Merge branch 'main' into feature/shardformer

docs/source/en/advanced_tutorials/train_vit_with_hybrid_parallelism.md

@@ -79,7 +79,7 @@ from colossalai.core import global_context as gpc
 from colossalai.logging import disable_existing_loggers, get_dist_logger
 from colossalai.nn.lr_scheduler import LinearWarmupLR
 from colossalai.nn.metric import Accuracy
-from colossalai.trainer import Trainer, hooks
+from colossalai.legacy.trainer import Trainer, hooks
 ```
 
 - Other modules
@@ -273,8 +273,8 @@ SEQ_LENGTH = (IMG_SIZE // PATCH_SIZE) ** 2 + 1  # add 1 for cls token
 
 ### Build pipeline model (`/hybrid_parallel/model/vit.py`)
 Colossal-AI provides two methods to build a pipeline model from the existing model.
-- `colossalai.builder.build_pipeline_model_from_cfg`
-- `colossalai.builder.build_pipeline_model`
+- `colossalai.legacy.builder.build_pipeline_model_from_cfg`
+- `colossalai.legacy.builder.build_pipeline_model`
 
 Besides, you can also build a pipeline model from scratch with Colossal-AI.
 ```python
@@ -284,11 +284,11 @@ from typing import Callable
 import inspect
 import torch
 from colossalai import nn as col_nn
-from colossalai.registry import LAYERS, MODELS
+from colossalai.legacy.registry import LAYERS, MODELS
 from colossalai.logging import get_dist_logger
 from colossalai.core import global_context as gpc
 from colossalai.context import ParallelMode
-from colossalai.builder.pipeline import partition_uniform
+from colossalai.legacy.builder.pipeline import partition_uniform
 from torch import dtype, nn
 from model_zoo.vit.vit import ViTBlock, ViTEmbedding, ViTHead
 
@@ -415,7 +415,7 @@ def build_pipeline_vit(num_layers, num_chunks, device=torch.device('cuda'), **kw
 
 #### Import modules
 ```python
-from colossalai.engine.schedule import (InterleavedPipelineSchedule,
+from colossalai.legacy.engine.schedule import (InterleavedPipelineSchedule,
                                         PipelineSchedule)
 from colossalai.utils import MultiTimer
 import os
@@ -644,3 +644,4 @@ torchrun --standalone --nproc_per_node <NUM_GPUs>  train_hybrid.py --config ./co
 # If your torch >= 1.9.0
 # python -m torch.distributed.run --standalone --nproc_per_node= <NUM_GPUs> train_hybrid.py --config ./configs/config_hybrid_parallel.py
 ```
+<!-- doc-test-command: echo  -->

docs/source/en/basics/engine_trainer.md

@@ -64,7 +64,7 @@ Trainer is a more high-level wrapper for the user to execute training with fewer
 
 ```python
 from colossalai.logging import get_dist_logger
-from colossalai.trainer import Trainer, hooks
+from colossalai.legacy.trainer import Trainer, hooks
 
 # build components and initialize with colossalai.initialize
 ...
@@ -107,7 +107,7 @@ If you want to customize your own hook class, you can inherit `hooks.BaseHook` a
 
 ```python
 from colossalai.logging import get_dist_logger
-from colossalai.trainer import hooks
+from colossalai.legacy.trainer import hooks
 
 class LogMessageHook(hooks.BaseHook):
 
@@ -345,7 +345,7 @@ If you wish to train with a trainer object, you can follow the code snippet belo
 
 ```python
 from colossalai.nn.metric import Accuracy
-from colossalai.trainer import Trainer, hooks
+from colossalai.legacy.trainer import Trainer, hooks
 
 
 # create a trainer object
@@ -387,3 +387,4 @@ python -m torch.distributed.launch --nproc_per_node <num_gpus> --master_addr loc
 # with trainer
 python -m torch.distributed.launch --nproc_per_node <num_gpus> --master_addr localhost --master_port 29500 run_resnet_cifar10_with_trainer.py
 ```
+<!-- doc-test-command: echo  -->

docs/source/en/basics/model_checkpoint.md

@@ -41,7 +41,7 @@ for epoch in range(num_epochs):
 
