heinsen_routing
RWKV-LM
heinsen_routing | RWKV-LM | |
---|---|---|
7 | 84 | |
160 | 11,657 | |
0.0% | - | |
2.7 | 8.8 | |
about 1 year ago | 9 days ago | |
Python | Python | |
MIT License | Apache License 2.0 |
Stars - the number of stars that a project has on GitHub. Growth - month over month growth in stars.
Activity is a relative number indicating how actively a project is being developed. Recent commits have higher weight than older ones.
For example, an activity of 9.0 indicates that a project is amongst the top 10% of the most actively developed projects that we are tracking.
heinsen_routing
-
What can LLMs never do?
At one point I experimented a little with transformers that had access to external memory searchable via KNN lookups https://github.com/lucidrains/memorizing-transformers-pytorc... or via routed queries with https://github.com/glassroom/heinsen_routing . Both approaches seemed to work for me, but I had to put that work on hold for reasons outside my control.
-
A Surprisingly Effective Way to Estimate Token Importance in LLM Prompts
Simple and in hindsight, obvious:
1. Run the text through a document embeddding model and save the embedding.
2. Remove one token at a time, and compute the cosine similarity of the new document embedding to the original one.
3. Compute importance as a function of the change in cosine similarity.
Nice.
Also check out https://github.com/glassroom/heinsen_routing . It takes n embeddings and outputs m embeddings, and also gives you an n×m matrix with credit assignments, without having to remove tokens one by one, which can be prohibitively slow for long texts.
-
Unlimiformer: Long-Range Transformers with Unlimited Length Input
After a very quick read, that's my understanding too: It's just KNN search. So I agree on points 1-3. When something works well, I don't care much about point 4.
I've had only mixed success with KNN search. Maybe I haven't done it right? Nothing seems to work quite as well for me as explicit token-token interactions by some form of attention, which as we all know is too costly for long sequences (O(n²)). Lately I've been playing with https://github.com/hazyresearch/safari , which uses a lot less compute and seems promising. Otherwise, for long sequences I've yet to find something better than https://github.com/HazyResearch/flash-attention for n×n interactions and https://github.com/glassroom/heinsen_routing for n×m interactions. If anyone here has other suggestions, I'd love to hear about them.
-
Scaling Transformer to 1M tokens and beyond with RMT
Here's a list of tools for scaling up transformer context that have github repos:
* FlashAttention: In my experience, the current best solution for n² attention, but it's very hard to scale it beyond the low tens of thousands of tokens. Code: https://github.com/HazyResearch/flash-attention
* Heinsen Routing: In my experience, the current best solution for n×m attention. I've used it to pull up more than a million tokens as context. It's not a substitute for n² attention. Code: https://github.com/glassroom/heinsen_routing
* RWKV: A sort-of-recurrent model which claims to have performance comparable to n² attention in transformers. In my limited experience, it doesn't. Others agree: https://twitter.com/arankomatsuzaki/status/16390003799784038... . Code: https://github.com/BlinkDL/RWKV-LM
* RMT (this method): I'm skeptical that the recurrent connections will work as well as n² attention in practice, but I'm going to give it a try. Code: https://github.com/booydar/t5-experiments/tree/scaling-repor...
In addition, there's a group at Stanford working on state-space models that looks promising to me. The idea is to approximate n² attention dynamically using only O(n log n) compute. There's no code available, but here's a blog post about it: https://hazyresearch.stanford.edu/blog/2023-03-27-long-learn...
If anyone here has other suggestions for working with long sequences (hundreds of thousands to millions of tokens), I'd love to learn about them.
-
From Deep to Long Learning
I imagine you could, maybe by using something like this https://github.com/glassroom/heinsen_routing#sequence-to-vec... ... but I doubt you'd be able to match the training efficiency of triangular masking in auto-regressive transformers. With routing, you'd have to train the model one time-step at a time, instead of all time-steps in parallel like a masked auto-regressive transformer.
- New algorithm can route sequences with 1M+ token embeddings in one GPU
RWKV-LM
-
Do LLMs need a context window?
https://github.com/BlinkDL/RWKV-LM#rwkv-discord-httpsdiscord... lists a number of implementations of various versions of RWKV.
https://github.com/BlinkDL/RWKV-LM#rwkv-parallelizable-rnn-w... :
> RWKV: Parallelizable RNN with Transformer-level LLM Performance (pronounced as "RwaKuv", from 4 major params: R W K V)
> RWKV is an RNN with Transformer-level LLM performance, which can also be directly trained like a GPT transformer (parallelizable). And it's 100% attention-free. You only need the hidden state at position t to compute the state at position t+1. You can use the "GPT" mode to quickly compute the hidden state for the "RNN" mode.
