A new non-transformer-architecture language model (LLM) called Eagle 7B has been released, standing out by being competitive with Mistral 7B models and excelling at handling multilingual tasks. It is the fastest known framework within the speculative sampling family, being 3x faster than vanilla decoding, 2x faster than Lookahead, and 1.6x faster than Medusa on MT-bench. The Receptance Weighted Key Value (RWKV) architecture behind Eagle-7B aims to reconcile the trade-off between computational efficiency and model performance in sequence processing tasks, combining aspects of both Transformers and RNNs. Multi-language support for Eagle includes Chinese, English, French, German, and Japanese.
π The Receptance Weighted Key Value (RWKV, the architecture behind Eagle-7B) introduced by Peng et al. aims to reconcile the trade-off between computational efficiency and model performance in sequence processing tasks. π RWKV combines aspects of both Transformers and RNNsβ¦ https://t.co/XOYvi3wDhK https://t.co/KncPEkfLmO
We are adding multi-language support for https://t.co/HW3jZmurTg - right now Chinese, English, French, German, and Japanese (in alphabetical order) are supported. Sometimes the LLM sticks with English, but overall it looks pretty good! Check out examples below. https://t.co/FA2PMgW6I2
Paper - "EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty" π EAGLE is the fastest known framework within the speculative sampling family. π On MT-bench, EAGLE is 3x faster than vanilla decoding, 2x faster than Lookahead, and 1.6x faster than Medusa. π₯ πβ¦ https://t.co/YD94rQKGgb
a new RWKV (non-transformer-architecture) LLM called Eagle 7B has just been released. this model stands out by being competitive with Mistral 7B models and excels, in particular, at handling multilingual tasks. model and demo links π https://t.co/MnlcEmgmYs
EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty paper page: https://t.co/SDvnBZEpmh On MT-bench, EAGLE is 3x faster than vanilla decoding, 2x faster than Lookahead, and 1.6x faster than Medusa. Using gpt-fast, EAGLE attains on average 160 tokens/s with⦠https://t.co/waXA6ZJvnF