pytorch/xla

torch_xla scan forces inputs to be differentiable

Open

#8,783 opened on Mar 4, 2025

View on GitHub
 (1 comment) (0 reactions) (1 assignee)C++ (575 forks)github user discovery
buggood first issue

Repository metrics

Stars
 (2,786 stars)
PR merge metrics
 (PR metrics pending)

Description

The snippet https://github.com/pytorch/xla/blob/00fac787e986a51e1d5d8b74603867cea4086477/torch_xla/experimental/scan.py#L217-L226 is probably wrong. It adds require_grads=True on all carry inputs and that won't work if one of the carry is a LongTensor.

The most obvious example is that if one of the input is an integer, then it can't possibly have gradients.

Contributor guide

Research direction
Examine the scan implementation in torch xla/experimental/scan.py around line 217-226. Understand why require grads=True is applied to all carry inputs and determine how to conditionally set it only for tensors that support gradients (e.g., not LongTensor). Consider using torch.is floating point or checking the dtype.
Tech stack
pythonpytorch
Domain
backendmachine learning
Issue type
Bug
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.
1-3 hours
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Clear
Prerequisites
PythonPyTorchautograd basics
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
65