|
--- |
|
license: apache-2.0 |
|
datasets: |
|
- c4 |
|
language: |
|
- en |
|
--- |
|
|
|
# MosaicBERT base model |
|
Our goal in developing MosaicBERT was to greatly reduce pretraining time. |
|
|
|
## Model description |
|
|
|
In order to build MosaicBERT, we adopted architectural choices from the recent transformer literature. |
|
These include [FlashAttention (Dao et al. 2022)](https://arxiv.org/pdf/2205.14135.pdf), [ALiBi (Press et al. 2021)](https://arxiv.org/abs/2108.12409), training in an unpadded manner, |
|
low precision LayerNorm, and [Gated Linear Units (Shazeer 2020)](https://arxiv.org/abs/2002.05202). |
|
|
|
### Modifications to the Attention Mechanism |
|
1. **FlashAttention**: Attention layers are core components of the transformer architecture. The recently proposed FlashAttention layer |
|
reduces the number of read/write operations between the GPU HBM (high bandwidth memory, i.e. long-term memory) and the GPU SRAM |
|
(i.e. short-term memory) [[Dao et al. 2022]](https://arxiv.org/pdf/2205.14135.pdf). We used the FlashAttention module built by |
|
[hazy research](https://github.com/HazyResearch/flash-attention) with [OpenAI’s triton library](https://github.com/openai/triton). |
|
|
|
2. **Attention with Linear Biases (ALiBi)**: In most BERT models, the positions of tokens in a sequence are encoded with a position embedding layer; |
|
this embedding allows subsequent layers to keep track of the order of tokens in a sequence. ALiBi eliminates position embeddings and |
|
instead conveys this information using a bias matrix in the attention operation. It modifies the attention mechanism such that nearby |
|
tokens strongly attend to one another [[Press et al. 2021]](https://arxiv.org/abs/2108.12409). In addition to improving the performance of the final model, ALiBi helps the |
|
model to handle sequences longer than it saw during training. Details on our ALiBi implementation can be found [in the mosaicml/examples repo here](https://github.com/mosaicml/examples/blob/d14a7c94a0f805f56a7c865802082bf6d8ac8903/examples/bert/src/bert_layers.py#L425). |
|
|
|
3. **Unpadding**: Standard NLP practice is to combine text sequences of different lengths into a batch, and pad the sequences with empty |
|
tokens so that all sequence lengths are the same. During training, however, this can lead to many superfluous operations on those |
|
padding tokens. In MosaicBERT, we take a different approach: we concatenate all the examples in a minibatch into a single sequence |
|
of batch size 1. Results from NVIDIA and others have shown that this approach leads to speed improvements during training, since |
|
operations are not performed on padding tokens (see for example [Zeng et al. 2022](https://arxiv.org/pdf/2208.08124.pdf)). |
|
Details on our “unpadding” implementation can be found [in the mosaicml/examples repo here](https://github.com/mosaicml/examples/blob/main/examples/bert/src/bert_padding.py). |
|
|
|
4. **Low Precision LayerNorm**: this small tweak forces LayerNorm modules to run in float16 or bfloat16 precision instead of float32, improving utilization. |
|
Our implementation can be found [in the mosaicml/examples repo here](https://docs.mosaicml.com/en/v0.12.1/method_cards/low_precision_layernorm.html). |
|
|
|
### Modifications to the Feedforward Layers |
|
|
|
5. **Gated Linear Units (GLU)**: We used Gated Linear Units for the feedforward sublayer of a transformer. GLUs were first proposed in 2016 [[Dauphin et al. 2016]](https://arxiv.org/abs/1612.08083), |
|
and incorporate an extra learnable matrix that “gates” the outputs of the feedforward layer. More recent work has shown that |
|
GLUs can improve performance quality in transformers [[Shazeer, 2020](https://arxiv.org/abs/2002.05202), [Narang et al. 2021](https://arxiv.org/pdf/2102.11972.pdf)]. We used the GeLU (Gaussian-error Linear Unit) |
|
activation function with GLU, which is sometimes referred to as GeGLU. The GeLU activation function is a smooth, fully differentiable |
|
approximation to ReLU; we found that this led to a nominal improvement over ReLU. More details on our implementation of GLU can be found here. |
|
The extra gating matrix in a GLU model potentially adds additional parameters to a model; we chose to augment our BERT-Base model with |
|
additional parameters due to GLU modules as it leads to a Pareto improvement across all timescales (which is not true of all larger |
|
models such as BERT-Large). While BERT-Base has 110 million parameters, MosaicBERT-Base has 137 million parameters. Note that |
|
MosaicBERT-Base trains faster than BERT-Base despite having more parameters. |
|
|
|
|
|
# How to use |
|
|
|
## Training data |
|
|
|
MosaicBERT is pretrained using a standard Masked Language Modeling (MLM) objective: the model is given a sequence of |
|
text with some tokens hidden, and it has to predict these masked tokens. MosaicBERT is trained on |
|
the English [“Colossal, Cleaned, Common Crawl” C4 dataset](https://github.com/allenai/allennlp/discussions/5056), which contains roughly 365 million curated text documents scraped |
|
from the internet (equivalent to 156 billion tokens). We used this more modern dataset in place of traditional BERT pretraining |
|
corpora like English Wikipedia and BooksCorpus. |
|
|
|
## Training procedure |
|
|
|
## Evaluation results |
|
|
|
When fine-tuned on downstream tasks, this model achieves the following results: |
|
|
|
GLUE test results: |
|
|
|
| Task | MNLI-(m/mm) | QQP | QNLI | SST-2 | CoLA | STS-B | MRPC | RTE | Average | |
|
|:----:|:-----------:|:----:|:----:|:-----:|:----:|:-----:|:----:|:----:|:-------:| |
|
| | | | | | | | | | | |
|
|
|
## Intended uses & limitations |
