[legacy] move builder and registry to legacy (hpcaitech#4603)

colossalai/auto_parallel/tensor_shard/node_handler/registry.py

@@ -1,5 +1,5 @@
 class Registry:
-    # TODO: refactor the registry classes used in colossalai.registry, colossalai.fx and here
+    # TODO: refactor the registry classes used in colossalai.legacy.registry, colossalai.fx and here
 
     def __init__(self, name):
         self.name = name

colossalai/context/parallel_context.py

@@ -15,8 +15,8 @@
 from colossalai.context.config import Config
 from colossalai.context.singleton_meta import SingletonMeta
 from colossalai.global_variables import tensor_parallel_env as env
+from colossalai.legacy.registry import DIST_GROUP_INITIALIZER
 from colossalai.logging import get_dist_logger
-from colossalai.registry import DIST_GROUP_INITIALIZER
 
 from .parallel_mode import ParallelMode
 from .random import add_seed, get_seeds, set_mode

colossalai/context/process_group_initializer/initializer_1d.py

@@ -2,8 +2,9 @@
 # -*- encoding: utf-8 -*-
 
 import torch.distributed as dist
+
 from colossalai.global_variables import tensor_parallel_env as env
-from colossalai.registry import DIST_GROUP_INITIALIZER
+from colossalai.legacy.registry import DIST_GROUP_INITIALIZER
 
 from ..parallel_mode import ParallelMode
 from .process_group_initializer import ProcessGroupInitializer

colossalai/context/process_group_initializer/initializer_2d.py

@@ -3,7 +3,7 @@
 import torch.distributed as dist
 
 from colossalai.global_variables import tensor_parallel_env as env
-from colossalai.registry import DIST_GROUP_INITIALIZER
+from colossalai.legacy.registry import DIST_GROUP_INITIALIZER
 
 from ..parallel_mode import ParallelMode
 from .process_group_initializer import ProcessGroupInitializer

colossalai/context/process_group_initializer/initializer_2p5d.py

@@ -4,9 +4,10 @@
 import math
 
 import torch.distributed as dist
+
 from colossalai.context import Config
 from colossalai.global_variables import tensor_parallel_env as env
-from colossalai.registry import DIST_GROUP_INITIALIZER
+from colossalai.legacy.registry import DIST_GROUP_INITIALIZER
 
 from ..parallel_mode import ParallelMode
 from .process_group_initializer import ProcessGroupInitializer

colossalai/context/process_group_initializer/initializer_3d.py

@@ -6,7 +6,7 @@
 import torch.distributed as dist
 
 from colossalai.global_variables import tensor_parallel_env as env
-from colossalai.registry import DIST_GROUP_INITIALIZER
+from colossalai.legacy.registry import DIST_GROUP_INITIALIZER
 
 from ..parallel_mode import ParallelMode
 from .process_group_initializer import ProcessGroupInitializer

colossalai/context/process_group_initializer/initializer_data.py

@@ -3,7 +3,7 @@
 
 from torch import distributed as dist
 
-from colossalai.registry import DIST_GROUP_INITIALIZER
+from colossalai.legacy.registry import DIST_GROUP_INITIALIZER
 
 from ..parallel_mode import ParallelMode
 from .process_group_initializer import ProcessGroupInitializer

colossalai/context/process_group_initializer/initializer_model.py

@@ -2,9 +2,11 @@
 # -*- encoding: utf-8 -*-
 
 import torch.distributed as dist
-from colossalai.registry import DIST_GROUP_INITIALIZER
-from .process_group_initializer import ProcessGroupInitializer
+
+from colossalai.legacy.registry import DIST_GROUP_INITIALIZER
+
 from ..parallel_mode import ParallelMode
+from .process_group_initializer import ProcessGroupInitializer
 