 #### Save when using trainer
 ```python
-from colossalai.trainer import Trainer, hooks
+from colossalai.legacy.trainer import Trainer, hooks
 model = ...
 engine, _, _, _ = colossalai.initialize(model=model, ...)
 trainer = Trainer(engine, ...)
@@ -61,3 +61,4 @@ model = ...
 load_checkpoint('xxx.pt', model)
 ... # train or test
 ```
+<!-- doc-test-command: echo  -->

docs/source/en/features/gradient_handler.md

@@ -28,8 +28,8 @@ To implement a customized gradient handler, you need to follow these steps.
 3. implement `handle_gradient` method.
 
 ```python
-from colossalai.registry import GRADIENT_HANDLER
-from colossalai.engine.gradient_handler import BaseGradientHandler
+from colossalai.legacy.registry import GRADIENT_HANDLER
+from colossalai.legacy.engine.gradient_handler import BaseGradientHandler
 
 
 @GRADIENT_HANDLER.register_module
@@ -61,3 +61,4 @@ to demonstrate the use of gradient handler. In this example, we used `DataParall
 ```shell
 python -m torch.distributed.launch --nproc_per_node 4 --master_addr localhost --master_port 29500  train_with_engine.py
 ```
+<!-- doc-test-command: echo  -->

docs/source/en/features/mixed_precision_training.md

@@ -267,7 +267,7 @@ from pathlib import Path
 from colossalai.core import global_context as gpc
 from colossalai.logging import get_dist_logger
 from colossalai.utils import get_dataloader
-from colossalai.trainer import Trainer, hooks
+from colossalai.legacy.trainer import Trainer, hooks
 from colossalai.nn.lr_scheduler import LinearWarmupLR
 from timm.models import vit_base_patch16_224
 from torchvision import datasets, transforms

docs/source/en/features/pipeline_parallel.md

@@ -79,7 +79,7 @@ import colossalai.nn as col_nn
 
 from colossalai.core import global_context as gpc
 from colossalai.logging import disable_existing_loggers, get_dist_logger
-from colossalai.trainer import Trainer, hooks
+from colossalai.legacy.trainer import Trainer, hooks
 from colossalai.utils import MultiTimer, get_dataloader
 from colossalai.context import ParallelMode
 from colossalai.pipeline.pipelinable import PipelinableContext
@@ -157,3 +157,4 @@ trainer.fit(train_dataloader=train_dataloader,
 ```
 
 We use `2` pipeline stages and the batch will be split into `4` micro batches.
+<!-- doc-test-command: echo  -->

docs/source/zh-Hans/advanced_tutorials/add_your_parallel.md

@@ -81,14 +81,14 @@ Colossal-AI 为用户提供了一个全局 context，使他们能够轻松地管
 ## 梯度 Handler
 
 梯度 handler 是对参数的梯度执行 all-reduce 操作的对象。由于不同的 all-reduce 策略或许在不同的并行中被执行，用户可以继承
-`colossalai.engine.gradient_handler.BaseGradientHandler` 来实现其策略。目前，Colossal-AI 使用普通的数据并行梯度 handler 在数据并行的 rank 间 all-reduce 梯度。
+`colossalai.legacy.engine.gradient_handler.BaseGradientHandler` 来实现其策略。目前，Colossal-AI 使用普通的数据并行梯度 handler 在数据并行的 rank 间 all-reduce 梯度。
 如果数据并行被检测到，梯度 handler 会被自动添加进 engine。
 
