facebookresearch/metaseq

Integrate LucidRain's RotaryEmbeddings

Open

#621 创建于 2023年1月27日

在 GitHub 查看
 (2 评论) (0 反应) (0 负责人)Python (6,195 star) (701 fork)batch import
enhancementgood first issue

描述

See https://github.com/lucidrains/rotary-embedding-torch/blob/main/rotary_embedding_torch/rotary_embedding_torch.py

And from PaLM paper:

We use RoPE embeddings (Su et al., 2021) rather than absolute or relative position embeddings, since RoPE embeddings have been shown to have better performance on long sequence lengths.

贡献者指南

技术栈
pythonpytorch
领域
machine learning
议题类型
feature
难度面向新贡献者的预计实现难度，1 表示很小改动，5 表示专家级工作。
3
预计时间有经验贡献者完成调查、实现、测试并准备 pull request 的粗略时间范围。
1-2 days
活动状态议题当前的可参与程度：新鲜、活跃、陈旧、阻塞或等待维护者输入。
stale
清晰度议题是否清楚说明期望改动、验收标准和下一步。
mostly clear
前置要求
PythonPyTorchrotary embeddings
新手友好度1-100 的估计分数，表示该议题对首次贡献者的友好程度。
30
研究方向
The issue requests integrating rotary embeddings from lucidrains's implementation into metaseq. Start by examining the existing positional encoding code in metaseq (likely in the model directory). Study the reference implementation and the PaLM paper to understand the integration points. The issue has no linked PRs or detailed discussion, so the first step is to propose a plan in the issue thread.