Paran0idy / Tenet

A DL Framework for Tensor Computation

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Tenet

A DL Framework for Tensor Computation, inspired by Needle Framework in CMU 10-414/714: Deep Learning Systems.

Architecture Overview

Architechture

Install

git clone

git clone https://github.com/Paran0idy/Tenet.git

build

cd ./Tenet
make

dependence

  • OpenAI Triton
  • Pytorch
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip3 install triton

env

# using ndarray backend cuda, cpu or triton
export PYTHONPATH=./python && export NEEDLE_BACKEND=nd

Frontend

Autograd

Module

Optim

Backend

OpenAI Triton

  • matmul
    • MMA instruction using Tensor Cores, like Ampere Arch
      • mma.sync.align.m16n8k16
      • ldmatrix.sync
      • ldmatrix.trans.sync
  • reduce
  • element-wise

NVIDIA CUDA

  • matmul
  • reduce
  • element-wise

X86 CPU

  • matmul
  • reduce
  • element-wise

About

A DL Framework for Tensor Computation


Languages

Language:Python 64.4%Language:Cuda 23.5%Language:C++ 10.3%Language:CMake 1.7%Language:Makefile 0.2%