 你可以添加你自己的梯度 handler，如下所示：
 
 ```python
-from colossalai.registry import GRADIENT_HANDLER
-from colossalai.engine import BaseGradientHandler
+from colossalai.legacy.registry import GRADIENT_HANDLER
+from colossalai.legacy.engine import BaseGradientHandler
 
 @GRADIENT_HANDLER.register_module
 class YourGradientHandler(BaseGradientHandler):
@@ -109,4 +109,5 @@ gradient_handlers = [
 ## Schedule
 
 Schedule 包含了如何执行前向和后向计算。目前， Colossal-AI 提供了流水和非流水的 schedule。
-如果你想修改前向和后向计算的执行方式，你可以继承 `colossalai.engine.schedule.BaseSchedule` 并实现 `forward_back_step` 函数。
+如果你想修改前向和后向计算的执行方式，你可以继承 `colossalai.legacy.engine.schedule.BaseSchedule` 并实现 `forward_back_step` 函数。
+<!-- doc-test-command: echo  -->

docs/source/zh-Hans/advanced_tutorials/train_gpt_using_hybrid_parallelism.md

@@ -36,14 +36,14 @@ import torch
 import torch.nn as nn
 from colossalai import nn as col_nn
 from colossalai.amp import AMP_TYPE
-from colossalai.builder.pipeline import partition_uniform
+from colossalai.legacy.builder.pipeline import partition_uniform
 from colossalai.context.parallel_mode import ParallelMode
 from colossalai.core import global_context as gpc
-from colossalai.engine.schedule import (InterleavedPipelineSchedule,
+from colossalai.legacy.engine.schedule import (InterleavedPipelineSchedule,
                                         PipelineSchedule)
 from colossalai.logging import disable_existing_loggers, get_dist_logger
 from colossalai.nn.layer.wrapper import PipelineSharedModuleWrapper
-from colossalai.trainer import Trainer, hooks
+from colossalai.legacy.trainer import Trainer, hooks
 from colossalai.utils.timer import MultiTimer
 from model_zoo.gpt import GPTLMLoss
 from torch.nn import functional as F
@@ -273,3 +273,4 @@ def train():
         return_output_label=False,
     )
 ```
+<!-- doc-test-command: echo  -->

docs/source/zh-Hans/advanced_tutorials/train_vit_using_pipeline_parallelism.md

@@ -32,11 +32,11 @@ import colossalai
 import colossalai.nn as col_nn
 import torch
 import torch.nn as nn
-from colossalai.builder import build_pipeline_model
-from colossalai.engine.schedule import (InterleavedPipelineSchedule,
+from colossalai.legacy.builder import build_pipeline_model
+from colossalai.legacy.engine.schedule import (InterleavedPipelineSchedule,
                                         PipelineSchedule)
 from colossalai.logging import disable_existing_loggers, get_dist_logger
-from colossalai.trainer import Trainer, hooks
+from colossalai.legacy.trainer import Trainer, hooks
 from colossalai.utils import MultiTimer, get_dataloader
 from timm.models import vision_transformer as vit
 from torchvision import transforms
@@ -48,17 +48,17 @@ from torchvision.datasets import CIFAR10
 
 总的来说, 我们提供3种方法来建立一个流水并行的模型:
 
-1. `colossalai.builder.build_pipeline_model_from_cfg`
-2. `colossalai.builder.build_pipeline_model`
+1. `colossalai.legacy.builder.build_pipeline_model_from_cfg`
+2. `colossalai.legacy.builder.build_pipeline_model`
 3. 自己按阶段拆分模型
 
 当你的内存能够容纳模型时，你可以使用前两种方法来建立你的模型，否则你必须自己分割模型。前两种方法首先在 CPU 上建立整个模型，然后分割模型，最后你可以直接把模型的相应部分移到 GPU 上。
 
-`colossalai.builder.build_pipeline_model_from_cfg()` 接收一个模型的配置文件，它可以均匀地（按层）或平衡地（按参数大小）分割模型。
+`colossalai.legacy.builder.build_pipeline_model_from_cfg()` 接收一个模型的配置文件，它可以均匀地（按层）或平衡地（按参数大小）分割模型。
 
-如果你熟悉 `PyTorch`, 你可以使用 `colossalai.builder.build_pipeline_model()` 它接收一个 `torch.nn.Sequential` 模型并按层均匀分割。
+如果你熟悉 `PyTorch`, 你可以使用 `colossalai.legacy.builder.build_pipeline_model()` 它接收一个 `torch.nn.Sequential` 模型并按层均匀分割。
 
-在本教程中，我们将修改 [TIMM/ViT](https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py) to `torch.nn.Sequential`，然后使用 `colossalai.builder.build_pipeline_model()` 来建立流水线模型。
+在本教程中，我们将修改 [TIMM/ViT](https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py) to `torch.nn.Sequential`，然后使用 `colossalai.legacy.builder.build_pipeline_model()` 来建立流水线模型。
 
 当数据是 **一个** `Tensor`, 你可以使用你的模型 `forward()` 中的位置参数来获得数据张量。对于流水线的第一阶段，`forward()` 的第一个位置参数是从数据加载器加载的数据张量。对于其他阶段，`forward()` 的第一个位置参数是上一阶段的输出张量。注意，如果该阶段不是最后一个阶段，则 `forward()` 的返回必须是一个 `Tensor`。
 
