data` after module init finished.
             This will reduce memory usage when initializing model.
@@ -97,8 +99,9 @@ class ZeroContextConfig(object):
             See torchvision resnet18. Defaults to False.
     """
 
-    def __init__(self, shard_param: bool = False, rm_torch_payload_on_the_fly: bool = False):
+    def __init__(self, replicated: bool = True, shard_param: bool = False, rm_torch_payload_on_the_fly: bool = False):
         super().__init__()
+        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
 
@@ -139,10 +142,15 @@ class ZeroInitContext(InsertPostInitMethodToModuleSubClasses):
         self.model_numel_tensor = model_numel_tensor
         self.dp_process_group = dp_process_group or gpc.get_group(ParallelMode.DATA)
 
-        self.config = ZeroContextConfig(shard_param=shard_param,
+        self.config = ZeroContextConfig(replicated=True,
+                                        shard_param=shard_param,
                                         rm_torch_payload_on_the_fly=rm_torch_payload_on_the_fly)
         ZeroContextMgr().current_context = self
 
+    @property
+    def is_replicated(self):
+        return self.config.is_replicated
+
     @property
     def shard_param(self):
         return self.config.shard_param
@@ -183,6 +191,9 @@ class ZeroInitContext(InsertPostInitMethodToModuleSubClasses):
 
             self.model_numel_tensor += param.numel()
 
+            # mark whether the param is replicated
+            param.is_replicated = self.is_replicated
+
             # convert parameters to half
             param_half = half_fn(param)
             param.data = param_half
@@ -224,14 +235,20 @@ class ZeroContextMgr(metaclass=SingletonMeta):
             self.current_context.config = old_config
 
 
-def no_shard_zero_context():
-    return ZeroContextMgr().hijack_context_config(shard_param=False, rm_torch_payload_on_the_fly=False)
+def no_shard_zero_context(is_replicated: bool = True):
+    return ZeroContextMgr().hijack_context_config(replicated=is_replicated,
+                                                  shard_param=False,
+                                                  rm_torch_payload_on_the_fly=False)
+
+
+def no_shard_zero_decrator(is_replicated: bool = True):
 
+    def _wrapper(init_func):
 
-def no_shard_zero_decrator(init_func):
+        def _no_shard(*args, **kwargs):
+            with no_shard_zero_context(is_replicated):
+                init_func(*args, **kwargs)
 
-    def _no_shard(*args, **kwargs):
-        with no_shard_zero_context():
-            init_func(*args, **kwargs)
+        return _no_shard
 
-    return _no_shard
+    return _wrapper

colossalai/zero/sharded_model/sharded_model_v2.py

@@ -10,6 +10,7 @@ from colossalai.core import global_context as gpc
 from colossalai.engine.ophooks import register_ophooks_recursively
 from colossalai.engine.ophooks.zero_hook import ZeroHook
 from colossalai.engine.paramhooks import BaseParamHookMgr
+from colossalai.engine.gradient_handler.utils import bucket_allreduce
 from colossalai.logging import get_dist_logger
 from colossalai.utils import get_current_device
 from colossalai.utils.memory_tracer.memstats_collector import MemStatsCollector
@@ -67,17 +68,27 @@ class ShardedModelV2(nn.Module):
         self.logger = get_dist_logger()
 
         # We force users to use ZeroInitContext
-        sharded = []
-        unsharded = []
+        for submodule in module.modules():
+            sharded_cnt = 0
+            unshard_cnt = 0
+            for param in submodule.parameters(recurse=False):
+                assert hasattr(param, 'colo_attr'), 'You must use ZeroInitContext to init your module first.'
+                if param.colo_attr.param_is_sharded:
+                    sharded_cnt += 1
+                else:
+                    unshard_cnt += 1
+            assert (not sharded_cnt) or (not unshard_cnt), 'nn.Module can not both have shard param and unshard param'
+            submodule.param_is_sharded = (sharded_cnt > 0)
+
+        self.sharded_params = []
+        self.unshard_params = []
         for param in module.parameters():
-            assert hasattr(param, 'colo_attr'), 'You must use ZeroInitContext to init your module first.'
-            sharded.append(param.colo_attr.param_is_sharded)
-            unsharded.append(not param.colo_attr.param_is_sharded)
-        assert all(sharded) or all(
-            unsharded), 'Parameters must be all sharded or all unsharded! Parameters are partially sharded now.'
-        self.shard_param = all(sharded)
-        self.module = module
+            if param.colo_attr.param_is_sharded:
+                self.sharded_params.append(param)
+            else:
+                self.unshard_params.append(param)
 
+        self.module = module
         self.process_group = process_group or gpc.get_group(ParallelMode.DATA)
         self.reduce_scatter_process_group = reduce_scatter_process_group or self.process_group
         self.world_size = dist.get_world_size(self.process_group)
@@ -95,8 +106,8 @@ class ShardedModelV2(nn.Module):
 
