# ⚡️ ShardFormer ## 📚 Table of Contents - [⚡️ ShardFormer](#️-shardformer) - [📚 Table of Contents](#-table-of-contents) - [🔗 Introduction](#-introduction) - [🔨 Usage](#-usage) - [Quick Start](#quick-start) - [Write your own policy](#write-your-own-policy) - [🗺 Roadmap](#-roadmap) - [💡 API Design](#-api-design) - [Distributed Modules](#distributed-modules) - [Shard Config](#shard-config) - [Policy](#policy) - [Model Sharder](#model-sharder) - [User-facing API](#user-facing-api) - [⌨️ Development Notes](#️-development-notes) - [Add New Policy to Shardformer](#add-new-policy-to-shardformer) - [Write Your Unit Testing](#write-your-unit-testing) - [📊 Benchmarking](#-benchmarking) - [System Performance](#system-performance) - [Convergence](#convergence) ## 🔗 Introduction **Shardformer** is a module that automatically parallelizes the mainstream models in libraries such as HuggingFace and TIMM. This module aims to make parallelization hassle-free for users who are not from the system background. ## 🔨 Usage ### Quick Start The sample API usage is given below(If you enable the use of flash attention, please install `flash_attn`. In addition, xformers's `cutlass_op` provide a supplementary optimization): ```python from colossalai.shardformer import ShardConfig, ShardFormer from transformers import BertForMaskedLM import colossalai # launch colossalai colossalai.launch_from_torch() # create model config = BertConfig.from_pretrained('bert-base-uncased') model = BertForMaskedLM.from_pretrained('bert-base-uncased', config=config) # create huggingface model as normal shard_config = ShardConfig(tensor_parallel_process_group=tp_group, pipeline_stage_manager=stage_manager, enable_tensor_parallelism=True, enable_fused_normalization=True, enable_flash_attention=True, enable_jit_fused=True, enable_sequence_parallelism=True, enable_sequence_overlap=True) shard_former = ShardFormer(shard_config=shard_config) sharded_model, shared_params = shard_former.optimize(model).to('cuda') # do everything like normal ... ``` Following are the description `ShardConfig`'s arguments: - `tensor_parallel_process_group`: The process group of tensor parallelism, it's necessary when using tensor parallel. Defaults to None, which is the global process group. - `pipeline_stage_manager`: If using pipeline parallelism, it's necessary to specify a pipeline stage manager for inter-process communication in pipeline parallelism. Defaults to None, which means not using pipeline parallelism. - `enable_tensor_parallelism`: Whether to use tensor parallelism. Defaults to True. - `enable_fused_normalization`: Whether to use fused layernorm. Defaults to False. - `enable_flash_attention`: Whether to switch on flash attention. Defaults to False. - `enable_jit_fused`: Whether to switch on JIT fused operators. Defaults to False. - `enable_sequence_parallelism`: Whether to turn on sequence parallelism, which partitions non-tensor-parallel regions along the sequence dimension. Defaults to False. - `enable_sequence_overlap`: Whether to turn on sequence overlap, which overlap the computation and communication in sequence parallelism. It can only be used when `enable_sequence_parallelism` is True. Defaults to False. - `enable_all_optimization`: Whether to turn on all optimization tools including `fused normalization`, `flash attention`, `JIT fused operators`, `sequence parallelism` and `sequence overlap`. Defaults to False. - `extra_kwargs`: A dict to store extra kwargs for ShardFormer. ### Write your own policy If you have a custom model, you can also use Shardformer to parallelize it by writing your own sharding policy. More information about the sharding policy can be found in [API Design](#-api-design). ```python from colossalai.shardformer import Policy class MyPolicy(Policy): # implement your own policy ... # init model and shard former ... # use customized policy to shard model my_policy = MyPolicy() shard_former.optimize(model, my_policy) ``` ## 🗺 Roadmap We will follow this roadmap to develop Shardformer: - [x] API Design - [x] API Implementation - [x] Unit Testing - [ ] Policy Implementation | model | tensor parallel | pipeline parallel | lazy initialization | xformer | flash attn2 | jit fused operator | fused layernorm | sequence parallel | overlap | |:-----------:|:---------------:|:-----------------:|:-------------------:|:-------:|:-----------:|:------------------:|:---------------:|:-----------------:|:-------:| | bert | [√] | [√] | [√] | [√] | [√] | [√] | [√] | [√] | [√] | | t5 | [√] | [√] | [√] | [√] | [√] | [√] | [√] | [ ] | [ ] | | llama V1/V2 | [√] | [√] | [√] | [√] | [√] | [√] | [√] | [ ] | [ ] | | gpt2 | [√] | [√] | [√] | [√] | [√] | [√] | [√] | [√] | [√] | | opt | [√] | [√] | [√] | [√] | [√] | [√] | [√] | [ ] | [ ] | | bloom | [√] | [√] | [√] | [√] | [√] | [√] | [√] | [√] | [√] | | chatglm2 | [√] | [√] | [√] | [√] | [√] | [√] | [√] | [√] | [√] | | vit | [√] | [√] | [ ] | [√] | [√] | [√] | [√] | [ ] | [ ] | | whisper | [√] | [√] | [√] | [√] | [√] | [ ] | [√] | [ ] | [ ] | | sam | [√] | [ ] | [ ] | [√] | [√] | [√] | [√] | [ ] | [ ] | | blip2 | [√] | [ ] | [ ] | [√] | [√] | [√] | [√] | [ ] | [ ] | | falcon | [√] | [√] | [√] | [√] | [√] | [ ] | [√] | [ ] | [ ] | | roberta | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | | albert | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | | ernie | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | | gpt-neo | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | | gpt-j | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | | beit | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | | swin | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | | swin V2 | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | | qwen | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | | mistral | [√] | [ ] | [ ] | [√] | [√] | [√] | [√] | [ ] | [ ] | ## 💡 API Design We will discuss the major components of `ShardFormer` below to help you better understand how things work. This section serves as the design doc for Shardformer and the function signature might differ from the actual implementation. Please refer to the code for more details.


### Distributed Modules `ShardFormer` replaces the original PyTorch module with a distributed module. The distributed module keeps the same attributes as the original module but replaces the original parameters with distributed parameters and defines a new `forward` function to execute distributed computation. Each distributed module implements its `from_native_module` static method to convert the PyTorch module to its corresponding distributed module. ````python class ParallelModule(torch.nn.Module): @abstractmethod def from_native_module(module: torch.nn.Module, process_group: Union[ProcessGroup, Tuple[ProcessGroup]]) -> ParallelModule """ Convert a native module to a parallelized Examples: ```python # replace module my_linear = Linear1D_Col.from_native_module(my_linear, process_group) ``` """ ```` ### Shard Config `ShardConfig` is a simple data class to tell `ShardFormer` how sharding will be performed. ```python @dataclass class ShardConfig: tensor_parallel_process_group: ProcessGroup = None enable_fused_normalization: bool = False ... # Some possible future config fields tensor_parallel_mode: Choice['1d', '2d', '2.5d', '3d'] # support different tensor parallel mode use_flash_attention: bool # whether to use flash attention to speed up attention extra_kwargs: Dict[str, Any] # extra kwargs for the shardformer ``` ### Policy The `Policy` class describes how to handle the model sharding. It is merely a description, the actual sharding will be performed by `ModelSharder`. We abstract the policy into four stages: 1. Preprocessing: call `Policy.preprocess` to do some prior work before sharding, for example, resizing the embedding 2. Providing `ModulePolicyDescription`: call `Policy.module_policy` to get a bunch of `ModulePolicyDescription` to tell `ModelSharder` how the submodules's attributes, child parameters, and deeper submodules will be substituted. 