@@ -244,3 +244,4 @@ def train():
                 hooks=hook_list,
                 display_progress=True)
 ```
+<!-- doc-test-command: echo  -->

docs/source/zh-Hans/advanced_tutorials/train_vit_with_hybrid_parallelism.md

@@ -74,7 +74,7 @@ from colossalai.core import global_context as gpc
 from colossalai.logging import disable_existing_loggers, get_dist_logger
 from colossalai.nn.lr_scheduler import LinearWarmupLR
 from colossalai.nn.metric import Accuracy
-from colossalai.trainer import Trainer, hooks
+from colossalai.legacy.trainer import Trainer, hooks
 ```
 
 - 其他模块
@@ -256,8 +256,8 @@ SEQ_LENGTH = (IMG_SIZE // PATCH_SIZE) ** 2 + 1  # add 1 for cls token
 
 ### 构建流水线模型 (`/hybrid_parallel/model/vit.py`)
 Colossal-AI 提供了两种从现有模型构建流水线模型的方法。
-- `colossalai.builder.build_pipeline_model_from_cfg`
-- `colossalai.builder.build_pipeline_model`
+- `colossalai.legacy.builder.build_pipeline_model_from_cfg`
+- `colossalai.legacy.builder.build_pipeline_model`
 
 此外，您还可以使用 Colossal-AI 从头开始构建流水线模型。
 ```python
@@ -266,11 +266,11 @@ from typing import Callable
 import inspect
 import torch
 from colossalai import nn as col_nn
-from colossalai.registry import LAYERS, MODELS
+from colossalai.legacy.registry import LAYERS, MODELS
 from colossalai.logging import get_dist_logger
 from colossalai.core import global_context as gpc
 from colossalai.context import ParallelMode
-from colossalai.builder.pipeline import partition_uniform
+from colossalai.legacy.builder.pipeline import partition_uniform
 from torch import dtype, nn
 from model_zoo.vit.vit import ViTBlock, ViTEmbedding, ViTHead
 @MODELS.register_module
@@ -380,7 +380,7 @@ def build_pipeline_vit(num_layers, num_chunks, device=torch.device('cuda'), **kw
 
 #### 导入模块
 ```python
-from colossalai.engine.schedule import (InterleavedPipelineSchedule,
+from colossalai.legacy.engine.schedule import (InterleavedPipelineSchedule,
                                         PipelineSchedule)
 from colossalai.utils import MultiTimer
 import os
@@ -589,3 +589,4 @@ torchrun --standalone --nproc_per_node <NUM_GPUs>  train_hybrid.py --config ./co
 # If your torch >= 1.9.0
 # python -m torch.distributed.run --standalone --nproc_per_node= <NUM_GPUs> train_hybrid.py --config ./configs/config_hybrid_parallel.py
 ```
+<!-- doc-test-command: echo  -->

docs/source/zh-Hans/basics/engine_trainer.md

@@ -61,7 +61,7 @@ Trainer 的参数 `schedule` 默认值是 `None` 。在大多数情况下，除
 
 ```python
 from colossalai.logging import get_dist_logger
-from colossalai.trainer import Trainer, hooks
+from colossalai.legacy.trainer import Trainer, hooks
 
 # build components and initialize with colossalai.initialize
 ...
@@ -104,7 +104,7 @@ trainer.fit(
 
 ```python
 from colossalai.logging import get_dist_logger
-from colossalai.trainer import hooks
+from colossalai.legacy.trainer import hooks
 
 class LogMessageHook(hooks.BaseHook):
 
@@ -341,7 +341,7 @@ for epoch in range(gpc.config.NUM_EPOCHS):
 
 ```python
 from colossalai.nn.metric import Accuracy
-from colossalai.trainer import Trainer, hooks
+from colossalai.legacy.trainer import Trainer, hooks
 
