Transformer networks as recurrent or state-space models
December 20, 2017 — January 16, 2025
The intersection of transformers and recurrent or state space models. Much to say here, but only scrapbook notes for now.
1 Tokens as recurrent state
TBD
2 RWKV
State-space, i.e. recurrent transformers, without (classic) attention. Suggestive connection to S4 models.
RWKV is inspired by Apple’s Attention Free Transformer (Zhai et al. 2021). …
How to combine the best of transformers and RNNs? The main drawback of transformer-based models is that it can become challenging to run a model with a context window that is larger than a certain value, as the attention scores are computed simultaneously for the entire sequence.
RNNs natively support very long context lengths—only limited by the context length seen in training, but this can be extended to millions of tokens with careful coding. Currently, there are RWKV models trained on a context length of 8192 (
ctx8192) and they are as fast asctx1024models and require the same amount of RAM.The major drawbacks of traditional RNN models and how RWKV is different:
- Traditional RNN models are unable to utilise very long contexts (LSTM can only manage ~100 tokens when used as a LM). However, RWKV can utilise thousands of tokens and beyond…
- Traditional RNN models cannot be parallelised when training. RWKV is similar to a “linearised GPT” and it trains faster than GPT.
By combining both advantages into a single architecture, the hope is that RWKV can grow to become more than the sum of its parts.
- Introducing RWKV - An RNN with the advantages of a transformer
- How the RWKV language model works
- The RWKV language model: An RNN with the advantages of a transformer
- BlinkDL/ChatRWKV: ChatRWKV is like ChatGPT but powered by RWKV (100% RNN) language model, and open source.
2.1 Mamba
Interesting connection to transformers.
Foundation models, now powering most of the exciting applications in deep learning, are almost universally based on the Transformer architecture and its core attention module. Many subquadratic-time architectures such as linear attention, gated convolution and recurrent models, and structured state space models (SSMs) have been developed to address Transformers’ computational inefficiency on long sequences, but they have not performed as well as attention on important modalities such as language. We identify that a key weakness of such models is their inability to perform content-based reasoning, and make several improvements. First, simply letting the SSM parameters be functions of the input addresses their weakness with discrete modalities, allowing the model to selectively propagate or forget information along the sequence length dimension depending on the current token. Second, even though this change prevents the use of efficient convolutions, we design a hardware-aware parallel algorithm in recurrent mode. We integrate these selective SSMs into a simplified end-to-end neural network architecture without attention or even MLP blocks (Mamba). Mamba enjoys fast inference (5× higher throughput than Transformers) and linear scaling in sequence length, and its performance improves on real data up to million-length sequences. As a general sequence model backbone, Mamba achieves state-of-the-art performance across several modalities such as language, audio, and genomics. On language modelling, our Mamba-3B model outperforms Transformers of the same size and matches Transformers twice its size, both in pretraining and downstream evaluation.
See
- canonical implementation: state-spaces/mamba: Mamba SSM architecture
- pedagogic implementation Implementation of Mamba in one file of PyTorch | Hacker News
- A Visual Guide to Mamba and State Space Models
- What is Mamba?
- An Introduction to the Mamba LLM Architecture: A New Paradigm in Machine Learning | DataCamp
and (Cai et al. 2024; Gu and Dao 2023; Hu et al. 2024; Patro and Agneeswaran 2024).
There is now a sequel, Mamba-2 (Dao and Gu 2024).
3 Memory
Memory-augmented models.
For example, Behrouz, Zhong, and Mirrokni (2024) introduced “Titans”.
Over more than a decade there has been an extensive research effort on how to effectively utilise recurrent models and attention. While recurrent models aim to compress the data into a fixed-size memory (called hidden state), attention allows attending to the entire context window, capturing the direct dependencies of all tokens. This more accurate modelling of dependencies, however, comes with a quadratic cost, limiting the model to a fixed-length context. We present a new neural long-term memory module that learns to memorise historical context and helps attention to attend to the current context while utilising long past information. We show that this neural memory has the advantage of fast parallelisable training while maintaining a fast inference. From a memory perspective, we argue that attention due to its limited context but accurate dependency modelling performs as a short-term memory, while neural memory due to its ability to memorise the data, acts as a long-term, more persistent, memory. Based on these two modules, we introduce a new family of architectures, called Titans, and present three variants to address how one can effectively incorporate memory into this architecture. Our experimental results on language modelling, common-sense reasoning, genomics, and time series tasks show that Titans are more effective than Transformers and recent modern linear recurrent models. They further can effectively scale to larger than 2M context window size with higher accuracy in needle-in-haystack tasks compared to baselines.
See also Google Research Unveils “Transformers 2.0” aka TITANS
4 RecurrentGPT
I am not sure what RecurrentGPT (Zhou et al. 2023) is. Let us find out
GitHub - aiwaves-cn/RecurrentGPT
RecurrentGPT replaces the vectorised elements (i.e., cell state, hidden state, input, and output) in a Long-short Term Memory RNN (LSTM) with natural language (i.e., paragraphs of texts), and simulates the recurrence mechanism with prompt engineering.
At each timestep \(t\), RecurrentGPT receives a paragraph of text and a brief plan of the next paragraph, which are both generated in step \(t − 1\). It then attends to the long-term memory, which contains the summaries of all previously generated paragraphs and can be stored on hard drives, and relevant paragraphs can be retrieved with semantic search.
RecurrentGPT also maintains a short-term memory that summarises key information within recent timesteps in natural language and is updated at each time step. RecurrentGPT combines all aforementioned inputs in a prompt and asks the backbone LLM to generate a new paragraph, a short plan for the next paragraph, and updates the long-short term memory by rewriting the short-term memory and appending the summary of the output paragraph to the long-term memory.
Hmm; I’m not convinced by that one.
5 Practicalities
For you and me, see AI democratizateion.