triton-lang/triton

how to implement backward propagation

Open

#1,189 opened on Feb 13, 2023

View on GitHubContributor guide
 (4 comments) (0 reactions) (0 assignees)MLIR (3,126 forks)github user discovery
help wanted

Repository metrics

Stars
 (19,982 stars)
PR merge metrics
 (Avg merge 2d 18h) (185 merged PRs in 30d)

Description

Hi. I am new to triton and cuda. From my understanding, when we implement a customized pytorch operator using cuda, we need to define both forward and backward function, so that the gradients are propagated properly. However, in the triton document (e.g., how to implement vector addition, fused softmax ...), there seems nothing about backward function? Does that mean the backward function is automatically generated by triton? Thx!

Contributor guide

Research direction
Read Triton documentation on autodiff and manually implementing backward functions. Check examples like fused softmax for backward pass patterns.
Tech stack
pythonpytorch
Domain
backendmachine learning
Issue type
Research
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
2
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
Under 1 hour
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Fresh
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Clear
Prerequisites
PyTorchCUDA basics
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
85