 
 # create a trainer object
@@ -384,3 +384,4 @@ python -m torch.distributed.launch --nproc_per_node <num_gpus> --master_addr loc
 # with trainer
 python -m torch.distributed.launch --nproc_per_node <num_gpus> --master_addr localhost --master_port 29500 run_resnet_cifar10_with_trainer.py
 ```
+<!-- doc-test-command: echo  -->

docs/source/zh-Hans/basics/model_checkpoint.md

@@ -41,7 +41,7 @@ for epoch in range(num_epochs):
 
 #### 用 trainer 保存
 ```python
-from colossalai.trainer import Trainer, hooks
+from colossalai.legacy.trainer import Trainer, hooks
 model = ...
 engine, _, _, _ = colossalai.initialize(model=model, ...)
 trainer = Trainer(engine, ...)
@@ -61,3 +61,4 @@ model = ...
 load_checkpoint('xxx.pt', model)
 ... # train or test
 ```
+<!-- doc-test-command: echo  -->

docs/source/zh-Hans/features/gradient_handler.md

@@ -25,8 +25,8 @@
 3. 实现 `handle_gradient`
 
 ```python
-from colossalai.registry import GRADIENT_HANDLER
-from colossalai.engine.gradient_handler import BaseGradientHandler
+from colossalai.legacy.registry import GRADIENT_HANDLER
+from colossalai.legacy.engine.gradient_handler import BaseGradientHandler
 
 
 @GRADIENT_HANDLER.register_module
@@ -57,3 +57,4 @@ gradient_handler = [dict(type='MyGradientHandler')]
 ```shell
 python -m torch.distributed.launch --nproc_per_node 4 --master_addr localhost --master_port 29500  train_with_engine.py
 ```
+<!-- doc-test-command: echo  -->

docs/source/zh-Hans/features/mixed_precision_training.md

@@ -245,7 +245,7 @@ from pathlib import Path
 from colossalai.core import global_context as gpc
 from colossalai.logging import get_dist_logger
 from colossalai.utils import get_dataloader
-from colossalai.trainer import Trainer, hooks
+from colossalai.legacy.trainer import Trainer, hooks
 from colossalai.nn.lr_scheduler import LinearWarmupLR
 from timm.models import vit_base_patch16_224
 from torchvision import datasets, transforms

