**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.
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):
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.
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).
`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.
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.
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.
- 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/<model-name>.py`**.
- You can implement the `ParallelModule` for primitive modules in the `shardformer/layer/<model-name>.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.
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.
#### 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.
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.
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.
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.
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.
