transformers/docs/source/en/model_doc/mega.md

MEGA

This model is in maintenance mode only, we don't accept any new PRs changing its code. If you run into any issues running this model, please reinstall the last version that supported this model: v4.40.2. You can do so by running the following command: pip install -U transformers==4.40.2.

Overview

The MEGA model was proposed in Mega: Moving Average Equipped Gated Attention by Xuezhe Ma, Chunting Zhou, Xiang Kong, Junxian He, Liangke Gui, Graham Neubig, Jonathan May, and Luke Zettlemoyer. MEGA proposes a new approach to self-attention with each encoder layer having a multi-headed exponential moving average in addition to a single head of standard dot-product attention, giving the attention mechanism stronger positional biases. This allows MEGA to perform competitively to Transformers on standard benchmarks including LRA while also having significantly fewer parameters. MEGA's compute efficiency allows it to scale to very long sequences, making it an attractive option for long-document NLP tasks.

The abstract from the paper is the following:

*The design choices in the Transformer attention mechanism, including weak inductive bias and quadratic computational complexity, have limited its application for modeling long sequences. In this paper, we introduce Mega, a simple, theoretically grounded, single-head gated attention mechanism equipped with (exponential) moving average to incorporate inductive bias of position-aware local dependencies into the position-agnostic attention mechanism. We further propose a variant of Mega that offers linear time and space complexity yet yields only minimal quality loss, by efficiently splitting the whole sequence into multiple chunks with fixed length. Extensive experiments on a wide range of sequence modeling benchmarks, including the Long Range Arena, neural machine translation, auto-regressive language modeling, and image and speech classification, show that Mega achieves significant improvements over other sequence models, including variants of Transformers and recent state space models. *

This model was contributed by mnaylor. The original code can be found here.

Usage tips

• MEGA can perform quite well with relatively few parameters. See Appendix D in the MEGA paper for examples of architectural specs which perform well in various settings. If using MEGA as a decoder, be sure to set bidirectional=False to avoid errors with default bidirectional.
• Mega-chunk is a variant of mega that reduces time and spaces complexity from quadratic to linear. Utilize chunking with MegaConfig.use_chunking and control chunk size with MegaConfig.chunk_size

Implementation Notes

• The original implementation of MEGA had an inconsistent expectation of attention masks for padding and causal self-attention between the softmax attention and Laplace/squared ReLU method. This implementation addresses that inconsistency.
• The original implementation did not include token type embeddings; this implementation adds support for these, with the option controlled by MegaConfig.add_token_type_embeddings

MegaConfig

autodoc MegaConfig

MegaModel

autodoc MegaModel - forward

MegaForCausalLM

autodoc MegaForCausalLM - forward

MegaForMaskedLM

autodoc MegaForMaskedLM - forward

MegaForSequenceClassification

autodoc MegaForSequenceClassification - forward

MegaForMultipleChoice

autodoc MegaForMultipleChoice - forward

MegaForTokenClassification

autodoc MegaForTokenClassification - forward

MegaForQuestionAnswering

autodoc MegaForQuestionAnswering - forward