Distributed Training¶

Sentence Transformers implements two forms of distributed training: Data Parallel (DP) and Distributed Data Parallel (DDP). Read the Data Parallelism documentation on Hugging Face for more details on these strategies. Some of the key differences include:

DDP is generally faster than DP because it has to communicate less data.
With DP, GPU 0 does the bulk of the work, while with DDP, the work is distributed more evenly across all GPUs.
DDP allows for training across multiple machines, while DP is limited to a single machine.

In short, DDP is generally recommended. You can use DDP by running your normal training scripts with torchrun or accelerate. For example, if you have a script called train_script.py, you can run it with DDP using the following command:

Via torchrun

torchrun documentation

torchrun --nproc_per_node=4 train_script.py

Via accelerate

accelerate documentation

accelerate launch --num_processes 4 train_script.py

Note

When performing distributed training, you have to wrap your code in a main function and call it with if __name__ == "__main__":. This is because each process will run the entire script, so you don’t want to run the same code multiple times. Here is an example of how to do this:

from sentence_transformers import SentenceTransformer, SentenceTransformerTrainingArguments, SentenceTransformerTrainer
# Other imports here

def main():
    # Your training code here

if __name__ == "__main__":
    main()

Note

When using an Evaluator, the evaluator only runs on the first device unlike the training and evaluation datasets, which are shared across all devices.

Comparison¶

The following table shows the speedup of DDP over DP and no parallelism given a certain hardware setup.

Hardware: a p3.8xlarge AWS instance, i.e. 4x V100 GPUs
Model being trained: microsoft/mpnet-base (133M parameters)
Maximum sequence length: 384 (following all-mpnet-base-v2)
Training datasets: MultiNLI, SNLI and STSB (note: these have short texts)
Losses: SoftmaxLoss for MultiNLI and SNLI, CosineSimilarityLoss for STSB
Batch size per device: 32

Strategy	Launcher	Samples per Second
No Parallelism	`CUDA_VISIBLE_DEVICES=0 python train_script.py`	2724
Data Parallel (DP)	`python train_script.py` (DP is used by default when launching a script with `python`)	3675 (1.349x speedup)
Distributed Data Parallel (DDP)	`torchrun --nproc_per_node=4 train_script.py` or `accelerate launch --num_processes 4 train_script.py`	6980 (2.562x speedup)

FSDP¶

Fully Sharded Data Parallelism (FSDP) is another distributed training strategy that is not fully supported by Sentence Transformers. It is a more advanced version of DDP that is particularly useful for very large models. Note that in the previous comparison, FSDP reaches 5782 samples per second (2.122x speedup), i.e. worse than DDP. FSDP only makes sense with very large models. If you want to use FSDP with Sentence Transformers, you have to be aware of the following limitations:

You can’t use the evaluator functionality with FSDP.
You have to save the trained model with trainer.accelerator.state.fsdp_plugin.set_state_dict_type("FULL_STATE_DICT") followed with trainer.save_model("output").
You have to use fsdp=["full_shard", "auto_wrap"] and fsdp_config={"transformer_layer_cls_to_wrap": "BertLayer"} in your SentenceTransformerTrainingArguments, where BertLayer is the repeated layer in the encoder that houses the multi-head attention and feed-forward layers, so e.g. BertLayer or MPNetLayer.

Read the FSDP documentation by Accelerate for more details.