> So it's combining the best of RNN and transformer - great performance, fast inference, saves VRAM, fast training, "infinite" ctx_len, and free sentence embedding (using the final hidden state).
> "Our latest version is RWKV-6,*
- People who've used RWKV, whats your wishlist for it?
- Paving the way to efficient architectures: StripedHyena-7B
-
Understanding Deep Learning
That is not true. There are RNNs with transformer/LLM-like performance. See https://github.com/BlinkDL/RWKV-LM.
-
Q-Transformer: Scalable Reinforcement Learning via Autoregressive Q-Functions
This is what RWKV (https://github.com/BlinkDL/RWKV-LM) was made for, and what it will be good at.
Wow. Pretty darn cool! <3 :'))))
-
Personal GPT: A tiny AI Chatbot that runs fully offline on your iPhone
Thanks for the support! Two weeks ago, I'd have said longer contexts on small on-device LLMs are at least a year away, but developments from last week seem to indicate that it's well within reach. Once the low hanging product features are done, I think it's a worthy problem to spend a couple of weeks or perhaps even months on. Speaking of context lengths, recurrent models like RWKV technically have infinite context lengths, but in practice the context slowly fades away after a few thousands of tokens.
-
"If you see a startup claiming to possess top-secret results leading to human level AI, they're lying or delusional. Don't believe them!" - Yann LeCun, on the conspiracy theories of "X company has reached AGI in secret"
This is the reason there are only a few AI labs, and they show little of the theoretical and scientific understanding you believe is required. Go check their code, there's nothing there. Even the transformer with it's heads and other architectural elements turns out to not do anything and it is less efficient than RNNs. (see https://github.com/BlinkDL/RWKV-LM)
-
The Secret Sauce behind 100K context window in LLMs: all tricks in one place
I've been pondering the same thing, as simply extending the context window in a straightforward manner would lead to a significant increase in computational resources. I've had the opportunity to experiment with Anthropics' 100k model, and it's evident that they're employing some clever techniques to make it work, albeit with some imperfections. One interesting observation is that their prompt guide recommends placing instructions after the reference text when inputting lengthy text bodies. I noticed that the model often disregarded the instructions if placed beforehand. It's clear that the model doesn't allocate the same level of "attention" to all parts of the input across the entire context window.
Moreover, the inability to cache transformers makes the use of large context windows quite costly, as all previous messages must be sent with each call. In this context, the RWKV-LM project on GitHub (https://github.com/BlinkDL/RWKV-LM) might offer a solution. They claim to achieve performance comparable to transformers using an RNN, which could potentially handle a 100-page document and cache it, thereby eliminating the need to process the entire document with each subsequent query. However, I suspect RWKV might fall short in handling complex tasks that require maintaining multiple variables in memory, such as mathematical computations, but it should suffice for many scenarios.
On a related note, I believe Anthropics' Claude is somewhat underappreciated. In some instances, it outperforms GPT4, and I'd rank it somewhere between GPT4 and Bard overall.
-
Meta's plan to offer free commercial AI models puts pressure on Google, OpenAI
> The only reason open-source LLMs have a heartbeat is they’re standing on Meta’s weights.
Not necessarily.
RWKV, for example, is a different architecture that wasn't based on Facebook's weights whatsoever. I don't know where BlinkDL (the author) got the training data, but they seem to have done everything mostly independently otherwise.
https://github.com/BlinkDL/RWKV-LM
disclaimer: I've been doing a lot of work lately on an implementation of CPU inference for this model, so I'm obviously somewhat biased since this is the model I have the most experience in.
-
Eliezer Yudkowsky - open letter on AI
I think the main concern is that, due to the resources put into LLM research for finding new ways to refine and improve them, that work can then be used by projects that do go the extra mile and create things that are more than just LLMs. For example, RWKV is similar to an LLM but will actually change its own model after every processed token, thus letting it remember things longer-term without the use of 'context tokens'.
What are some alternatives?
safari - Convolutions for Sequence Modeling
llama - Inference code for Llama models
flash-attention - Fast and memory-efficient exact attention
alpaca-lora - Instruct-tune LLaMA on consumer hardware
block-recurrent-transformer-pytorch - Implementation of Block Recurrent Transformer - Pytorch
iris - Transformers are Sample-Efficient World Models. ICLR 2023, notable top 5%.
koboldcpp - A simple one-file way to run various GGML and GGUF models with KoboldAI's UI
block-recurrent-transformer-py
gpt4all - gpt4all: run open-source LLMs anywhere
recurrent-memory-transformer - [NeurIPS 22] [AAAI 24] Recurrent Transformer-based long-context architecture.
RWKV-CUDA - The CUDA version of the RWKV language model ( https://github.com/BlinkDL/RWKV-LM )