transfomers-silicon-research

transfomers-silicon-research

Transformer模型硬件实现研究进展

本项目汇集了Transformer模型硬件实现的研究资料,包括BERT及其优化方案。内容涵盖算法-硬件协同设计、神经网络加速器、量化和剪枝等技术。项目提供了详细的论文列表,涉及FPGA实现、功耗优化等多个领域,全面展示了Transformer硬件加速的最新研究进展。

TransformerBERT自然语言处理硬件加速神经网络Github开源项目

Transformer Models Silicon Research

Research and Materials on Hardware implementation of Transformer Models

<!-- <p align="center"> <img src="https://img.shields.io/badge/-WIP-ff69b4?style=flat-square"/> </p> <p align="center"> <img src="https://img.shields.io/badge/Progress-%2599-ef6c00?labelColor=1565c0&style=flat-square"/> </p> -->

How to Contribute

You can add new papers via pull requests, Please check data/papers.yaml and if your paper is not in list, add entity at the last item and create pull request.

Transformer and BERT Model

  • BERT is a method of pre-training language representations, meaning that we train a general-purpose language understanding model on a large text corpus (like Wikipedia) and then use that model for downstream NLP tasks.

  • BERT was created and published in 2018 by Jacob Devlin and his colleagues from Google. BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks.

<p align="center"> <img src="./data/img/BERT-ARCH.png" width='480' /> </p>
  • BERT is a Transformer-based model.
    • The architecture of BERT is similar to the original Transformer model, except that BERT has two separate Transformer models: one for the left-to-right direction (the “encoder”) and one for the right-to-left direction (the “encoder”).
    • The output of each model is the hidden state output by the final Transformer layer. The two models are pre-trained jointly on a large corpus of unlabeled text. The pre-training task is a simple and straightforward masked language modeling objective.
    • The pre-trained BERT model can then be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications.

Reference Papers

1. Attention Is All You Need

DOI-Link PDF-Download

Code-Link Code-Link

<details> <summary><img src="https://img.shields.io/badge/ABSTRACT-9575cd?&style=plastic"/></summary> The dominant sequence transduction models are based on complex recurrent or convolutional neural networks that include an encoder and a decoder. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 Englishto-German translation task, improving over the existing best results, including ensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data. </details>

2. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

DOI-Link PDF-Download Code-Link Code-Link

<details> <summary><img src="https://img.shields.io/badge/ABSTRACT-9575cd?&style=plastic"/></summary> We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models (Peters et al., 2018a; Radford et al., 2018), BERT is designed to pretrain deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be finetuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial taskspecific architecture modifications. <br> BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement). </details>

Hardware Research

2018

Algorithm-Hardware Co-Design of Single Shot Detector for Fast Object Detection on FPGAs

DOI-Link

SparseNN: An energy-efficient neural network accelerator exploiting input and output sparsity

DOI-Link PDF-Link


2019

A Power Efficient Neural Network Implementation on Heterogeneous FPGA and GPU Devices

DOI-Link

A Simple and Effective Approach to Automatic Post-Editing with Transfer Learning

DOI-Link

An Evaluation of Transfer Learning for Classifying Sales Engagement Emails at Large Scale

DOI-Link

MAGNet: A Modular Accelerator Generator for Neural Networks

DOI-Link PDF-Link

mRNA: Enabling Efficient Mapping Space Exploration for a Reconfiguration Neural Accelerator

DOI-Link PDF-Link

Pre-trained bert-gru model for relation extraction

DOI-Link

Q8BERT: Quantized 8Bit BERT

DOI-Link PDF-Link

Structured pruning of a BERT-based question answering model

DOI-Link PDF-Link

Structured pruning of large language models

DOI-Link PDF-Link

Tinybert: Distilling bert for natural language understanding

DOI-Link PDF-Link


2020

A Low-Cost Reconfigurable Nonlinear Core for Embedded DNN Applications

DOI-Link

A Multi-Neural Network Acceleration Architecture

DOI-Link

A Primer in BERTology: What We Know About How BERT Works

DOI-Link

A Reconfigurable DNN Training Accelerator on FPGA

DOI-Link

A^3: Accelerating Attention Mechanisms in Neural Networks with Approximation

DOI-Link

Emerging Neural Workloads and Their Impact on Hardware

DOI-Link

Accelerating event detection with DGCNN and FPGAS

DOI-Link

An Empirical Analysis of BERT Embedding for Automated Essay Scoring

DOI-Link

**An investigation on different underlying quantization schemes for pre-trained language

编辑推荐精选

openai-agents-python

openai-agents-python

OpenAI Agents SDK,助力开发者便捷使用 OpenAI 相关功能。

openai-agents-python 是 OpenAI 推出的一款强大 Python SDK,它为开发者提供了与 OpenAI 模型交互的高效工具,支持工具调用、结果处理、追踪等功能,涵盖多种应用场景,如研究助手、财务研究等,能显著提升开发效率,让开发者更轻松地利用 OpenAI 的技术优势。