 
 @DIST_GROUP_INITIALIZER.register_module

colossalai/context/process_group_initializer/initializer_pipeline.py

@@ -3,7 +3,7 @@
 
 from torch import distributed as dist
 
-from colossalai.registry import DIST_GROUP_INITIALIZER
+from colossalai.legacy.registry import DIST_GROUP_INITIALIZER
 
 from ..parallel_mode import ParallelMode
 from .process_group_initializer import ProcessGroupInitializer

colossalai/context/process_group_initializer/initializer_sequence.py

@@ -2,7 +2,7 @@
 # -*- encoding: utf-8 -*-
 import torch.distributed as dist
 
-from colossalai.registry import DIST_GROUP_INITIALIZER
+from colossalai.legacy.registry import DIST_GROUP_INITIALIZER
 
 from ..parallel_mode import ParallelMode
 from .initializer_tensor import Initializer_Tensor

colossalai/context/process_group_initializer/initializer_tensor.py

@@ -3,9 +3,10 @@
 
 import torch.distributed as dist
 
-from colossalai.registry import DIST_GROUP_INITIALIZER
-from .process_group_initializer import ProcessGroupInitializer
+from colossalai.legacy.registry import DIST_GROUP_INITIALIZER
+
 from ..parallel_mode import ParallelMode
+from .process_group_initializer import ProcessGroupInitializer
 
 
 @DIST_GROUP_INITIALIZER.register_module

colossalai/initialize.py

@@ -17,10 +17,10 @@
 
 from colossalai.amp import AMP_TYPE, convert_to_amp
 from colossalai.amp.naive_amp import NaiveAMPModel
-from colossalai.builder.builder import build_gradient_handler
 from colossalai.context import Config, ConfigException, ParallelMode
 from colossalai.context.moe_context import MOE_CONTEXT
 from colossalai.core import global_context as gpc
+from colossalai.legacy.builder.builder import build_gradient_handler
 from colossalai.legacy.engine import Engine
 from colossalai.legacy.engine.gradient_accumulation import accumulate_gradient
 from colossalai.legacy.engine.schedule import (

colossalai/builder/builder.py → colossalai/legacy/builder/builder.py

@@ -3,7 +3,7 @@
 
 import inspect
 
-from colossalai.registry import *
+from colossalai.legacy.registry import *
 
 
 def build_from_config(module, config: dict):

colossalai/legacy/engine/gradient_handler/_data_parallel_gradient_handler.py

@@ -1,6 +1,6 @@
 from colossalai.context.parallel_mode import ParallelMode
 from colossalai.core import global_context as gpc
-from colossalai.registry import GRADIENT_HANDLER
+from colossalai.legacy.registry import GRADIENT_HANDLER
 
 from ._base_gradient_handler import BaseGradientHandler
 from .utils import bucket_allreduce

colossalai/legacy/engine/gradient_handler/_moe_gradient_handler.py

@@ -1,7 +1,7 @@
 from colossalai.context.moe_context import MOE_CONTEXT
 from colossalai.context.parallel_mode import ParallelMode
 from colossalai.core import global_context as gpc
-from colossalai.registry import GRADIENT_HANDLER
+from colossalai.legacy.registry import GRADIENT_HANDLER
 from colossalai.utils.moe import get_moe_epsize_param_dict
 
 from ._base_gradient_handler import BaseGradientHandler

colossalai/legacy/engine/gradient_handler/_pipeline_parallel_gradient_handler.py

@@ -7,7 +7,7 @@
 from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors
 
 from colossalai.core import global_context as gpc
-from colossalai.registry import GRADIENT_HANDLER
+from colossalai.legacy.registry import GRADIENT_HANDLER
 
 from ._base_gradient_handler import BaseGradientHandler
 

colossalai/legacy/engine/gradient_handler/_sequence_parallel_gradient_handler.py

@@ -1,6 +1,6 @@
 from colossalai.context.parallel_mode import ParallelMode
 from colossalai.core import global_context as gpc
-from colossalai.registry import GRADIENT_HANDLER
+from colossalai.legacy.registry import GRADIENT_HANDLER
 
 from ._base_gradient_handler import BaseGradientHandler
 from .utils import bucket_allreduce

colossalai/legacy/engine/gradient_handler/_zero_gradient_handler.py

@@ -1,4 +1,4 @@
-from colossalai.registry import GRADIENT_HANDLER
+from colossalai.legacy.registry import GRADIENT_HANDLER
 
 from ._base_gradient_handler import BaseGradientHandler
 

colossalai/registry/registry.py → colossalai/legacy/registry/registry.py

@@ -6,7 +6,7 @@
 
 
 class Registry:
-    """This is a registry class used to register classes and modules so that a universal 
+    """This is a registry class used to register classes and modules so that a universal
     object builder can be enabled.
 
     Args:
@@ -42,7 +42,7 @@ def register_module(self, module_class):
         return module_class
 
     def get_module(self, module_name: str):
-        """Retrieves a module with name `module_name` and returns the module if it has 
+        """Retrieves a module with name `module_name` and returns the module if it has
         already been registered before.
 