3. Postprocessing: call `Policy.postprocess` to perform some postprocessing work, for example, binding the embedding and classifier head weights of the BERT model. ```python @dataclass class ModulePolicyDescription: r""" Describe how the attributes and parameters will be transformed in a policy. Args: attribute_replacement (Dict[str, Any]): key is the attribute name, value is the attribute value after sharding param_replacement (List[Callable]): a list of functions to perform in-place param replacement. The function must receive only one arguments: module. sub_module_replacement (List[SubModuleReplacementDescription]): each element in the list is a ParamReplacementDescription object which specifies the module to be replaced and the target module used to replacement. method_replace (Dict[str, Callable]): key is the method name, value is the method for replacement """ attribute_replacement: Dict[str, Any] = None param_replacement: List[Callable] = None sub_module_replacement: List[SubModuleReplacementDescription] = None method_replacement: Dict[str, Callable] = None @dataclass class SubModuleReplacementDescription: r""" Describe how a submodule will be replaced Args: suffix (str): used to get the submodule object target_module (ParallelModule): specifies the module class used to replace to submodule kwargs (Dict[str, Any]): the dictionary used to pass extra arguments to the `ParallelModule.from_native_module` method. ignore_if_not_exist (bool): if the submodule does not exist, ignore it or raise an exception """ suffix: str target_module: ParallelModule kwargs: Dict[str, Any] = None ignore_if_not_exist: bool = False class Policy(ABC): r""" The base class for all the policies. For each different model, it should have a different policy class, like BertPolicy for Bert Model or OPTPolicy for OPT model. Shardformer has provided many built-in sharding policies for the mainstream models. You can use the built-in policies by setting `policy = None`, which is already the default argument for `Shardformer.optimize`. If you want to define your own policy, you can inherit from this class and overwrite the methods you want to modify. """ def __init__(self) self.model = None def set_model(self, model: nn.Module) -> None: """ Set model as an attribute of the Policy object so that we can access the model's attributes. """ self.model = model def set_shard_config(self, shard_config: ShardConfig) -> None: r""" Set shard config as an attribute of the Policy object. Args: shard_config (:class:`ShardConfig`): The shard config to be perform """ self.shard_config = shard_config self.config_sanity_check() @abstractmethod def preprocess(self) -> nn.Module: """ Perform some preprocessing on the model, such as resizing the embedding size """ ... @abstractmethod def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: """ Return the dict for the modify policy, the key is the original layer class and the value is the argument for the modify layer """ ... @abstractmethods def postprocess(self) -> nn.Module: """ Perform some postprocessing on the model, such as binding the embedding with the weight of the classifier head """ ... ``` ### Model Sharder `ModelSharder` is the class in charge of sharding the model based on the given policy. ```python class ModelSharder: def __init__(self, model: torch.nn.Module, shard_config: ShardConfig, Policy: ShardPolicy = None): #TODO: input is a cls or a obj ... def shard(self) -> None: """ Shard model with parallelism with the help of pre-processing, replace_model_class, replace_module, and post-processing. """ ... def replace_module(self) -> None: """ Replace the layer according to the policy. Call Policy.module_policy() to get the module. Call _replace_module recursively. """ ... ``` ### User-facing API We only expose a limited number of APIs to the user to keep their user experience simple and clean. ```python class ShardFormer: """ Parallelize model based on the given config and policy Example: org_model = BertForMaskedLM.from_pretrained('bert-base-uncased') shard_config = ShardConfig() shard_former = ShardFormer(shard_config=shard_config) model, shared_params = shard_former.optimize(org_model) """ def __init__(self, shard_config: ShardConfig): """ Do two things: 1. Create a distribute coordinator 2. serve as a store for shard config """ self.shard_config = shard_config self.coordinator = DistCoordinator() def optimize(self, model: nn.Module, policy: Policy = None) -> Tuple[nn.Module, List[Dict[int, Tensor]]]: r""" This method will optimize the model based on the given policy. Args: model (`torch.nn.Model`): the origin huggingface model shard_config (`ShardConfig`): the config for distribute information policy (`Policy`): the custom policy for sharding Returns: the sharded model and the shared parameters """ sharder = ModelSharder(model=model, shard_config=self.shard_config, policy=policy) shared_params = sharder.shard() return model, shared_params ``` ## ⌨️ Development Notes ### Add New Policy to Shardformer This section serves as the guideline for writing new policies and register them into `shardformer`. - Step 1. Write your own model policy You can create a new file in the `colossalai/shardformer/policies` folder and name the file with the model name. You can implement your policy in this file. You should not import the any model zoo library at the header section of the file because we do not want to import the library when we do not use the policy. Libraries such as `transformers` should be imported only in the function body when needed. Please follow the following protocols when writing your policy: - You have to make a clear decision what you want to replace exactly in the original PyTorch module - Use `ModulePolicyDescription.attribute_replacement` to replace the module attributes - Use `ModulePolicyDescription.param_replacement` to replace the module parameters - Use `ModulePolicyDescription.sub_module_replacement` to replace the submodules completely. The target module should implement the `from_native_module` for the replacement. - Use `ModulePolicyDescription.method_replacement` to replace the module methods. **These replacement methods should be put in the `shardformer/modeling/.py`**. - You can implement the `ParallelModule` for primitive modules in the `shardformer/layer/.py` file. Primitive modules refer to modules which are not composed of other modules. For example, the `torch.nn.Linear` module is a primitive module while modules such as `BertEncoder` module in the `transformers` library is a composite module. Primitive modules do not nested inner `nn.Module` members. For composite modules, you should consider using `ModulePolicyDescription` to implement your replacement. - `ParallelModule` is meant to be used in two ways: `ParallelModule.from_native_module` to convert native PyTorch module to the `ParallelModule` and `ParallelModule(...)` to instantiate the module directly just like a normal PyTorch module. `ParallelModule` should be only implemented for modules whose weights are sharded. If you want to make your module compatible with the `ModulePolicyDescription.sub_module_replacement` and there is no weight sharding in your module, you can just implement the `from_native_module` method without inheriting the `ParallelModule` like `colossalai/shardformer/layer/normalization.py`. - **Do not import any file in the `colossalai/shardformer/policies` and `colossalai/shardformer/modeling` to avoid unwanted import error**. For example, a file in these folders accidentally imports `transformers` library at the top of the file, then the user will have to install `transformers` library even if they do not use this file. Any file in the `modeling` folder should be only imported by the policy file. A policy implementation should be only imported dynamically via the autopolicy or manually via the `ShardFormer` module. - Try to keep your import statement on third-party libraries such as `transformers` within the function body instead of the header section of the file. This is because we do not want to import the library when we do not use the policy. - Step 2. Register your policy to the autopolicy Next, you need to register your policy in the `colossalai/shardformer/policies/autopolicy.py` file. For example, if we register the policy for the BERT model, we just add a key-value in the `_POLICY_LIST` dictionary. The key if the `qualname` of the model object (you can get it by model.\_\_class\_\_.