docs/source/zh-Hans/features/pipeline_parallel.md

@@ -78,7 +78,7 @@ import colossalai.nn as col_nn
 
 from colossalai.core import global_context as gpc
 from colossalai.logging import disable_existing_loggers, get_dist_logger
-from colossalai.trainer import Trainer, hooks
+from colossalai.legacy.trainer import Trainer, hooks
 from colossalai.utils import MultiTimer, get_dataloader
 from colossalai.context import ParallelMode
 from colossalai.pipeline.pipelinable import PipelinableContext
@@ -156,3 +156,4 @@ trainer.fit(train_dataloader=train_dataloader,
 ```
 
 我们使用 `2` 个流水段，并且 batch 将被切分为 `4` 个 micro batches。
+<!-- doc-test-command: echo  -->

examples/language/gpt/titans/dataset/webtext.py

@@ -6,7 +6,7 @@ import torch
 from torch.utils.data import Dataset
 from transformers import GPT2Tokenizer
 
-from colossalai.registry import DATASETS
+from colossalai.legacy.registry import DATASETS
 
 
 @DATASETS.register_module

examples/language/gpt/titans/model/embed.py

@@ -8,11 +8,11 @@ from torch.nn.parameter import Parameter
 
 from colossalai.context import ParallelMode, seed
 from colossalai.core import global_context as gpc
+from colossalai.legacy.registry import LAYERS, LOSSES, MODELS
 from colossalai.nn.layer.base_layer import ParallelLayer
 from colossalai.nn.layer.parallel_1d._utils import gather_forward_split_backward, reduce_grad, reduce_input
 from colossalai.nn.layer.parallel_1d.layers import Linear1D_Row
 from colossalai.nn.layer.utils import divide
-from colossalai.registry import LAYERS, LOSSES, MODELS
 from colossalai.utils import get_current_device
 
 

examples/language/gpt/titans/train_gpt.py

@@ -10,9 +10,9 @@ import colossalai
 import colossalai.utils as utils
 from colossalai.context.parallel_mode import ParallelMode
 from colossalai.core import global_context as gpc
+from colossalai.legacy.trainer import Trainer, hooks
 from colossalai.logging import disable_existing_loggers, get_dist_logger
 from colossalai.nn import LinearWarmupLR
-from colossalai.trainer import Trainer, hooks
 from colossalai.utils import colo_set_process_memory_fraction, is_using_pp
 from colossalai.utils.timer import MultiTimer
 from colossalai.zero.legacy.init_ctx import ZeroInitContext

examples/tutorial/sequence_parallel/data/datasets/indexed_dataset.py

@@ -3,17 +3,16 @@
 # This source code is licensed under the MIT license found in the
 # LICENSE file in the root directory of this source tree.
 
-
 # copied from fairseq/fairseq/data/indexed_dataset.py
 # Removed IndexedRawTextDataset since it relied on Fairseq dictionary
 # other slight modifications to remove fairseq dependencies
 # Added document index to index file and made it accessible.
 #    An empty sentence no longer separates documents.
 
-from functools import lru_cache
 import os
 import shutil
 import struct
+from functools import lru_cache
 from itertools import accumulate
 
 import numpy as np
@@ -88,16 +87,7 @@ def write_longs(f, a):
     f.write(np.array(a, dtype=np.int64))
 
 
-dtypes = {
-    1: np.uint8,
-    2: np.int8,
-    3: np.int16,
-    4: np.int32,
-    5: np.int64,
-    6: np.float,
-    7: np.double,
-    8: np.uint16
-}
+dtypes = {1: np.uint8, 2: np.int8, 3: np.int16, 4: np.int32, 5: np.int64, 6: float, 7: np.double, 8: np.uint16}
 
 
 def code(dtype):
@@ -136,10 +126,8 @@ class IndexedDataset(torch.utils.data.Dataset):
     def read_index(self, path):
         with open(index_file_path(path), 'rb') as f:
             magic = f.read(8)
-            assert magic == self._HDR_MAGIC, (
-                'Index file doesn\'t match expected format. '
-                'Make sure that --dataset-impl is configured properly.'
-            )
+            assert magic == self._HDR_MAGIC, ('Index file doesn\'t match expected format. '
+                                              'Make sure that --dataset-impl is configured properly.')
             version = f.read(8)
             assert struct.unpack('<Q', version) == (1,)
             code, self.element_size = struct.unpack('<QQ', f.read(16))
@@ -198,13 +186,11 @@ class IndexedDataset(torch.utils.data.Dataset):
 
     @staticmethod
     def exists(path):
-        return (
-            os.path.exists(index_file_path(path)) and os.path.exists(data_file_path(path))
-        )
+        return (os.path.exists(index_file_path(path)) and os.path.exists(data_file_path(path)))
 
     @property
     def supports_prefetch(self):
-        return False  # avoid prefetching to save memory
+        return False    # avoid prefetching to save memory
 
 
 class IndexedCachedDataset(IndexedDataset):
@@ -233,7 +219,7 @@ class IndexedCachedDataset(IndexedDataset):
         for i in indices:
             self.cache_index[i] = ptx
             size = self.data_offsets[i + 1] - self.data_offsets[i]
-            a = self.cache[ptx: ptx + size]
+            a = self.cache[ptx:ptx + size]
             self.data_file.seek(self.data_offsets[i] * self.element_size)
             self.data_file.readinto(a)
             ptx += size
@@ -250,7 +236,7 @@ class IndexedCachedDataset(IndexedDataset):
             tensor_size = self.sizes[self.dim_offsets[i]:self.dim_offsets[i + 1]]
             a = np.empty(tensor_size, dtype=self.dtype)
             ptx = self.cache_index[i]
-            np.copyto(a, self.cache[ptx: ptx + a.size])
+            np.copyto(a, self.cache[ptx:ptx + a.size])