         Args:

colossalai/legacy/trainer/hooks/_checkpoint_hook.py

@@ -2,9 +2,9 @@
 # -*- encoding: utf-8 -*-
 import torch
 
+from colossalai.legacy.registry import HOOKS
 from colossalai.legacy.trainer.hooks import BaseHook
 from colossalai.logging import get_dist_logger
-from colossalai.registry import HOOKS
 from colossalai.utils.checkpointing import save_checkpoint
 
 from ._lr_scheduler_hook import LRSchedulerHook

colossalai/legacy/trainer/hooks/_log_hook.py

@@ -7,9 +7,9 @@
 
 from colossalai.context import ParallelMode
 from colossalai.core import global_context as gpc
+from colossalai.legacy.registry import HOOKS
 from colossalai.legacy.trainer.hooks._metric_hook import ThroughputMetric
 from colossalai.logging import DistributedLogger
-from colossalai.registry import HOOKS
 from colossalai.utils import MultiTimer, is_dp_rank_0, is_no_pp_or_last_stage, is_tp_rank_0, report_memory_usage
 
 from ._base_hook import BaseHook

colossalai/legacy/trainer/hooks/_lr_scheduler_hook.py

@@ -1,6 +1,6 @@
 from torch import Tensor
 
-from colossalai.registry import HOOKS
+from colossalai.legacy.registry import HOOKS
 
 from ._metric_hook import LearningRateMetric, MetricHook
 

colossalai/legacy/trainer/hooks/_metric_hook.py

@@ -10,7 +10,7 @@
 from colossalai.communication import all_reduce
 from colossalai.context import ParallelMode
 from colossalai.core import global_context as gpc
-from colossalai.registry import HOOKS
+from colossalai.legacy.registry import HOOKS
 from colossalai.utils import get_current_device, is_no_pp_or_last_stage
 
 from ._base_hook import BaseHook
@@ -356,7 +356,7 @@ def get_last_step_value(self) -> float:
             self.last_step_num_samples *= gpc.get_world_size(ParallelMode.DATA)
         else:
             self.last_step_used_time = all_reduce(self.last_step_used_time, ParallelMode.DATA) / \
-                 gpc.get_world_size(ParallelMode.DATA)
+                gpc.get_world_size(ParallelMode.DATA)
             self.last_step_num_samples = all_reduce(self.last_step_num_samples, ParallelMode.DATA)
 
         sample_per_sec = _format_number(self.last_step_num_samples / (self.last_step_used_time + 1e-12).item())
@@ -367,7 +367,7 @@ def get_last_step_info(self) -> str:
             self.last_step_num_samples *= gpc.get_world_size(ParallelMode.DATA)
         else:
             self.last_step_used_time = all_reduce(self.last_step_used_time, ParallelMode.DATA) / \
-                 gpc.get_world_size(ParallelMode.DATA)
+                gpc.get_world_size(ParallelMode.DATA)
             self.last_step_num_samples = all_reduce(self.last_step_num_samples, ParallelMode.DATA)
 
         sample_per_sec = _format_number(self.last_step_num_samples / (self.last_step_used_time + 1e-12).item())

colossalai/nn/layer/parallel_1d/layers.py

@@ -15,8 +15,8 @@
 from colossalai.core import global_context as gpc
 from colossalai.global_variables import tensor_parallel_env as env
 from colossalai.kernel import LayerNorm
+from colossalai.legacy.registry import LAYERS
 from colossalai.nn import init as init
-from colossalai.registry import LAYERS
 from colossalai.utils.checkpointing import (
     broadcast_state_dict,
     gather_tensor_parallel_state_dict,

colossalai/nn/layer/parallel_2d/layers.py

@@ -5,21 +5,30 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+from torch import Tensor
+from torch.nn import Parameter
+
 from colossalai.communication import broadcast
 from colossalai.context import ParallelMode, seed
 from colossalai.core import global_context as gpc
 from colossalai.global_variables import tensor_parallel_env as env
+from colossalai.legacy.registry import LAYERS
 from colossalai.nn import init as init
-from colossalai.registry import LAYERS
 from colossalai.utils.checkpointing import gather_tensor_parallel_state_dict, partition_tensor_parallel_state_dict
 from colossalai.utils.cuda import get_current_device
-from torch import Tensor
-from torch.nn import Parameter
 
