# triton

**Repository Path**: shengbao/triton

## Basic Information

- **Project Name**: triton
- **Description**: triton learn
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-02-16
- **Last Updated**: 2026-02-16

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# 01triton



# Triton基础

## Triton是什么

Triton：An Intermediate Language and Compiler for Tiled Neural Network Computations（一种用于分块神经网络计算的中间语言与编译器。）

关键字：基于python的DSL，用于GPU的分开神经网络，编译器。

![image-20260216075853526](./README.assets/image-20260216075853526.png)

其中triton处于Kernel环节。triton的编译过程如下：

![image-20260216080158513](./README.assets/image-20260216080158513.png)

Triton编译器可以兼容Pytorch等框架，通过inductor后端降级为Triton kernel，Pytorch中深入学习模型经过TorchDynamo捕获得到FX计算图，再经过inductor后端生成triton代码。

> Philippe Tillet;H. T. Kung;David Cox.Triton: an intermediate language and compiler for tiled neural network computations[A].MAPL 2019: Proceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages[C],2019
> https://www.jokeren.tech/slides/triton_next.pdf
> https://www.jokeren.tech/slides/Tritonbsc.pdf
> https://openai.com/index/triton/
> https://superjomn.github.io/posts/triton-mlir-publish/

## Triton安装

环境：Ubuntu 22.04，Python 3.10，Cuda 12.1，以安装目前最新版本triton3.0.0为例。

1. pip 安装：pip install triton

![image-20260216081522349](./README.assets/image-20260216081522349.png)

2. 源码安装

```
git clone https://github.com/triton-lang/triton.git
cd triton

pip install ninja cmake wheel; # build-time dependencies
pip install -e python
```

![image-20260216081902263](./README.assets/image-20260216081902263.png)

![image-20260216081940530](./README.assets/image-20260216081940530.png)

注：需要在官网找到对应llvm的版本

![image-20260216082117589](./README.assets/image-20260216082117589.png)

![image-20260216082135643](./README.assets/image-20260216082135643.png)

![image-20260216082152679](./README.assets/image-20260216082152679.png)

# Triton编写程序

## Triton语法

参考：https://triton-lang.org/main/index.html

常见的triton语法：

```
jit: Decorator for JIT-compiling a function using the Triton compiler.
autotune: Decorator for auto-tuning a triton. jit 'd function.
heuristics: Decorator for specifying how the values of certain meta-parameters may be computed.
config: An object that represents a possible kernel configuration for the auto-tuner to try.
```

**jit**

一个jit装饰器，用于使用triton编译器，对函数进行jit编译。该函数将在gpu上运行。使用jit装饰的函数只能访问python基元、triton包里的内置函数、该函数参数、其他jit函数。

**Triton.autotune**

用于评估所有配置，内核将使用多次，最终使用最好的配置进行运行。Decorator for auto-tuning a triton.jit'd function.

```

@triton.autotune(configs=[
    triton.Config(kwargs={'BLOCK_SIZE': 128}, num_warps=4),
    triton.Config(kwargs={'BLOCK_SIZE': 1024}, num_warps=8),
],
key=['x_size'] # the two above configs will be evaluated anytime
# the value of x_size changes
)
@triton.jit
def kernel(x_ptr, x_size, **META):
    BLOCK_SIZE = META['BLOCK_SIZE']
```