Eugene Cheah

Bo Peng, Daniel Goldstein, Quentin Anthony et al. · harvard

We present Eagle (RWKV-5) and Finch (RWKV-6), sequence models improving upon the RWKV (RWKV-4) architecture. Our architectural design advancements include multi-headed matrix-valued states and a dynamic recurrence mechanism that improve expressivity while maintaining the inference efficiency characteristics of RNNs. We introduce a new multilingual corpus with 1.12 trillion tokens and a fast tokenizer based on greedy matching for enhanced multilinguality. We trained four Eagle models, ranging from 0.46 to 7.5 billion parameters, and two Finch models with 1.6 and 3.1 billion parameters and find that they achieve competitive performance across a wide variety of benchmarks. We release all our models on HuggingFace under the Apache 2.0 license. Models at: https://huggingface.co/RWKV Training code at: https://github.com/RWKV/RWKV-LM Inference code at: https://github.com/RWKV/ChatRWKV Time-parallel training code at: https://github.com/RWKV/RWKV-infctx-trainer

Daniel Goldstein, Eugene Cheah

We present Key-Value Means ("KVM"), a novel block-recurrence for attention that can accommodate either fixed-size or growing state. Equipping a strong transformer baseline with fixed-size KVM attention layers yields a strong $O(N)$ chunked RNN, while adding only an insignificant number of new parameters. We train a transformer with a growable KVM cache and show it performs competitively on long-context tests with only subquadratic prefill time and sublinear state growth. KVM is implementable with standard operations and without custom kernels, and supports chunk-wise parallelizable training and prefill. It provides many of the benefits of both traditional transformers (expandable context memory, chunk-wise parallelizable training and prefill) and linear RNNs in a single unified package. It can be used on every layer, saving KV-cache memory, and allowing a continuous range of choices of prefill time complexity between $O(N)$ and $O(N^2)$. It can also be implemented in a hybrid solution in tandem with LRNN layers in place of traditional attention, to supplement the LRNN with improved sublinear memory growth context length usage and long context decoding. We release our code at https://github.com/recursal/KVM-paper and trained models at https://huggingface.co/collections/recursal/key-value-means under the Apache 2.0 license.

Daniel Goldstein, Fares Obeid, Eric Alcaide et al.

We introduce GoldFinch, a hybrid Linear Attention/Transformer sequence model that uses a new technique to efficiently generate a highly compressed and reusable KV-Cache in linear time and space with respect to sequence length. GoldFinch stacks our new GOLD transformer on top of an enhanced version of the Finch (RWKV-6) architecture. We train up to 1.5B parameter class models of the Finch, Llama, and GoldFinch architectures, and find dramatically improved modeling performance relative to both Finch and Llama. Our cache size savings increase linearly with model layer count, ranging from 756-2550 times smaller than the traditional transformer cache for common sizes, enabling inference of extremely large context lengths even on limited hardware. Although autoregressive generation has O(n) time complexity per token because of attention, pre-fill computation of the entire initial cache state for a submitted context costs only O(1) time per token due to the use of a recurrent neural network (RNN) to generate this cache. We release our trained weights and training code under the Apache 2.0 license for community use.

Daniel Goldstein, Eric Alcaide, Janna Lu et al.

We present Rapid Attention Distillation to Linear Attention Decoders at Scale (RADLADS), a protocol for rapidly converting softmax attention transformers into linear attention decoder models, along with two new RWKV-variant architectures, and models converted from popular Qwen2.5 open source models in 7B, 32B, and 72B sizes. Our conversion process requires only 350-700M tokens, less than 0.005% of the token count used to train the original teacher models. Converting to our 72B linear attention model costs less than \$2,000 USD at today's prices, yet quality at inference remains close to the original transformer. These models achieve state-of-the-art downstream performance across a set of standard benchmarks for linear attention models of their size. We release all our models on HuggingFace under the Apache 2.0 license, with the exception of our 72B models which are also governed by the Qwen License Agreement. Models at https://huggingface.co/collections/recursal/radlads-6818ee69e99e729ba8a87102 Training Code at https://github.com/recursal/RADLADS-paper

Shuhang Wang, Vivek Kumar Singh, Alex Benjamin et al.

Conventional transfer learning leverages weights of pre-trained networks, but mandates the need for similar neural architectures. Alternatively, knowledge distillation can transfer knowledge between heterogeneous networks but often requires access to the original training data or additional generative networks. Knowledge transfer between networks can be improved by being agnostic to the choice of network architecture and reducing the dependence on original training data. We propose a knowledge transfer approach from a teacher to a student network wherein we train the student on an independent transferal dataset, whose annotations are generated by the teacher. Experiments were conducted on five state-of-the-art networks for semantic segmentation and seven datasets across three imaging modalities. We studied knowledge transfer from a single teacher, combination of knowledge transfer and fine-tuning, and knowledge transfer from multiple teachers. The student model with a single teacher achieved similar performance as the teacher; and the student model with multiple teachers achieved better performance than the teachers. The salient features of our algorithm include: 1)no need for original training data or generative networks, 2) knowledge transfer between different architectures, 3) ease of implementation for downstream tasks by using the downstream task dataset as the transferal dataset, 4) knowledge transfer of an ensemble of models, trained independently, into one student model. Extensive experiments demonstrate that the proposed algorithm is effective for knowledge transfer and easily tunable.

Shuhang Wang, Szu-Yeu Hu, Eugene Cheah et al.

This paper proposes a novel U-Net variant using stacked dilated convolutions for medical image segmentation (SDU-Net). SDU-Net adopts the architecture of vanilla U-Net with modifications in the encoder and decoder operations (an operation indicates all the processing for feature maps of the same resolution). Unlike vanilla U-Net which incorporates two standard convolutions in each encoder/decoder operation, SDU-Net uses one standard convolution followed by multiple dilated convolutions and concatenates all dilated convolution outputs as input to the next operation. Experiments showed that SDU-Net outperformed vanilla U-Net, attention U-Net (AttU-Net), and recurrent residual U-Net (R2U-Net) in all four tested segmentation tasks while using parameters around 40% of vanilla U-Net's, 17% of AttU-Net's, and 15% of R2U-Net's.