 from ..base_layer import ParallelLayer
 from ..utils import divide, set_tensor_parallel_attribute_by_partition, to_2tuple
-from ._operation import (Matmul_AB_2D, Matmul_ABT_2D, add_bias_2d, all_gather_tensor_2d, classifier_2d, layernorm_2d,
-                         reduce_scatter_tensor_2d, split_batch_2d)
+from ._operation import (
+    Matmul_AB_2D,
+    Matmul_ABT_2D,
+    add_bias_2d,
+    all_gather_tensor_2d,
+    classifier_2d,
+    layernorm_2d,
+    reduce_scatter_tensor_2d,
+    split_batch_2d,
+)
 from ._utils import assert_summa_initialization, get_summa_dim_from_env
 
 

colossalai/nn/layer/parallel_2p5d/layers.py

@@ -5,22 +5,34 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+from torch import Tensor
+from torch.nn import Parameter
+
 from colossalai.communication import broadcast
 from colossalai.context import ParallelMode, seed
 from colossalai.core import global_context as gpc
 from colossalai.global_variables import tensor_parallel_env as env
+from colossalai.legacy.registry import LAYERS
 from colossalai.nn import init as init
-from colossalai.registry import LAYERS
-from colossalai.utils.checkpointing import (broadcast_state_dict, gather_tensor_parallel_state_dict,
-                                            partition_tensor_parallel_state_dict)
+from colossalai.utils.checkpointing import (
+    broadcast_state_dict,
+    gather_tensor_parallel_state_dict,
+    partition_tensor_parallel_state_dict,
+)
 from colossalai.utils.cuda import get_current_device
-from torch import Tensor
-from torch.nn import Parameter
 
 from ..base_layer import ParallelLayer
 from ..utils import divide, set_tensor_parallel_attribute_by_partition, to_2tuple
-from ._operation import (Matmul_AB_2p5D, Matmul_ABT_2p5D, add_bias_2p5d, all_gather_tensor_2p5d, classifier_2p5d,
-                         layernorm_2p5d, reduce_scatter_tensor_2p5d, split_batch_2p5d)
+from ._operation import (
+    Matmul_AB_2p5D,
+    Matmul_ABT_2p5D,
+    add_bias_2p5d,
+    all_gather_tensor_2p5d,
+    classifier_2p5d,
+    layernorm_2p5d,
+    reduce_scatter_tensor_2p5d,
+    split_batch_2p5d,
+)
 from ._utils import assert_tesseract_initialization, get_tesseract_dim_dep_from_env
 
 

colossalai/nn/layer/parallel_3d/layers.py

@@ -13,9 +13,9 @@
 from colossalai.context import ParallelMode, seed
 from colossalai.core import global_context as gpc
 from colossalai.global_variables import tensor_parallel_env as env
+from colossalai.legacy.registry import LAYERS
 from colossalai.nn import init as init
 from colossalai.nn.layer.base_layer import ParallelLayer
-from colossalai.registry import LAYERS
 from colossalai.utils.checkpointing import (
     broadcast_state_dict,
     gather_tensor_parallel_state_dict,

colossalai/nn/layer/parallel_sequence/layers.py

@@ -2,20 +2,20 @@
 # -*- encoding: utf-8 -*-
 
 import math
-import colossalai
 
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from torch.nn import Parameter
 
+import colossalai
+from colossalai.context import seed
 from colossalai.context.parallel_mode import ParallelMode
 from colossalai.core import global_context as gpc
-from colossalai.nn.layer.parallel_sequence._operation import RingQK, RingAV
-from colossalai.registry import LAYERS
-from colossalai.kernel.cuda_native.scaled_softmax import AttnMaskType
 from colossalai.kernel import FusedScaleMaskSoftmax
-from colossalai.context import seed
+from colossalai.kernel.cuda_native.scaled_softmax import AttnMaskType
+from colossalai.legacy.registry import LAYERS
+from colossalai.nn.layer.parallel_sequence._operation import RingAV, RingQK
 
 
 @LAYERS.register_module

colossalai/nn/layer/vanilla/layers.py

@@ -8,8 +8,8 @@
 from torch.nn.parameter import Parameter
 
 from colossalai.context import seed
+from colossalai.legacy.registry import LAYERS
 from colossalai.nn import init as init
-from colossalai.registry import LAYERS
 from colossalai.utils.cuda import get_current_device
 
 from ..utils import to_2tuple