             return a
         elif isinstance(idx, slice):
             # Hack just to make this work, can optimizer later if necessary
@@ -261,15 +247,7 @@ class IndexedCachedDataset(IndexedDataset):
 
 
 class IndexedDatasetBuilder(object):
-    element_sizes = {
-        np.uint8: 1,
-        np.int8: 1,
-        np.int16: 2,
-        np.int32: 4,
-        np.int64: 8,
-        np.float: 4,
-        np.double: 8
-    }
+    element_sizes = {np.uint8: 1, np.int8: 1, np.int16: 2, np.int32: 4, np.int64: 8, float: 4, np.double: 8}
 
     def __init__(self, out_file, dtype=np.int32):
         self.out_file = open(out_file, 'wb')
@@ -332,12 +310,15 @@ def _warmup_mmap_file(path):
 
 
 class MMapIndexedDataset(torch.utils.data.Dataset):
+
     class Index(object):
         _HDR_MAGIC = b'MMIDIDX\x00\x00'
 
         @classmethod
         def writer(cls, path, dtype):
+
             class _Writer(object):
+
                 def __enter__(self):
                     self._file = open(path, 'wb')
 
@@ -384,10 +365,8 @@ class MMapIndexedDataset(torch.utils.data.Dataset):
         def __init__(self, path, skip_warmup=False):
             with open(path, 'rb') as stream:
                 magic_test = stream.read(9)
-                assert self._HDR_MAGIC == magic_test, (
-                    'Index file doesn\'t match expected format. '
-                    'Make sure that --dataset-impl is configured properly.'
-                )
+                assert self._HDR_MAGIC == magic_test, ('Index file doesn\'t match expected format. '
+                                                       'Make sure that --dataset-impl is configured properly.')
                 version = struct.unpack('<Q', stream.read(8))
                 assert (1,) == version
 
@@ -406,16 +385,16 @@ class MMapIndexedDataset(torch.utils.data.Dataset):
             self._bin_buffer_mmap = np.memmap(path, mode='r', order='C')
             self._bin_buffer = memoryview(self._bin_buffer_mmap)
             print("    reading sizes...")
-            self._sizes = np.frombuffer(
-                self._bin_buffer,
-                dtype=np.int32,
-                count=self._len,
-                offset=offset)
+            self._sizes = np.frombuffer(self._bin_buffer, dtype=np.int32, count=self._len, offset=offset)
             print("    reading pointers...")
-            self._pointers = np.frombuffer(self._bin_buffer, dtype=np.int64, count=self._len,
+            self._pointers = np.frombuffer(self._bin_buffer,
+                                           dtype=np.int64,
+                                           count=self._len,
                                            offset=offset + self._sizes.nbytes)
             print("    reading document index...")
-            self._doc_idx = np.frombuffer(self._bin_buffer, dtype=np.int64, count=self._doc_count,
+            self._doc_idx = np.frombuffer(self._bin_buffer,
+                                          dtype=np.int64,
+                                          count=self._doc_count,
                                           offset=offset + self._sizes.nbytes + self._pointers.nbytes)
 
         def __del__(self):
@@ -480,8 +459,7 @@ class MMapIndexedDataset(torch.utils.data.Dataset):
     def __getitem__(self, idx):
         if isinstance(idx, int):
             ptr, size = self._index[idx]
-            np_array = np.frombuffer(self._bin_buffer, dtype=self._index.dtype,
-                                     count=size, offset=ptr)
+            np_array = np.frombuffer(self._bin_buffer, dtype=self._index.dtype, count=size, offset=ptr)
             return np_array
         elif isinstance(idx, slice):
             start, stop, step = idx.indices(len(self))
@@ -491,8 +469,7 @@ class MMapIndexedDataset(torch.utils.data.Dataset):
             sizes = self._index._sizes[idx]
             offsets = list(accumulate(sizes))
             total_size = sum(sizes)
-            np_array = np.frombuffer(self._bin_buffer, dtype=self._index.dtype,
-                                     count=total_size, offset=ptr)
+            np_array = np.frombuffer(self._bin_buffer, dtype=self._index.dtype, count=total_size, offset=ptr)
             sents = np.split(np_array, offsets[:-1])
             return sents
 
@@ -506,8 +483,7 @@ class MMapIndexedDataset(torch.utils.data.Dataset):
         if length is None:
             length = size - offset
         ptr += offset * np.dtype(self._index.dtype).itemsize
-        np_array = np.frombuffer(self._bin_buffer, dtype=self._index.dtype,
-                                 count=length, offset=ptr)
+        np_array = np.frombuffer(self._bin_buffer, dtype=self._index.dtype, count=length, offset=ptr)
         return np_array
 
     @property
@@ -530,12 +506,11 @@ class MMapIndexedDataset(torch.utils.data.Dataset):
 
     @staticmethod
     def exists(path):
-        return (
-            os.path.exists(index_file_path(path)) and os.path.exists(data_file_path(path))
-        )
+        return (os.path.exists(index_file_path(path)) and os.path.exists(data_file_path(path)))
 
 
 class MMapIndexedDatasetBuilder(object):
+
     def __init__(self, out_file, dtype=np.int64):
         self._data_file = open(out_file, 'wb')
         self._dtype = dtype

examples/tutorial/sequence_parallel/requirements.txt

 colossalai
 torch
+six