\_\_qualname\_\_). The value is a `PolicyLocation` object, which contains the file name and the class name of the policy. We do not import the policy directly because the policy file may contain libraries (such as `transformers`) which we do not want to import when we do not use the policy. ```python _POLICY_LIST = { # BERT "transformers.models.bert.modeling_bert.BertModel": PolicyLocation(file_name="bert", class_name="BertModelPolicy"), } ``` #### How to support those models in huggingface model hub but not in the transformers library There are two cases: 1. the modeling file is in the `transformers` library but the model weight is not in the `transformers` library. E.g. model structure of "01-ai/Yi-34B" is the same as LLaMA but the weight is not in the `transformers` library. In this case, we should support llama as usual and Yi-34B is also supported by the llama policy. We do not need to add a new policy for Yi-34B. 2. the modeling file is not in the `transformers` library, such as the "THUDM/chatglm2-6b". Take "THUDM/chatglm2-6b" as an example, we clearly illustrate how to support this model in the `shardformer`. Unlike llama which is in `transformers` library, we cannot import chatglm2 model directly. Thus, the key in policy should be str of class name, rather than class itself. E.g. for llama: ```python policy[LlamaDecoderLayer] = ModulePolicyDescription(...) ``` for chatglm2: ```python policy["GLMBlock"] = ModulePolicyDescription(...) ``` Then when registering such models in the autopolicy, we should follow below format: ```python "transformers_modules..": PolicyLocation( file_name="", class_name="" ) ``` As for chatglm2 model, it should be: ```python "transformers_modules.modeling_chatglm.ChatGLMForConditionalGeneration": PolicyLocation( file_name="chatglm2", class_name="ChatGLMForConditionalGenerationPolicy" ) ``` When using such models, `AutoModel` is supported as usual. The policy will be automatically loaded by the autopolicy. ### Write Your Unit Testing This section serves as the guideline for testing the `shardformer` module. - Step 1. Add your model to the model zoo in the test kits. Add your model to the `tests/kit/model_zoo` file. This allows you to define test-related components for this model. You can take `tests/kit/model_zoo/transformers/llama.py` as an example for reference. - Step 2. Write your unit testing for the model Next, implement your unit test in the `tests/test_shardformer` folder. Please refer to other similar tests for style consistency. - Step 3. Execute your test When you run tests locally, you should run tests for both your newly-added test file and the whole `shardformer` module tests. ```bash # test for your own test file pytest tests/test_shardformer/test_model/.py # test for the whole shardformer module pytest tests/test_shardformer ``` ## 📊 Benchmarking ### System Performance We conducted [benchmark tests](./examples/performance_benchmark.py) to evaluate the performance improvement of Shardformer. We compared the training time between the original model and the shard model. We set the batch size to 4, the number of attention heads to 8, and the head dimension to 64. 'N_CTX' refers to the sequence length. In the case of using 2 GPUs, the training times are as follows. | N_CTX | org_model | shard_model | |:-----:|:---------:|:-----------:| | 256 | 11.2ms | 17.2ms | | 512 | 9.8ms | 19.5ms | | 1024 | 19.6ms | 18.9ms | | 2048 | 46.6ms | 30.8ms | | 4096 | 160.5ms | 90.4ms |


In the case of using 4 GPUs, the training times are as follows. | N_CTX | org_model | shard_model | |:-----:|:---------:|:-----------:| | 256 | 10.0ms | 21.1ms | | 512 | 11.5ms | 20.2ms | | 1024 | 22.1ms | 20.6ms | | 2048 | 46.9ms | 24.8ms | | 4096 | 160.4ms | 68.0ms |


As shown in the figures above, when the sequence length is around 1000 or greater, the parallel optimization of Shardformer for long sequences starts to become evident. ### Convergence To validate that training the model using shardformers does not impact its convergence. We [fine-tuned the BERT model](./examples/convergence_benchmark.py) using both shardformer and non-shardformer approaches. The example that utilizes Shardformer simultaneously with Pipeline Parallelism and Data Parallelism (Zero1). We then compared the accuracy, loss, and F1 score of the training results. the configurations are as follows: ```python batch_size = 2 epoch = 3 lr = 2.4e-5 accumulation_steps = 8 warmup_fraction = 0.03 ``` | accuracy | f1 | loss | GPU number | model sharded | |:--------:|:-------:|:-------:|:----------:|:-------------:| | 0.82971 | 0.87713 | 0.23194 | 4 | True | | 0.83797 | 0.88006 | 0.22683 | 2 | True | | 0.84521 | 0.88700 | 0.21822 | 1 | False | Overall, the results demonstrate that using shardformers during model training does not affect the convergence.