Hunyuan3D-2

Hunyuan3D-2

高分辨率纹理 3D 资产生成

Hunyuan3D-2 是腾讯开发的用于 3D 资产生成的强大工具,支持从文本描述、单张图片或多视角图片生成 3D 模型,具备快速形状生成能力,可生成带纹理的高质量 3D 模型,适用于多个领域,为 3D 创作提供了高效解决方案。

3FS

3FS

一个具备存储、管理和客户端操作等多种功能的分布式文件系统相关项目。

3FS 是一个功能强大的分布式文件系统项目,涵盖了存储引擎、元数据管理、客户端工具等多个模块。它支持多种文件操作,如创建文件和目录、设置布局等,同时具备高效的事件循环、节点选择和协程池管理等特性。适用于需要大规模数据存储和管理的场景,能够提高系统的性能和可靠性,是分布式存储领域的优质解决方案。

TRELLIS

TRELLIS

用于可扩展和多功能 3D 生成的结构化 3D 潜在表示

TRELLIS 是一个专注于 3D 生成的项目,它利用结构化 3D 潜在表示技术,实现了可扩展且多功能的 3D 生成。项目提供了多种 3D 生成的方法和工具,包括文本到 3D、图像到 3D 等,并且支持多种输出格式,如 3D 高斯、辐射场和网格等。通过 TRELLIS,用户可以根据文本描述或图像输入快速生成高质量的 3D 资产,适用于游戏开发、动画制作、虚拟现实等多个领域。

ai-agents-for-beginners

ai-agents-for-beginners

10 节课教你开启构建 AI 代理所需的一切知识

AI Agents for Beginners 是一个专为初学者打造的课程项目,提供 10 节课程,涵盖构建 AI 代理的必备知识,支持多种语言,包含规划设计、工具使用、多代理等丰富内容,助您快速入门 AI 代理领域。

AEE

AEE

AI Excel全自动制表工具

AEE 在线 AI 全自动 Excel 编辑器,提供智能录入、自动公式、数据整理、图表生成等功能,高效处理 Excel 任务,提升办公效率。支持自动高亮数据、批量计算、不规则数据录入,适用于企业、教育、金融等多场景。

UI-TARS-desktop

UI-TARS-desktop

基于 UI-TARS 视觉语言模型的桌面应用,可通过自然语言控制计算机进行多模态操作。

UI-TARS-desktop 是一款功能强大的桌面应用,基于 UI-TARS(视觉语言模型)构建。它具备自然语言控制、截图与视觉识别、精确的鼠标键盘控制等功能,支持跨平台使用(Windows/MacOS),能提供实时反馈和状态显示,且数据完全本地处理,保障隐私安全。该应用集成了多种大语言模型和搜索方式,还可进行文件系统操作。适用于需要智能交互和自动化任务的场景,如信息检索、文件管理等。其提供了详细的文档,包括快速启动、部署、贡献指南和 SDK 使用说明等,方便开发者使用和扩展。

Wan2.1

Wan2.1

开源且先进的大规模视频生成模型项目

Wan2.1 是一个开源且先进的大规模视频生成模型项目,支持文本到图像、文本到视频、图像到视频等多种生成任务。它具备丰富的配置选项,可调整分辨率、扩散步数等参数,还能对提示词进行增强。使用了多种先进技术和工具,在视频和图像生成领域具有广泛应用前景,适合研究人员和开发者使用。

爱图表

爱图表

全流程 AI 驱动的数据可视化工具,助力用户轻松创作高颜值图表

爱图表(aitubiao.com)就是AI图表,是由镝数科技推出的一款创新型智能数据可视化平台,专注于为用户提供便捷的图表生成、数据分析和报告撰写服务。爱图表是中国首个在图表场景接入DeepSeek的产品。通过接入前沿的DeepSeek系列AI模型,爱图表结合强大的数据处理能力与智能化功能,致力于帮助职场人士高效处理和表达数据,提升工作效率和报告质量。

Qwen2.5-VL

Qwen2.5-VL

一款强大的视觉语言模型,支持图像和视频输入

Qwen2.5-VL 是一款强大的视觉语言模型,支持图像和视频输入,可用于多种场景,如商品特点总结、图像文字识别等。项目提供了 OpenAI API 服务、Web UI 示例等部署方式,还包含了视觉处理工具,有助于开发者快速集成和使用,提升工作效率。

下拉加载更多