Abstract: We introduce the Block-Recurrent Transformer, which applies a transformer layer in a recurrent fashion along a sequence, and has linear complexity with respect to sequence length. Our recurrent cell operates on blocks of tokens rather than single tokens, and leverages parallel computation within a block in order to make efficient use of accelerator hardware. The cell itself is strikingly simple. It is merely a transformer layer: it uses self-attention and cross-attention to efficiently compute a recurrent function over a large set of state vectors and tokens. Our design was inspired in part by LSTM cells, and it uses LSTM-style gates, but it scales the typical LSTM cell up by several orders of magnitude.
Our implementation of recurrence has the same cost in both computation time and parameter count as a conventional transformer layer, but offers dramatically improved perplexity in language modeling tasks over very long sequences. Our model out-performs a long-range Transformer XL baseline by a wide margin, while running twice as fast. We demonstrate its effectiveness on PG19 (books), arXiv papers, and GitHub source code.
Block Recurrent Transformer
Dockerfile, requirements.txt, and environment.yaml because I love chaos.
Differences from the Paper (as of 2022/05/04)
- Keys and values are not shared between the "vertical" and "horizontal" directions (the standard input -> output information flow and the recurrent state flow, respectively).
- The state vectors are augmented with Rotary Embeddings for positional encoding, instead of using learned embeddings.
- The special LSTM gate initialization is not yet implemented.