         # Register hooks
         self._ophook_list = [ZeroHook(self.shard_strategy, self._memstats_collector, self.process_group)]
-        register_ophooks_recursively(self.module, self._ophook_list)
-        self.param_hook_mgr = BaseParamHookMgr(list(self.module.parameters()))
+        register_ophooks_recursively(self.module, self._ophook_list, filter_fn=lambda m: not m.param_is_sharded)
+        self.param_hook_mgr = BaseParamHookMgr(self.sharded_params)
         self.param_hook_mgr.register_backward_hooks(self._grad_post_backward_hook)
 
         self.fp32_reduce_scatter = fp32_reduce_scatter
@@ -185,7 +196,6 @@ class ShardedModelV2(nn.Module):
 
     def backward_by_grad(self, tensor, grad):
         torch.autograd.backward(tensors=tensor, grad_tensors=grad)
-
         self._post_backward_operations()
         for ophook in self._ophook_list:
             ophook.post_iter()
@@ -224,17 +234,21 @@ class ShardedModelV2(nn.Module):
                 # Wait for the non-blocking GPU -> CPU grad transfers to finish.
                 torch.cuda.current_stream().synchronize()
         self.reducer.free()
+
+        # all reduce gradients for unsharded parameters
+        reduce_list = [p for p in self.unshard_params if p.is_replicated]
+        bucket_allreduce(reduce_list, self.process_group)
+
         # 3. shard tensors not dealed in the zero hook
-        if self.shard_param:
-            tensor_list = []
-            for p in self.module.parameters():
-                if not p.colo_attr.param_is_sharded:
-                    tensor_list.append(p.colo_attr.sharded_data_tensor)
-                    p.colo_attr.sharded_data_tensor.trans_state(TensorState.HOLD_AFTER_BWD)
-                    p.colo_attr.remove_torch_payload()
-            self.shard_strategy.shard(tensor_list, self.process_group)
-
-        # 4. move sharded param grad payload to param.grad
+        tensor_list = []
+        for p in self.sharded_params:
+            if not p.colo_attr.param_is_sharded:
+                tensor_list.append(p.colo_attr.sharded_data_tensor)
+                p.colo_attr.sharded_data_tensor.trans_state(TensorState.HOLD_AFTER_BWD)
+                p.colo_attr.remove_torch_payload()
+        self.shard_strategy.shard(tensor_list, self.process_group)
+
+        # 4. set all parameters' grad to None
         for p in self.module.parameters():
             if not p.requires_grad:
                 continue
@@ -245,6 +259,16 @@ class ShardedModelV2(nn.Module):
             # We also allows to interleave no-sync pass with sync passes, if desired.
             if not self._require_backward_grad_sync:
                 continue
+
+            # move unsharded param grad to saved_grad
+            if not p.colo_attr.param_is_sharded:
+                if p.colo_attr.offload_grad:
+                    colo_model_data_move_to_cpu(p.grad)
+                if p.colo_attr.saved_grad.is_null():
+                    p.colo_attr.saved_grad.reset_payload(p.grad.data)
+                else:
+                    p.colo_attr.saved_grad.payload.add_(p.grad.data)
+
             p.grad = None
 
     @torch.no_grad()
@@ -320,16 +344,14 @@ class ShardedModelV2(nn.Module):
         param.colo_attr.saved_grad.trans_state(TensorState.HOLD)
 
     def state_dict(self, destination=None, prefix='', keep_vars=False) -> 'OrderedDict[str, torch.Tensor]':
-        self.shard_strategy.gather([p.colo_attr.sharded_data_tensor for p in self.module.parameters()],
-                                   self.process_group)
+        self.shard_strategy.gather([p.colo_attr.sharded_data_tensor for p in self.sharded_params], self.process_group)
         prev_params = {}
-        for p in self.module.parameters():
+        for p in self.sharded_params:
             prev_params[p] = p.data
             p.data = p.colo_attr.sharded_data_tensor.payload
         gathered_state_dict = self.module.state_dict(destination, prefix, keep_vars)
-        self.shard_strategy.shard([p.colo_attr.sharded_data_tensor for p in self.module.parameters()],
-                                  self.process_group)
-        for p in self.module.parameters():
+        self.shard_strategy.shard([p.colo_attr.sharded_data_tensor for p in self.sharded_params], self.process_group)
+        for p in self.sharded_params:
             p.data = prev_params[p]
         return gathered_state_dict
 

tests/test_moe/test_moe_zero_init.py

@@ -22,19 +22,14 @@ class MoeModel(nn.Module):
 
     def __init__(self):
         super().__init__()
-        self.proj1 = nn.Linear(4, 8)
+        self.proj1 = nn.Linear(4, 16)
         expert_cls = nn.Linear
-        expert_args_dict = dict(in_features=8, out_features=8)
-        self.moe = MoeModule(dim_model=8,
-                             num_experts=8,
-                             noisy_policy='Jitter',
-                             use_residual=True,
-                             expert_cls=expert_cls,
-                             **expert_args_dict)
-        self.proj2 = nn.Linear(8, 4)
+        expert_args_dict = dict(in_features=16, out_features=16)
+        self.moe = MoeModule(dim_model=16, num_experts=8, use_residual=True, expert_cls=expert_cls, **expert_args_dict)
+        self.proj2 = nn.Linear(16, 4)
 
     def forward(self, x):
-        x = self.proj(x)
+        x = self.proj1(x)
         x = self.moe(x)
         x = self.proj2(x)
         return x
@@ -75,6 +70,12 @@ def run_moe_zero_init(init_device_type, shard_strategy_class):
             else:
                 assert param.colo_attr.sharded_data_tensor.is_sharded
 
+            # the parameters in moe experts is not replicated
+            if 'experts' in name:
+                assert not param.is_replicated
+            else:
+                assert param.is_replicated
+
             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}'
 

tests/test_moe/test_moe_zero_model.py

@@ -0,0 +1,78 @@
+from functools import partial
+
+import colossalai
+import pytest
+import torch
+import torch.multiprocessing as mp
+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 cast_tensor_to_fp16
+from colossalai.zero.sharded_model.utils import col_model_deepcopy
+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_grads_padding, run_fwd_bwd
+from tests.test_moe.test_moe_zero_init import MoeModel
+
+
+@parameterize("enable_autocast", [False])
+@parameterize("shard_strategy_class", [TensorShardStrategy, BucketTensorShardStrategy])
+def run_model_test(enable_autocast, shard_strategy_class):
+    shard_strategy = shard_strategy_class()
+
+    get_components_func = non_distributed_component_funcs.get_callable('no_leaf_module')
+    _, train_dataloader, _, _, criterion = get_components_func()
+
+    rm_torch_payload_on_the_fly = False
+
+    with ZeroInitContext(target_device=torch.cuda.current_device(),
+                         shard_strategy=shard_strategy,
+                         shard_param=True,
+                         rm_torch_payload_on_the_fly=rm_torch_payload_on_the_fly):
+        zero_model = MoeModel()
+    zero_model = ShardedModelV2(zero_model, shard_strategy, use_memory_tracer=True)
+
+    # 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)
+
+    model = MoeModel().half()
+    col_model_deepcopy(zero_model, model)
+    model = model.cuda()
+    grad_handler = MoeGradientHandler(model)
+
+    for i, (data, label) in enumerate(train_dataloader):
+        if i > 5:
+            break
+
+        data, label = cast_tensor_to_fp16(data).cuda(), label.cuda()
+        run_fwd_bwd(model, data, label, criterion, enable_autocast)
+        run_fwd_bwd(zero_model, data, label, criterion, enable_autocast)
+        grad_handler.handle_gradient()
+
+        check_grads_padding(model, zero_model, loose=True)
+
+
+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)
+    MOE_CONTEXT.reset_loss()
+    run_model_test()
+
+
+@pytest.mark.dist
+@pytest.mark.parametrize("world_size", [1, 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())
+    mp.spawn(run_func, nprocs=world_size)
+
+
+if __name__ == '__main__':
+    test_moe_zero_model(world_size=2)

tests/test_zero_data_parallel/common.py

@@ -91,15 +91,19 @@ def check_params(model, zero_model, loose=False):
 
 def check_grads_padding(model, zero_model, loose=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()):
         # zero_grad = zero_p.grad.clone().to(p.device)
-        zero_grad = zero_p.colo_attr.saved_grad.payload.clone().to(p.device)
-        chunks = torch.flatten(p.grad).chunk(dist.get_world_size())
-        if rank >= len(chunks):
-            continue
-        grad = chunks[rank].float()
-        if zero_grad.size(0) > grad.size(0):
-            zero_grad = zero_grad[:grad.size(0)]
+        if zero_p.colo_attr.param_is_sharded:
+            zero_grad = zero_p.colo_attr.saved_grad.payload.clone().to(p.device)
+            chunks = torch.flatten(p.grad).chunk(dist.get_world_size())
+            if rank >= len(chunks):
+                continue
+            grad = chunks[rank].float()
+            if zero_grad.size(0) > grad.size(0):
+                zero_grad = zero_grad[:grad.size(0)]
+        else:
+            grad = p.grad
+            zero_grad = zero_p.colo_attr.saved_grad.payload
         assert grad.dtype == zero_grad.dtype
         assert allclose(grad, zero_grad, loose=loose), f'diff: {grad - zero_grad}'